Probabilistic Power Forecasting in a Microgrid for the Integration of Distributed Energy Resources

IFAC PapersOnLine 58-2 (2024) 150–155 ScienceDirect Available online at www.sciencedirect.com 2405-8963 Copyright © 2024 The Authors. This is an open access article under the CC BY-NC-ND license . Peer review under responsibility of International Federation of Automatic Control. 10.1016/j.ifacol.2024.07.106 10.1016/j.ifacol.2024.07.106 2405-8963 Probabilistic Power Forecasting in a Microgrid for the Integration of Distributed Energy Resources Antonio Parejo* Stefano Bracco ** Sebastián García* Enrique Personal* Juan I. Guerrero* Carlos León* * Department of Electronic Technology, Escuela Politécnica Superior, Universidad de Sevilla, Sevilla, 41011, Spain (e-mail: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]). ** Department of Electrical, Electronic, Telecommunication Engineering and Naval Architecture, University of Genoa, Genova, 16145, Italy (e-mail: [email protected]) Abstract: The integration of distributed energy resources in the power grid supposes a great challenge, considering the expected growing in the electric mobility and the renewable generation. Therefore, to support the successful deployment of these resources, energy management according to the expected behaviors of the elements of power grids is required. In this sense, the probabilistic forecasting methods are gaining acceptance, as they provide richer information than deterministic methods thanks to the consideration of expected probabilities. One of their problems is that the application of metrics for choosing the best models is not simple, depending on the objective of the type of probabilistic prediction that must be performed. In this paper, diverse types of forecasting are performed, and two possible approaches for selecting interval forecasting models are compared in a case study centered on the Smart Polygeneration Microgrid (SPM) located on the Savona Campus of University of Genova (Italy). The results show that, for the machine learning models that were tested, it is better to obtain a model for each of the different types of forecasting that are required (quantiles, and the prediction intervals) to have better-quality predictions. The average improvement on the quality of the tested prediction intervals for the three forecasted variables were a 27.40%, 7.39%, and 28.65% in their Winkler score respectively. Keywords: Stochastic forecasting, probabilistic forecasting, microgrids, distributed energy resources, demand. 1. INTRODUCTION Microgrids are bound to be one of the key elements for the future of the power system. A microgrid can be defined as a small electric network that is designed for a reliable and massive integration of distributed generations (DG) at the low (or medium) voltage level, especially the renewables (solar, wind, and other low carbon technologies), as done by Gouveia et al. (2013) and Harmouch et al. (2018) and even hydrogen, as Ennassiri et al. (2023) say. Regarding its operation modes, as it is states by Ma & Ma (2018), a microgrid system normally operates in two modes, which are called “grid-connected mode” and “island mode”. Under the grid-connected mode, the microgrid system adjusts dispatchable units according to different purposes (e.g., reducing economic costs or to maintain electrical values between feasible bounds). Beyond the power needs, some microgrids could be equipped with power quality monitoring, being able to disconnect themselves from the main grid if the power quality is below a certain level and preventing in this way possible failures and technical problems due to instabilities from the main grid, as Ma & Ma (2018) say. For taking advantage from the microgrid assets and controllability, it is required to perform a scheduling for the operation of the microgrid. This means to decide which generation units should be connected and disconnected, when it is convenient to charge a storage system, or even when the microgrid should swap to islanding mode and disconnect from the main grid. This process is usually referred to as microgrid scheduling, or microgrid optimization. Two main general types of microgrid optimization methods can be found in the literature taking into account the nature of their uncertainties data, i.e., the forecast of the power variables of interest that have to be considered to perform the optimization (expected load consumption, expected generation, etc.). These can be deterministic or probabilistic (sometimes also called stochastic). Deterministic scheduling methods, also called deterministic Model Predictive Control (deterministic MPC) methods by some authors as Franc et al. (2021) are some of the most famous lookahead techniques. Deterministic methods require a value for each unknown input variable (i.e., a prediction or estimation). Stochastic scheduling methods requires that the inputs are given in the shape of quantiles, interval probabilities, scenarios, or others. Some of these methods, according to Franc et al. (2021), are the open loop feedback control (OLFC, Probabilistic Power Forecasting in a Microgrid for the Integration of Distributed Energy Resources Antonio Parejo* Stefano Bracco ** Sebastián García* Enrique Personal* Juan I. Guerrero* Carlos León* * Department of Electronic Technology, Escuela Politécnica Superior, Universidad de Sevilla, Sevilla, 41011, Spain (e-mail: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]). ** Department of Electrical, Electronic, Telecommunication Engineering and Naval Architecture, University of Genoa, Genova, 16145, Italy (e-mail: [email protected]) Abstract: The integration of distributed energy resources in the power grid supposes a great challenge, considering the expected growing in the electric mobility and the renewable generation. Therefore, to support the successful deployment of these resources, energy management according to the expected behaviors of the elements of power grids is required. In this sense, the probabilistic forecasting methods are gaining acceptance, as they provide richer information than deterministic methods thanks to the consideration of expected probabilities. One of their problems is that the application of metrics for choosing the best models is not simple, depending on the objective of the type of probabilistic prediction that must be performed. In this paper, diverse types of forecasting are performed, and two possible approaches for selecting interval forecasting models are compared in a case study centered on the Smart Polygeneration Microgrid (SPM) located on the Savona Campus of University of Genova (Italy). The results show that, for the machine learning models that were tested, it is better to obtain a model for each of the different types of forecasting that are required (quantiles, and the prediction intervals) to have better-quality predictions. The average improvement on the quality of the tested prediction intervals for the three forecasted variables were a 27.40%, 7.39%, and 28.65% in their Winkler score respectively. Keywords: Stochastic forecasting, probabilistic forecasting, microgrids, distributed energy resources, demand. 1. INTRODUCTION Microgrids are bound to be one of the key elements for the future of the power system. A microgrid can be defined as a small electric network that is designed for a reliable and massive integration of distributed generations (DG) at the low (or medium) voltage level, especially the renewables (solar, wind, and other low carbon technologies), as done by Gouveia et al. (2013) and Harmouch et al. (2018) and even hydrogen, as Ennassiri et al. (2023) say. Regarding its operation modes, as it is states by Ma & Ma (2018), a microgrid system normally operates in two modes, which are called “grid-connected mode” and “island mode”. Under the grid-connected mode, the microgrid system adjusts dispatchable units according to different purposes (e.g., reducing economic costs or to maintain electrical values between feasible bounds). Beyond the power needs, some microgrids could be equipped with power quality monitoring, being able to disconnect themselves from the main grid if the power quality is below a certain level and preventing in this way possible failures and technical problems due to instabilities from the main grid, as Ma & Ma (2018) say. For taking advantage from the microgrid assets and controllability, it is required to perform a scheduling for the operation of the microgrid. This means to decide which generation units should be connected and disconnected, when it is convenient to charge a storage system, or even when the microgrid should swap to islanding mode and disconnect from the main grid. This process is usually referred to as microgrid scheduling, or microgrid optimization. Two main general types of microgrid optimization methods can be found in the literature taking into account the nature of their uncertainties data, i.e., the forecast of the power variables of interest that have to be considered to perform the optimization (expected load consumption, expected generation, etc.). These can be deterministic or probabilistic (sometimes also called stochastic). Deterministic scheduling methods, also called deterministic Model Predictive Control (deterministic MPC) methods by some authors as Franc et al. (2021) are some of the most famous lookahead techniques. Deterministic methods require a value for each unknown input variable (i.e., a prediction or estimation). Stochastic scheduling methods requires that the inputs are given in the shape of quantiles, interval probabilities, scenarios, or others. Some of these methods, according to Franc et al. (2021), are the open loop feedback control (OLFC, Probabilistic Power Forecasting in a Microgrid for the Integration of Distributed Energy Resources Antonio Parejo* Stefano Bracco ** Sebastián García* Enrique Personal* Juan I. Guerrero* Carlos León* * Department of Electronic Technology, Escuela Politécnica Superior, Universidad de Sevilla, Sevilla, 41011, Spain (e-mail: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]). ** Department of Electrical, Electronic, Telecommunication Engineering and Naval Architecture, University of Genoa, Genova, 16145, Italy (e-mail: [email protected]) Abstract: The integration of distributed energy resources in the power grid supposes a great challenge, considering the expected growing in the electric mobility and the renewable generation. Therefore, to support the successful deployment of these resources, energy management according to the expected behaviors of the elements of power grids is required. In this sense, the probabilistic forecasting methods are gaining acceptance, as they provide richer information than deterministic methods thanks to the consideration of expected probabilities. One of their problems is that the application of metrics for choosing the best models is not simple, depending on the objective of the type of probabilistic prediction that must be performed. In this paper, diverse types of forecasting are performed, and two possible approaches for selecting interval forecasting models are compared in a case study centered on the Smart Polygeneration Microgrid (SPM) located on the Savona Campus of University of Genova (Italy). The results show that, for the machine learning models that were tested, it is better to obtain a model for each of the different types of forecasting that are required (quantiles, and the prediction intervals) to have better-quality predictions. The average improvement on the quality of the tested prediction intervals for the three forecasted variables were a 27.40%, 7.39%, and 28.65% in their Winkler score respectively. Keywords: Stochastic forecasting, probabilistic forecasting, microgrids, distributed energy resources, demand. 1. INTRODUCTION Microgrids are bound to be one of the key elements for the future of the power system. A microgrid can be defined as a small electric network that is designed for a reliable and massive integration of distributed generations (DG) at the low (or medium) voltage level, especially the renewables (solar, wind, and other low carbon technologies), as done by Gouveia et al. (2013) and Harmouch et al. (2018) and even hydrogen, as Ennassiri et al. (2023) say. Regarding its operation modes, as it is states by Ma & Ma (2018), a microgrid system normally operates in two modes, which are called “grid-connected mode” and “island mode”. Under the grid-connected mode, the microgrid system adjusts dispatchable units according to different purposes (e.g., reducing economic costs or to maintain electrical values between feasible bounds). Beyond the power needs, some microgrids could be equipped with power quality monitoring, being able to disconnect themselves from the main grid if the power quality is below a certain level and preventing in this way possible failures and technical problems due to instabilities from the main grid, as Ma & Ma (2018) say. For taking advantage from the microgrid assets and controllability, it is required to perform a scheduling for the operation of the microgrid. This means to decide which generation units should be connected and disconnected, when it is convenient to charge a storage system, or even when the microgrid should swap to islanding mode and disconnect from the main grid. This process is usually referred to as microgrid scheduling, or microgrid optimization. Two main general types of microgrid optimization methods can be found in the literature taking into account the nature of their uncertainties data, i.e., the forecast of the power variables of interest that have to be considered to perform the optimization (expected load consumption, expected generation, etc.). These can be deterministic or probabilistic (sometimes also called stochastic). Deterministic scheduling methods, also called deterministic Model Predictive Control (deterministic MPC) methods by some authors as Franc et al. (2021) are some of the most famous lookahead techniques. Deterministic methods require a value for each unknown input variable (i.e., a prediction or estimation). Stochastic scheduling methods requires that the inputs are given in the shape of quantiles, interval probabilities, scenarios, or others. Some of these methods, according to Franc et al. (2021), are the open loop feedback control (OLFC, Probabilistic Power Forecasting in a Microgrid for the Integration of Distributed Energy Resources Antonio Parejo* Stefano Bracco ** Sebastián García* Enrique Personal* Juan I. Guerrero* Carlos León* * Department of Electronic Technology, Escuela Politécnica Superior, Universidad de Sevilla, Sevilla, 41011, Spain (e-mail: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]). ** Department of Electrical, Electronic, Telecommunication Engineering and Naval Architecture, University of Genoa, Genova, 16145, Italy (e-mail: [email protected]) Abstract: The integration of distributed energy resources in the power grid supposes a great challenge, considering the expected growing in the electric mobility and the renewable generation. Therefore, to support the successful deployment of these resources, energy management according to the expected behaviors of the elements of power grids is required. In this sense, the probabilistic forecasting methods are gaining acceptance, as they provide richer information than deterministic methods thanks to the consideration of expected probabilities. One of their problems is that the application of metrics for choosing the best models is not simple, depending on the objective of the type of probabilistic prediction that must be performed. In this paper, diverse types of forecasting are performed, and two possible approaches for selecting interval forecasting models are compared in a case study centered on the Smart Polygeneration Microgrid (SPM) located on the Savona Campus of University of Genova (Italy). The results show that, for the machine learning models that were tested, it is better to obtain a model for each of the different types of forecasting that are required (quantiles, and the prediction intervals) to have better-quality predictions. The average improvement on the quality of the tested prediction intervals for the three forecasted variables were a 27.40%, 7.39%, and 28.65% in their Winkler score respectively. Keywords: Stochastic forecasting, probabilistic forecasting, microgrids, distributed energy resources, demand. 1. INTRODUCTION Microgrids are bound to be one of the key elements for the future of the power system. A microgrid can be defined as a small electric network that is designed for a reliable and massive integration of distributed generations (DG) at the low (or medium) voltage level, especially the renewables (solar, wind, and other low carbon technologies), as done by Gouveia et al. (2013) and Harmouch et al. (2018) and even hydrogen, as Ennassiri et al. (2023) say. Regarding its operation modes, as it is states by Ma & Ma (2018), a microgrid system normally operates in two modes, which are called “grid-connected mode” and “island mode”. Under the grid-connected mode, the microgrid system adjusts dispatchable units according to different purposes (e.g., reducing economic costs or to maintain electrical values between feasible bounds). Beyond the power needs, some microgrids could be equipped with power quality monitoring, being able to disconnect themselves from the main grid if the power quality is below a certain level and preventing in this way possible failures and technical problems due to instabilities from the main grid, as Ma & Ma (2018) say. For taking advantage from the microgrid assets and controllability, it is required to perform a scheduling for the operation of the microgrid. This means to decide which generation units should be connected and disconnected, when it is convenient to charge a storage system, or even when the microgrid should swap to islanding mode and disconnect from the main grid. This process is usually referred to as microgrid scheduling, or microgrid optimization. Two main general types of microgrid optimization methods can be found in the literature taking into account the nature of their uncertainties data, i.e., the forecast of the power variables of interest that have to be considered to perform the optimization (expected load consumption, expected generation, etc.). These can be deterministic or probabilistic (sometimes also called stochastic). Deterministic scheduling methods, also called deterministic Model Predictive Control (deterministic MPC) methods by some authors as Franc et al. (2021) are some of the most famous lookahead techniques. Deterministic methods require a value for each unknown input variable (i.e., a prediction or estimation). Stochastic scheduling methods requires that the inputs are given in the shape of quantiles, interval probabilities, scenarios, or others. Some of these methods, according to Franc et al. (2021), are the open loop feedback control (OLFC, Probabilistic Power Forecasting in a Microgrid for the Integration of Distributed Energy Resources Antonio Parejo* Stefano Bracco ** Sebastián García* Enrique Personal* Juan I. Guerrero* Carlos León* * Department of Electronic Technology, Escuela Politécnica Superior, Universidad de Sevilla, Sevilla, 41011, Spain (e-mail: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]). ** Department of Electrical, Electronic, Telecommunication Engineering and Naval Architecture, University of Genoa, Genova, 16145, Italy (e-mail: [email protected]) Abstract: The integration of distributed energy resources in the power grid supposes a great challenge, considering the expected growing in the electric mobility and the renewable generation. Therefore, to support the successful deployment of these resources, energy management according to the expected behaviors of the elements of power grids is required. In this sense, the probabilistic forecasting methods are gaining acceptance, as they provide richer information than deterministic methods thanks to the consideration of expected probabilities. One of their problems is that the application of metrics for choosing the best models is not simple, depending on the objective of the type of probabilistic prediction that must be performed. In this paper, diverse types of forecasting are performed, and two possible approaches for selecting interval forecasting models are compared in a case study centered on the Smart Polygeneration Microgrid (SPM) located on the Savona Campus of University of Genova (Italy). The results show that, for the machine learning models that were tested, it is better to obtain a model for each of the different types of forecasting that are required (quantiles, and the prediction intervals) to have better-quality predictions. The average improvement on the quality of the tested prediction intervals for the three forecasted variables were a 27.40%, 7.39%, and 28.65% in their Winkler score respectively. Keywords: Stochastic forecasting, probabilistic forecasting, microgrids, distributed energy resources, demand. 1. INTRODUCTION Microgrids are bound to be one of the key elements for the future of the power system. A microgrid can be defined as a small electric network that is designed for a reliable and massive integration of distributed generations (DG) at the low (or medium) voltage level, especially the renewables (solar, wind, and other low carbon technologies), as done by Gouveia et al. (2013) and Harmouch et al. (2018) and even hydrogen, as Ennassiri et al. (2023) say. Regarding its operation modes, as it is states by Ma & Ma (2018), a microgrid system normally operates in two modes, which are called “grid-connected mode” and “island mode”. Under the grid-connected mode, the microgrid system adjusts dispatchable units according to different purposes (e.g., reducing economic costs or to maintain electrical values between feasible bounds). Beyond the power needs, some microgrids could be equipped with power quality monitoring, being able to disconnect themselves from the main grid if the power quality is below a certain level and preventing in this way possible failures and technical problems due to instabilities from the main grid, as Ma & Ma (2018) say. For taking advantage from the microgrid assets and controllability, it is required to perform a scheduling for the operation of the microgrid. This means to decide which generation units should be connected and disconnected, when it is convenient to charge a storage system, or even when the microgrid should swap to islanding mode and disconnect from the main grid. This process is usually referred to as microgrid scheduling, or microgrid optimization. Two main general types of microgrid optimization methods can be found in the literature taking into account the nature of their uncertainties data, i.e., the forecast of the power variables of interest that have to be considered to perform the optimization (expected load consumption, expected generation, etc.). These can be deterministic or probabilistic (sometimes also called stochastic). Deterministic scheduling methods, also called deterministic Model Predictive Control (deterministic MPC) methods by some authors as Franc et al. (2021) are some of the most famous lookahead techniques. Deterministic methods require a value for each unknown input variable (i.e., a prediction or estimation). Stochastic scheduling methods requires that the inputs are given in the shape of quantiles, interval probabilities, scenarios, or others. Some of these methods, according to Franc et al. (2021), are the open loop feedback control (OLFC, Probabilistic Power Forecasting in a Microgrid for the Integration of Distributed Energy Resources Antonio Parejo* Stefano Bracco ** Sebastián García* Enrique Personal* Juan I. Guerrero* Carlos León* * Department of Electronic Technology, Escuela Politécnica Superior, Universidad de Sevilla, Sevilla, 41011, Spain (e-mail: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]). ** Department of Electrical, Electronic, Telecommunication Engineering and Naval Architecture, University of Genoa, Genova, 16145, Italy (e-mail: [email protected]) Abstract: The integration of distributed energy resources in the power grid supposes a great challenge, considering the expected growing in the electric mobility and the renewable generation. Therefore, to support the successful deployment of these resources, energy management according to the expected behaviors of the elements of power grids is required. In this sense, the probabilistic forecasting methods are gaining acceptance, as they provide richer information than deterministic methods thanks to the consideration of expected probabilities. One of their problems is that the application of metrics for choosing the best models is not simple, depending on the objective of the type of probabilistic prediction that must be performed. In this paper, diverse types of forecasting are performed, and two possible approaches for selecting interval forecasting models are compared in a case study centered on the Smart Polygeneration Microgrid (SPM) located on the Savona Campus of University of Genova (Italy). The results show that, for the machine learning models that were tested, it is better to obtain a model for each of the different types of forecasting that are required (quantiles, and the prediction intervals) to have better-quality predictions. The average improvement on the quality of the tested prediction intervals for the three forecasted variables were a 27.40%, 7.39%, and 28.65% in their Winkler score respectively. Keywords: Stochastic forecasting, probabilistic forecasting, microgrids, distributed energy resources, demand. 1. INTRODUCTION Microgrids are bound to be one of the key elements for the future of the power system. A microgrid can be defined as a small electric network that is designed for a reliable and massive integration of distributed generations (DG) at the low (or medium) voltage level, especially the renewables (solar, wind, and other low carbon technologies), as done by Gouveia et al. (2013) and Harmouch et al. (2018) and even hydrogen, as Ennassiri et al. (2023) say. Regarding its operation modes, as it is states by Ma & Ma (2018), a microgrid system normally operates in two modes, which are called “grid-connected mode” and “island mode”. Under the grid-connected mode, the microgrid system adjusts dispatchable units according to different purposes (e.g., reducing economic costs or to maintain electrical values between feasible bounds). Beyond the power needs, some microgrids could be equipped with power quality monitoring, being able to disconnect themselves from the main grid if the power quality is below a certain level and preventing in this way possible failures and technical problems due to instabilities from the main grid, as Ma & Ma (2018) say. For taking advantage from the microgrid assets and controllability, it is required to perform a scheduling for the operation of the microgrid. This means to decide which generation units should be connected and disconnected, when it is convenient to charge a storage system, or even when the microgrid should swap to islanding mode and disconnect from the main grid. This process is usually referred to as microgrid scheduling, or microgrid optimization. Two main general types of microgrid optimization methods can be found in the literature taking into account the nature of their uncertainties data, i.e., the forecast of the power variables of interest that have to be considered to perform the optimization (expected load consumption, expected generation, etc.). These can be deterministic or probabilistic (sometimes also called stochastic). Deterministic scheduling methods, also called deterministic Model Predictive Control (deterministic MPC) methods by some authors as Franc et al. (2021) are some of the most famous lookahead techniques. Deterministic methods require a value for each unknown input variable (i.e., a prediction or estimation). Stochastic scheduling methods requires that the inputs are given in the shape of quantiles, interval probabilities, scenarios, or others. Some of these methods, according to Franc et al. (2021), are the open loop feedback control (OLFC, Probabilistic Power Forecasting in a Microgrid for the Integration of Distributed Energy Resources Antonio Parejo* Stefano Bracco ** Sebastián García* Enrique Personal* Juan I. Guerrero* Carlos León* * Department of Electronic Technology, Escuela Politécnica Superior, Universidad de Sevilla, Sevilla, 41011, Spain (e-mail: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]). ** Department of Electrical, Electronic, Telecommunication Engineering and Naval Architecture, University of Genoa, Genova, 16145, Italy (e-mail: [email protected]) Abstract: The integration of distributed energy resources in the power grid supposes a great challenge, considering the expected growing in the electric mobility and the renewable generation. Therefore, to support the successful deployment of these resources, energy management according to the expected behaviors of the elements of power grids is required. In this sense, the probabilistic forecasting methods are gaining acceptance, as they provide richer information than deterministic methods thanks to the consideration of expected probabilities. One of their problems is that the application of metrics for choosing the best models is not simple, depending on the objective of the type of probabilistic prediction that must be performed. In this paper, diverse types of forecasting are performed, and two possible approaches for selecting interval forecasting models are compared in a case study centered on the Smart Polygeneration Microgrid (SPM) located on the Savona Campus of University of Genova (Italy). The results show that, for the machine learning models that were tested, it is better to obtain a model for each of the different types of forecasting that are required (quantiles, and the prediction intervals) to have better-quality predictions. The average improvement on the quality of the tested prediction intervals for the three forecasted variables were a 27.40%, 7.39%, and 28.65% in their Winkler score respectively. Keywords: Stochastic forecasting, probabilistic forecasting, microgrids, distributed energy resources, demand. 1. INTRODUCTION Microgrids are bound to be one of the key elements for the future of the power system. A microgrid can be defined as a small electric network that is designed for a reliable and massive integration of distributed generations (DG) at the low (or medium) voltage level, especially the renewables (solar, wind, and other low carbon technologies), as done by Gouveia et al. (2013) and Harmouch et al. (2018) and even hydrogen, as Ennassiri et al. (2023) say. Regarding its operation modes, as it is states by Ma & Ma (2018), a microgrid system normally operates in two modes, which are called “grid-connected mode” and “island mode”. Under the grid-connected mode, the microgrid system adjusts dispatchable units according to different purposes (e.g., reducing economic costs or to maintain electrical values between feasible bounds). Beyond the power needs, some microgrids could be equipped with power quality monitoring, being able to disconnect themselves from the main grid if the power quality is below a certain level and preventing in this way possible failures and technical problems due to instabilities from the main grid, as Ma & Ma (2018) say. For taking advantage from the microgrid assets and controllability, it is required to perform a scheduling for the operation of the microgrid. This means to decide which generation units should be connected and disconnected, when it is convenient to charge a storage system, or even when the microgrid should swap to islanding mode and disconnect from the main grid. This process is usually referred to as microgrid scheduling, or microgrid optimization. Two main general types of microgrid optimization methods can be found in the literature taking into account the nature of their uncertainties data, i.e., the forecast of the power variables of interest that have to be considered to perform the optimization (expected load consumption, expected generation, etc.). These can be deterministic or probabilistic (sometimes also called stochastic). Deterministic scheduling methods, also called deterministic Model Predictive Control (deterministic MPC) methods by some authors as Franc et al. (2021) are some of the most famous lookahead techniques. Deterministic methods require a value for each unknown input variable (i.e., a prediction or estimation). Stochastic scheduling methods requires that the inputs are given in the shape of quantiles, interval probabilities, scenarios, or others. Some of these methods, according to Franc et al. (2021), are the open loop feedback control (OLFC, Copyright © 2024 The Authors. This is an open access article under the CC BY-NC-ND license ( https://creativecommons.org/licenses/by-nc-nd/4.0/ ) Antonio Parejo et al. / IFAC PapersOnLine 58-2 (2024) 150–155 151 Probabilistic Power Forecasting in a Microgrid for the Integration of Distributed Energy Resources Antonio Parejo* Stefano Bracco ** Sebastián García* Enrique Personal* Juan I. Guerrero* Carlos León* * Department of Electronic Technology, Escuela Politécnica Superior, Universidad de Sevilla, Sevilla, 41011, Spain (e-mail: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]). ** Department of Electrical, Electronic, Telecommunication Engineering and Naval Architecture, University of Genoa, Genova, 16145, Italy (e-mail: [email protected]) Abstract: The integration of distributed energy resources in the power grid supposes a great challenge, considering the expected growing in the electric mobility and the renewable generation. Therefore, to support the successful deployment of these resources, energy management according to the expected behaviors of the elements of power grids is required. In this sense, the probabilistic forecasting methods are gaining acceptance, as they provide richer information than deterministic methods thanks to the consideration of expected probabilities. One of their problems is that the application of metrics for choosing the best models is not simple, depending on the objective of the type of probabilistic prediction that must be performed. In this paper, diverse types of forecasting are performed, and two possible approaches for selecting interval forecasting models are compared in a case study centered on the Smart Polygeneration Microgrid (SPM) located on the Savona Campus of University of Genova (Italy). The results show that, for the machine learning models that were tested, it is better to obtain a model for each of the different types of forecasting that are required (quantiles, and the prediction intervals) to have better-quality predictions. The average improvement on the quality of the tested prediction intervals for the three forecasted variables were a 27.40%, 7.39%, and 28.65% in their Winkler score respectively. Keywords: Stochastic forecasting, probabilistic forecasting, microgrids, distributed energy resources, demand. 1. INTRODUCTION Microgrids are bound to be one of the key elements for the future of the power system. A microgrid can be defined as a small electric network that is designed for a reliable and massive integration of distributed generations (DG) at the low (or medium) voltage level, especially the renewables (solar, wind, and other low carbon technologies), as done by Gouveia et al. (2013) and Harmouch et al. (2018) and even hydrogen, as Ennassiri et al. (2023) say. Regarding its operation modes, as it is states by Ma & Ma (2018), a microgrid system normally operates in two modes, which are called “grid-connected mode” and “island mode”. Under the grid-connected mode, the microgrid system adjusts dispatchable units according to different purposes (e.g., reducing economic costs or to maintain electrical values between feasible bounds). Beyond the power needs, some microgrids could be equipped with power quality monitoring, being able to disconnect themselves from the main grid if the power quality is below a certain level and preventing in this way possible failures and technical problems due to instabilities from the main grid, as Ma & Ma (2018) say. For taking advantage from the microgrid assets and controllability, it is required to perform a scheduling for the operation of the microgrid. This means to decide which generation units should be connected and disconnected, when it is convenient to charge a storage system, or even when the microgrid should swap to islanding mode and disconnect from the main grid. This process is usually referred to as microgrid scheduling, or microgrid optimization. Two main general types of microgrid optimization methods can be found in the literature taking into account the nature of their uncertainties data, i.e., the forecast of the power variables of interest that have to be considered to perform the optimization (expected load consumption, expected generation, etc.). These can be deterministic or probabilistic (sometimes also called stochastic). Deterministic scheduling methods, also called deterministic Model Predictive Control (deterministic MPC) methods by some authors as Franc et al. (2021) are some of the most famous lookahead techniques. Deterministic methods require a value for each unknown input variable (i.e., a prediction or estimation). Stochastic scheduling methods requires that the inputs are given in the shape of quantiles, interval probabilities, scenarios, or others. Some of these methods, according to Franc et al. (2021), are the open loop feedback control (OLFC, Probabilistic Power Forecasting in a Microgrid for the Integration of Distributed Energy Resources Antonio Parejo* Stefano Bracco ** Sebastián García* Enrique Personal* Juan I. Guerrero* Carlos León* * Department of Electronic Technology, Escuela Politécnica Superior, Universidad de Sevilla, Sevilla, 41011, Spain (e-mail: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]). ** Department of Electrical, Electronic, Telecommunication Engineering and Naval Architecture, University of Genoa, Genova, 16145, Italy (e-mail: [email protected]) Abstract: The integration of distributed energy resources in the power grid supposes a great challenge, considering the expected growing in the electric mobility and the renewable generation. Therefore, to support the successful deployment of these resources, energy management according to the expected behaviors of the elements of power grids is required. In this sense, the probabilistic forecasting methods are gaining acceptance, as they provide richer information than deterministic methods thanks to the consideration of expected probabilities. One of their problems is that the application of metrics for choosing the best models is not simple, depending on the objective of the type of probabilistic prediction that must be performed. In this paper, diverse types of forecasting are performed, and two possible approaches for selecting interval forecasting models are compared in a case study centered on the Smart Polygeneration Microgrid (SPM) located on the Savona Campus of University of Genova (Italy). The results show that, for the machine learning models that were tested, it is better to obtain a model for each of the different types of forecasting that are required (quantiles, and the prediction intervals) to have better-quality predictions. The average improvement on the quality of the tested prediction intervals for the three forecasted variables were a 27.40%, 7.39%, and 28.65% in their Winkler score respectively. Keywords: Stochastic forecasting, probabilistic forecasting, microgrids, distributed energy resources, demand. 1. INTRODUCTION Microgrids are bound to be one of the key elements for the future of the power system. A microgrid can be defined as a small electric network that is designed for a reliable and massive integration of distributed generations (DG) at the low (or medium) voltage level, especially the renewables (solar, wind, and other low carbon technologies), as done by Gouveia et al. (2013) and Harmouch et al. (2018) and even hydrogen, as Ennassiri et al. (2023) say. Regarding its operation modes, as it is states by Ma & Ma (2018), a microgrid system normally operates in two modes, which are called “grid-connected mode” and “island mode”. Under the grid-connected mode, the microgrid system adjusts dispatchable units according to different purposes (e.g., reducing economic costs or to maintain electrical values between feasible bounds). Beyond the power needs, some microgrids could be equipped with power quality monitoring, being able to disconnect themselves from the main grid if the power quality is below a certain level and preventing in this way possible failures and technical problems due to instabilities from the main grid, as Ma & Ma (2018) say. For taking advantage from the microgrid assets and controllability, it is required to perform a scheduling for the operation of the microgrid. This means to decide which generation units should be connected and disconnected, when it is convenient to charge a storage system, or even when the microgrid should swap to islanding mode and disconnect from the main grid. This process is usually referred to as microgrid scheduling, or microgrid optimization. Two main general types of microgrid optimization methods can be found in the literature taking into account the nature of their uncertainties data, i.e., the forecast of the power variables of interest that have to be considered to perform the optimization (expected load consumption, expected generation, etc.). These can be deterministic or probabilistic (sometimes also called stochastic). Deterministic scheduling methods, also called deterministic Model Predictive Control (deterministic MPC) methods by some authors as Franc et al. (2021) are some of the most famous lookahead techniques. Deterministic methods require a value for each unknown input variable (i.e., a prediction or estimation). Stochastic scheduling methods requires that the inputs are given in the shape of quantiles, interval probabilities, scenarios, or others. Some of these methods, according to Franc et al. (2021), are the open loop feedback control (OLFC, Probabilistic Power Forecasting in a Microgrid for the Integration of Distributed Energy Resources Antonio Parejo* Stefano Bracco ** Sebastián García* Enrique Personal* Juan I. Guerrero* Carlos León* * Department of Electronic Technology, Escuela Politécnica Superior, Universidad de Sevilla, Sevilla, 41011, Spain (e-mail: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]). ** Department of Electrical, Electronic, Telecommunication Engineering and Naval Architecture, University of Genoa, Genova, 16145, Italy (e-mail: [email protected]) Abstract: The integration of distributed energy resources in the power grid supposes a great challenge, considering the expected growing in the electric mobility and the renewable generation. Therefore, to support the successful deployment of these resources, energy management according to the expected behaviors of the elements of power grids is required. In this sense, the probabilistic forecasting methods are gaining acceptance, as they provide richer information than deterministic methods thanks to the consideration of expected probabilities. One of their problems is that the application of metrics for choosing the best models is not simple, depending on the objective of the type of probabilistic prediction that must be performed. In this paper, diverse types of forecasting are performed, and two possible approaches for selecting interval forecasting models are compared in a case study centered on the Smart Polygeneration Microgrid (SPM) located on the Savona Campus of University of Genova (Italy). The results show that, for the machine learning models that were tested, it is better to obtain a model for each of the different types of forecasting that are required (quantiles, and the prediction intervals) to have better-quality predictions. The average improvement on the quality of the tested prediction intervals for the three forecasted variables were a 27.40%, 7.39%, and 28.65% in their Winkler score respectively. Keywords: Stochastic forecasting, probabilistic forecasting, microgrids, distributed energy resources, demand. 1. INTRODUCTION Microgrids are bound to be one of the key elements for the future of the power system. A microgrid can be defined as a small electric network that is designed for a reliable and massive integration of distributed generations (DG) at the low (or medium) voltage level, especially the renewables (solar, wind, and other low carbon technologies), as done by Gouveia et al. (2013) and Harmouch et al. (2018) and even hydrogen, as Ennassiri et al. (2023) say. Regarding its operation modes, as it is states by Ma & Ma (2018), a microgrid system normally operates in two modes, which are called “grid-connected mode” and “island mode”. Under the grid-connected mode, the microgrid system adjusts dispatchable units according to different purposes (e.g., reducing economic costs or to maintain electrical values between feasible bounds). Beyond the power needs, some microgrids could be equipped with power quality monitoring, being able to disconnect themselves from the main grid if the power quality is below a certain level and preventing in this way possible failures and technical problems due to instabilities from the main grid, as Ma & Ma (2018) say. For taking advantage from the microgrid assets and controllability, it is required to perform a scheduling for the operation of the microgrid. This means to decide which generation units should be connected and disconnected, when it is convenient to charge a storage system, or even when the microgrid should swap to islanding mode and disconnect from the main grid. This process is usually referred to as microgrid scheduling, or microgrid optimization. Two main general types of microgrid optimization methods can be found in the literature taking into account the nature of their uncertainties data, i.e., the forecast of the power variables of interest that have to be considered to perform the optimization (expected load consumption, expected generation, etc.). These can be deterministic or probabilistic (sometimes also called stochastic). Deterministic scheduling methods, also called deterministic Model Predictive Control (deterministic MPC) methods by some authors as Franc et al. (2021) are some of the most famous lookahead techniques. Deterministic methods require a value for each unknown input variable (i.e., a prediction or estimation). Stochastic scheduling methods requires that the inputs are given in the shape of quantiles, interval probabilities, scenarios, or others. Some of these methods, according to Franc et al. (2021), are the open loop feedback control (OLFC, Probabilistic Power Forecasting in a Microgrid for the Integration of Distributed Energy Resources Antonio Parejo* Stefano Bracco ** Sebastián García* Enrique Personal* Juan I. Guerrero* Carlos León* * Department of Electronic Technology, Escuela Politécnica Superior, Universidad de Sevilla, Sevilla, 41011, Spain (e-mail: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]). ** Department of Electrical, Electronic, Telecommunication Engineering and Naval Architecture, University of Genoa, Genova, 16145, Italy (e-mail: [email protected]) Abstract: The integration of distributed energy resources in the power grid supposes a great challenge, considering the expected growing in the electric mobility and the renewable generation. Therefore, to support the successful deployment of these resources, energy management according to the expected behaviors of the elements of power grids is required. In this sense, the probabilistic forecasting methods are gaining acceptance, as they provide richer information than deterministic methods thanks to the consideration of expected probabilities. One of their problems is that the application of metrics for choosing the best models is not simple, depending on the objective of the type of probabilistic prediction that must be performed. In this paper, diverse types of forecasting are performed, and two possible approaches for selecting interval forecasting models are compared in a case study centered on the Smart Polygeneration Microgrid (SPM) located on the Savona Campus of University of Genova (Italy). The results show that, for the machine learning models that were tested, it is better to obtain a model for each of the different types of forecasting that are required (quantiles, and the prediction intervals) to have better-quality predictions. The average improvement on the quality of the tested prediction intervals for the three forecasted variables were a 27.40%, 7.39%, and 28.65% in their Winkler score respectively. Keywords: Stochastic forecasting, probabilistic forecasting, microgrids, distributed energy resources, demand. 1. INTRODUCTION Microgrids are bound to be one of the key elements for the future of the power system. A microgrid can be defined as a small electric network that is designed for a reliable and massive integration of distributed generations (DG) at the low (or medium) voltage level, especially the renewables (solar, wind, and other low carbon technologies), as done by Gouveia et al. (2013) and Harmouch et al. (2018) and even hydrogen, as Ennassiri et al. (2023) say. Regarding its operation modes, as it is states by Ma & Ma (2018), a microgrid system normally operates in two modes, which are called “grid-connected mode” and “island mode”. Under the grid-connected mode, the microgrid system adjusts dispatchable units according to different purposes (e.g., reducing economic costs or to maintain electrical values between feasible bounds). Beyond the power needs, some microgrids could be equipped with power quality monitoring, being able to disconnect themselves from the main grid if the power quality is below a certain level and preventing in this way possible failures and technical problems due to instabilities from the main grid, as Ma & Ma (2018) say. For taking advantage from the microgrid assets and controllability, it is required to perform a scheduling for the operation of the microgrid. This means to decide which generation units should be connected and disconnected, when it is convenient to charge a storage system, or even when the microgrid should swap to islanding mode and disconnect from the main grid. This process is usually referred to as microgrid scheduling, or microgrid optimization. Two main general types of microgrid optimization methods can be found in the literature taking into account the nature of their uncertainties data, i.e., the forecast of the power variables of interest that have to be considered to perform the optimization (expected load consumption, expected generation, etc.). These can be deterministic or probabilistic (sometimes also called stochastic). Deterministic scheduling methods, also called deterministic Model Predictive Control (deterministic MPC) methods by some authors as Franc et al. (2021) are some of the most famous lookahead techniques. Deterministic methods require a value for each unknown input variable (i.e., a prediction or estimation). Stochastic scheduling methods requires that the inputs are given in the shape of quantiles, interval probabilities, scenarios, or others. Some of these methods, according to Franc et al. (2021), are the open loop feedback control (OLFC, Probabilistic Power Forecasting in a Microgrid for the Integration of Distributed Energy Resources Antonio Parejo* Stefano Bracco ** Sebastián García* Enrique Personal* Juan I. Guerrero* Carlos León* * Department of Electronic Technology, Escuela Politécnica Superior, Universidad de Sevilla, Sevilla, 41011, Spain (e-mail: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]). ** Department of Electrical, Electronic, Telecommunication Engineering and Naval Architecture, University of Genoa, Genova, 16145, Italy (e-mail: [email protected]) Abstract: The integration of distributed energy resources in the power grid supposes a great challenge, considering the expected growing in the electric mobility and the renewable generation. Therefore, to support the successful deployment of these resources, energy management according to the expected behaviors of the elements of power grids is required. In this sense, the probabilistic forecasting methods are gaining acceptance, as they provide richer information than deterministic methods thanks to the consideration of expected probabilities. One of their problems is that the application of metrics for choosing the best models is not simple, depending on the objective of the type of probabilistic prediction that must be performed. In this paper, diverse types of forecasting are performed, and two possible approaches for selecting interval forecasting models are compared in a case study centered on the Smart Polygeneration Microgrid (SPM) located on the Savona Campus of University of Genova (Italy). The results show that, for the machine learning models that were tested, it is better to obtain a model for each of the different types of forecasting that are required (quantiles, and the prediction intervals) to have better-quality predictions. The average improvement on the quality of the tested prediction intervals for the three forecasted variables were a 27.40%, 7.39%, and 28.65% in their Winkler score respectively. Keywords: Stochastic forecasting, probabilistic forecasting, microgrids, distributed energy resources, demand. 1. INTRODUCTION Microgrids are bound to be one of the key elements for the future of the power system. A microgrid can be defined as a small electric network that is designed for a reliable and massive integration of distributed generations (DG) at the low (or medium) voltage level, especially the renewables (solar, wind, and other low carbon technologies), as done by Gouveia et al. (2013) and Harmouch et al. (2018) and even hydrogen, as Ennassiri et al. (2023) say. Regarding its operation modes, as it is states by Ma & Ma (2018), a microgrid system normally operates in two modes, which are called “grid-connected mode” and “island mode”. Under the grid-connected mode, the microgrid system adjusts dispatchable units according to different purposes (e.g., reducing economic costs or to maintain electrical values between feasible bounds). Beyond the power needs, some microgrids could be equipped with power quality monitoring, being able to disconnect themselves from the main grid if the power quality is below a certain level and preventing in this way possible failures and technical problems due to instabilities from the main grid, as Ma & Ma (2018) say. For taking advantage from the microgrid assets and controllability, it is required to perform a scheduling for the operation of the microgrid. This means to decide which generation units should be connected and disconnected, when it is convenient to charge a storage system, or even when the microgrid should swap to islanding mode and disconnect from the main grid. This process is usually referred to as microgrid scheduling, or microgrid optimization. Two main general types of microgrid optimization methods can be found in the literature taking into account the nature of their uncertainties data, i.e., the forecast of the power variables of interest that have to be considered to perform the optimization (expected load consumption, expected generation, etc.). These can be deterministic or probabilistic (sometimes also called stochastic). Deterministic scheduling methods, also called deterministic Model Predictive Control (deterministic MPC) methods by some authors as Franc et al. (2021) are some of the most famous lookahead techniques. Deterministic methods require a value for each unknown input variable (i.e., a prediction or estimation). Stochastic scheduling methods requires that the inputs are given in the shape of quantiles, interval probabilities, scenarios, or others. Some of these methods, according to Franc et al. (2021), are the open loop feedback control (OLFC, Probabilistic Power Forecasting in a Microgrid for the Integration of Distributed Energy Resources Antonio Parejo* Stefano Bracco ** Sebastián García* Enrique Personal* Juan I. Guerrero* Carlos León* * Department of Electronic Technology, Escuela Politécnica Superior, Universidad de Sevilla, Sevilla, 41011, Spain (e-mail: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]). ** Department of Electrical, Electronic, Telecommunication Engineering and Naval Architecture, University of Genoa, Genova, 16145, Italy (e-mail: [email protected]) Abstract: The integration of distributed energy resources in the power grid supposes a great challenge, considering the expected growing in the electric mobility and the renewable generation. Therefore, to support the successful deployment of these resources, energy management according to the expected behaviors of the elements of power grids is required. In this sense, the probabilistic forecasting methods are gaining acceptance, as they provide richer information than deterministic methods thanks to the consideration of expected probabilities. One of their problems is that the application of metrics for choosing the best models is not simple, depending on the objective of the type of probabilistic prediction that must be performed. In this paper, diverse types of forecasting are performed, and two possible approaches for selecting interval forecasting models are compared in a case study centered on the Smart Polygeneration Microgrid (SPM) located on the Savona Campus of University of Genova (Italy). The results show that, for the machine learning models that were tested, it is better to obtain a model for each of the different types of forecasting that are required (quantiles, and the prediction intervals) to have better-quality predictions. The average improvement on the quality of the tested prediction intervals for the three forecasted variables were a 27.40%, 7.39%, and 28.65% in their Winkler score respectively. Keywords: Stochastic forecasting, probabilistic forecasting, microgrids, distributed energy resources, demand. 1. INTRODUCTION Microgrids are bound to be one of the key elements for the future of the power system. A microgrid can be defined as a small electric network that is designed for a reliable and massive integration of distributed generations (DG) at the low (or medium) voltage level, especially the renewables (solar, wind, and other low carbon technologies), as done by Gouveia et al. (2013) and Harmouch et al. (2018) and even hydrogen, as Ennassiri et al. (2023) say. Regarding its operation modes, as it is states by Ma & Ma (2018), a microgrid system normally operates in two modes, which are called “grid-connected mode” and “island mode”. Under the grid-connected mode, the microgrid system adjusts dispatchable units according to different purposes (e.g., reducing economic costs or to maintain electrical values between feasible bounds). Beyond the power needs, some microgrids could be equipped with power quality monitoring, being able to disconnect themselves from the main grid if the power quality is below a certain level and preventing in this way possible failures and technical problems due to instabilities from the main grid, as Ma & Ma (2018) say. For taking advantage from the microgrid assets and controllability, it is required to perform a scheduling for the operation of the microgrid. This means to decide which generation units should be connected and disconnected, when it is convenient to charge a storage system, or even when the microgrid should swap to islanding mode and disconnect from the main grid. This process is usually referred to as microgrid scheduling, or microgrid optimization. Two main general types of microgrid optimization methods can be found in the literature taking into account the nature of their uncertainties data, i.e., the forecast of the power variables of interest that have to be considered to perform the optimization (expected load consumption, expected generation, etc.). These can be deterministic or probabilistic (sometimes also called stochastic). Deterministic scheduling methods, also called deterministic Model Predictive Control (deterministic MPC) methods by some authors as Franc et al. (2021) are some of the most famous lookahead techniques. Deterministic methods require a value for each unknown input variable (i.e., a prediction or estimation). Stochastic scheduling methods requires that the inputs are given in the shape of quantiles, interval probabilities, scenarios, or others. Some of these methods, according to Franc et al. (2021), are the open loop feedback control (OLFC, Probabilistic Power Forecasting in a Microgrid for the Integration of Distributed Energy Resources Antonio Parejo* Stefano Bracco ** Sebastián García* Enrique Personal* Juan I. Guerrero* Carlos León* * Department of Electronic Technology, Escuela Politécnica Superior, Universidad de Sevilla, Sevilla, 41011, Spain (e-mail: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]). ** Department of Electrical, Electronic, Telecommunication Engineering and Naval Architecture, University of Genoa, Genova, 16145, Italy (e-mail: [email protected]) Abstract: The integration of distributed energy resources in the power grid supposes a great challenge, considering the expected growing in the electric mobility and the renewable generation. Therefore, to support the successful deployment of these resources, energy management according to the expected behaviors of the elements of power grids is required. In this sense, the probabilistic forecasting methods are gaining acceptance, as they provide richer information than deterministic methods thanks to the consideration of expected probabilities. One of their problems is that the application of metrics for choosing the best models is not simple, depending on the objective of the type of probabilistic prediction that must be performed. In this paper, diverse types of forecasting are performed, and two possible approaches for selecting interval forecasting models are compared in a case study centered on the Smart Polygeneration Microgrid (SPM) located on the Savona Campus of University of Genova (Italy). The results show that, for the machine learning models that were tested, it is better to obtain a model for each of the different types of forecasting that are required (quantiles, and the prediction intervals) to have better-quality predictions. The average improvement on the quality of the tested prediction intervals for the three forecasted variables were a 27.40%, 7.39%, and 28.65% in their Winkler score respectively. Keywords: Stochastic forecasting, probabilistic forecasting, microgrids, distributed energy resources, demand. 1. INTRODUCTION Microgrids are bound to be one of the key elements for the future of the power system. A microgrid can be defined as a small electric network that is designed for a reliable and massive integration of distributed generations (DG) at the low (or medium) voltage level, especially the renewables (solar, wind, and other low carbon technologies), as done by Gouveia et al. (2013) and Harmouch et al. (2018) and even hydrogen, as Ennassiri et al. (2023) say. Regarding its operation modes, as it is states by Ma & Ma (2018), a microgrid system normally operates in two modes, which are called “grid-connected mode” and “island mode”. Under the grid-connected mode, the microgrid system adjusts dispatchable units according to different purposes (e.g., reducing economic costs or to maintain electrical values between feasible bounds). Beyond the power needs, some microgrids could be equipped with power quality monitoring, being able to disconnect themselves from the main grid if the power quality is below a certain level and preventing in this way possible failures and technical problems due to instabilities from the main grid, as Ma & Ma (2018) say. For taking advantage from the microgrid assets and controllability, it is required to perform a scheduling for the operation of the microgrid. This means to decide which generation units should be connected and disconnected, when it is convenient to charge a storage system, or even when the microgrid should swap to islanding mode and disconnect from the main grid. This process is usually referred to as microgrid scheduling, or microgrid optimization. Two main general types of microgrid optimization methods can be found in the literature taking into account the nature of their uncertainties data, i.e., the forecast of the power variables of interest that have to be considered to perform the optimization (expected load consumption, expected generation, etc.). These can be deterministic or probabilistic (sometimes also called stochastic). Deterministic scheduling methods, also called deterministic Model Predictive Control (deterministic MPC) methods by some authors as Franc et al. (2021) are some of the most famous lookahead techniques. Deterministic methods require a value for each unknown input variable (i.e., a prediction or estimation). Stochastic scheduling methods requires that the inputs are given in the shape of quantiles, interval probabilities, scenarios, or others. Some of these methods, according to Franc et al. (2021), are the open loop feedback control (OLFC, Copyright © 2024 The Authors. This is an open access article under the CC BY-NC-ND license ( https://creativecommons.org/licenses/by-nc-nd/4.0/ ) sometimes referred to as stochastic MPC), which suppose a modification on the deterministic MPC to allow the introduction of stochastic uncertainties. Another one is the stochastic dynamic programming (SDP), which apply (as it is stated in its name) dynamic programming for solving the problem of optimization considering multiple scenarios with their inherent probabilities of occurrence, as Franc et al. (2021) do. The optimization of electrical systems that consider electric mobility can also be found in the literature. Aicardi et al. (2023) perform the assignation of charging stations for vehicles by means of the stochastic user equilibrium conditions. The optimal location for charging stations for buses is done by Ferro et al. (2023). A method for optimizing the management of fleets of electric vehicles is proposed by Fresia & Bracco (2023), where the estimated generation of plants and the expected building load are considered as inputs for the optimization. From the perspective of the distribution grid management, an optimization method for including electric vehicles as resources that participate via aggregators is proposed by Valderrama et al. (2023). The impact of the electric mobility on the optimal design of energy systems is evaluated by Piazza et al. (2023). Once established the types of optimization methods to be applied in microgrids and energy systems, it is necessary to review some aspects of the state of the art on deterministic and probabilistic forecasting. As it can be appreciated, a deterministic forecasting model provides less information about the variable, while the probabilistic models give much more detail about the expected values and their characteristics. However, evaluating a probabilistic forecast involves a totally different process than in the case of deterministic forecast. This is due to the non-unique nature of the probabilistic forecasting, being their results not a single point, but various of them or even some intervals, having each one an associated probability value. Some metrics has been developed for evaluating the performance, which are the Pinball Loss Function and the Winkler Score, as Hong & Fan (2016) indicate. For the formulation of a probabilistic forecasting problem, four approaches are distinguished by Nowotarski & Weron (2018), which are historical simulation (or empirical/sample prediction intervals, PIs), distribution-based probabilistic forecast, bootstrapped PIs, and Quantile Regression Averaging (QRA). Other methods that are commonly found in the literature are Quantile Regression, bootstrapped quantiles, and scenario sets. Some of the most popular techniques when performing forecasting in the power sector are machine learning methods, such as neural networks, random forests, and transformers, which are applied by Domínguez-Cid et al. (2023). Forecasting techniques are applied to diverse types of loads and power demands, for example Zhou et al. (2024) apply them in industrial environments and Al-Rajab & Loucif (2024) apply them in residential buildings. As it can be seen, in order to improve the energy management of microgrids, the probabilistic forecasting and stochastic optimization are powerful tools that can lead to a more efficient power system. For example, Yang (2020) predicts solar generation power by means of nonparametric predictive distributions and prediction intervals. In the case of Mazhari et al. (2019), they apply forecasting techniques for predicting unstable groups of generators inside the power system. To do so, they use quantile regression techniques for obtaining prediction intervals, and then they use such intervals to obtain empirical cumulative distribution functions. In the referred literature, while some authors perform prediction of quantiles and prediction intervals (with their associated probability), when intervals are obtained from distributions, they do not study the effect of the evaluation metrics in the quantiles and intervals and their relationship to improve the forecasting quality. Considering this, the proposal of this paper is to apply a forecasting system over a microgrid and define how to handle the model selection for predicting while considering the evaluation metrics in quantiles and intervals. The present paper is an extension of the work done by Parejo et al. (2021). In that work, only deterministic methods were applied to a microgrid. In the present one, probabilistic methods have been added, and the procedure for model selection is also discussed. Some of the probabilistic methods, and the way of creating prediction intervals from quantile distributions, are described by Parejo et al. (2024), where real data of secondary distribution substations were used to apply these methods. The main contributions of the paper are: • A forecasting case study that includes deterministic and probabilistic methods of diverse variables of a microgrid. Diverse techniques are applied and evaluated using error metrics. • A comparison between two approaches for selecting models for probability intervals, showing which approach provides better results. The organization of this paper is as follows. Section 2 describes the characteristics of the microgrid and the case study that is evaluated. Section 3 presents and discusses the results of the case study. Section 4 exposes the conclusions and outlines the directions for future developments. 2. MATERIALS AND METHODS A deterministic and probabilistic forecasting procedure is performed over real power data of the Savona Campus of the University of Genoa (Italy), where there is a Smart Polygeneration Microgrid (SPM), an innovative facility used to provide electricity and heat to the campus, as Delfino et al. (2019) indicate. 2.1 Datasets from the microgrid In this study, three electric power variables of the microgrid are forecasted. These correspond to the demand of the campus (variable ge2), the load consumption of the main buildings inside the campus (variable lc), and PV generation (variable pv). These variables have been chosen because they are 152 Antonio Parejo et al. / IFAC PapersOnLine 58-2 (2024) 150–155 directly related to the energy management of the campus. In this sense, these variables can be used for optimizing the management of the existing controllable elements, such as EV chargers, microturbines, batteries, etc. Moreover, the global demand of the campus is useful to evaluate the interaction with the distribution network where it is connected. 2.2 Training, validation, and test periods For applying machine learning forecasting models, a procedure of training, validation, and test is applied. Therefore, the available historical data has been divided into smaller periods in which the models are successively retrained, evaluated (for ordering them according to their quality), and then executed (for obtaining the required forecasting) during a certain number of days. After some time, the models are retrained for incorporating new historical data in their training (and the older model instances are discarded). In this sense, that period of time in which the models are trained, evaluated, ordered in a ranking, and then used for predicting (using the best possible model instance among those available) is hereinafter called a "ranking cycle". A period of a whole year (2016) has been taken as the initial data for training and validating the first models (for the cycle 0), and after that the next month (January 2017) was used to perform the testing. Therefore, once a month, new models are trained, and a new ranking is created for them to be executed during the month to perform the forecasting tasks. Then, for the cycle 1, the period from 1st January 2016 to 31st January 2017 was used for training and validating the models, and February 2017 was used to perform the testing. This has been successively repeated until the cycle 23, in which the training period was from 1st January 2016 to 30th November 2018, and the testing period was December 2018. The results that are presented correspond with the average value of the error metrics during the testing period. In summary, the models that are trained and validated in a cycle are used to predict the next month. After that time, a new cycle starts, and all models are retrained. The selection of the training and validation sets for training the models and constructing the ranking for the selection of models to make predictions is as follows. The validation set corresponds to the last 15 days of data of the historical dataset that is available in the system when the modeling process is performed. All data except these 15 days are used for training the models, while the 15 last days are used for validating, i.e., choosing hyperparameters and creating the ordered ranking of models. The main advantage of this approach is that the models are selected based on their performance on the moments near to the present, which can increase the quality of the prediction in the nearby days. However, the models would be less generalist when applied to periods of time that are far away from the period used for validation, so a more frequent model retraining could be required to keep up-to-date models. 2.3 Forecasting techniques The machine learning techniques that are applied for performing deterministic forecasting are multi-layer perceptron regressor (MLPR), Random Forest Regressor (RFR), Autoregressive Integrated Moving Average (ARIMA), and Naive. For probabilistic forecasting, the techniques Quantile Random Forest (QRF) and the Random Forest Regressor (RFR) considering a normal distribution are applied for obtaining quantile distributions. The distributions are then processed to obtain the required prediction intervals using pairs of quantiles, in a similar way than it is done by Parejo et al. (2024). As Khosravi et al. (2013) say, a prediction interval (PI) is composed of an upper and lower bound. The probability that a variable is between these bounds can be described as (1 – α). 3. RESULTS AND DISCUSSION This section exposes the metrics of the forecasting obtained following the aforementioned procedure. The deterministic forecasting is evaluated using RMSE, as done by Parejo et al. (2021), the quantiles by means of Pinball, and the prediction intervals by means of Winkler, as done by Parejo et al. (2024). 3.1 Deterministic forecasting The deterministic forecasting was generated applying RFR and MLPR techniques with different combinations of inputs and selecting the best ones with the validation period. ARIMA and Naive models have also been applied. The results can be seen in Table 1. Table 1. Metrics for deterministic forecasting Variable Average Daily RMSE (kW) Naïve ARIMA ML techniques, Naive, and ARIMA Mean Std Mean Std Mean Std ge2 41.10 29.86 38.13 26.06 26.89 16.36 lc 34.33 31.85 22.54 16.18 18.79 11.38 pv 8.30 6.06 8.65 5.27 6.80 5.88 It can be observed that the inclusion of ML techniques improves the forecasting quality. As an example, the deterministic forecasting that has been obtained for the variable lc for the five days from 12th September 2018 to 16th September 2018 is represented in the Figure 1. Antonio Parejo et al. / IFAC PapersOnLine 58-2 (2024) 150–155 153 directly related to the energy management of the campus. In this sense, these variables can be used for optimizing the management of the existing controllable elements, such as EV chargers, microturbines, batteries, etc. Moreover, the global demand of the campus is useful to evaluate the interaction with the distribution network where it is connected. 2.2 Training, validation, and test periods For applying machine learning forecasting models, a procedure of training, validation, and test is applied. Therefore, the available historical data has been divided into smaller periods in which the models are successively retrained, evaluated (for ordering them according to their quality), and then executed (for obtaining the required forecasting) during a certain number of days. After some time, the models are retrained for incorporating new historical data in their training (and the older model instances are discarded). In this sense, that period of time in which the models are trained, evaluated, ordered in a ranking, and then used for predicting (using the best possible model instance among those available) is hereinafter called a "ranking cycle". A period of a whole year (2016) has been taken as the initial data for training and validating the first models (for the cycle 0), and after that the next month (January 2017) was used to perform the testing. Therefore, once a month, new models are trained, and a new ranking is created for them to be executed during the month to perform the forecasting tasks. Then, for the cycle 1, the period from 1st January 2016 to 31st January 2017 was used for training and validating the models, and February 2017 was used to perform the testing. This has been successively repeated until the cycle 23, in which the training period was from 1st January 2016 to 30th November 2018, and the testing period was December 2018. The results that are presented correspond with the average value of the error metrics during the testing period. In summary, the models that are trained and validated in a cycle are used to predict the next month. After that time, a new cycle starts, and all models are retrained. The selection of the training and validation sets for training the models and constructing the ranking for the selection of models to make predictions is as follows. The validation set corresponds to the last 15 days of data of the historical dataset that is available in the system when the modeling process is performed. All data except these 15 days are used for training the models, while the 15 last days are used for validating, i.e., choosing hyperparameters and creating the ordered ranking of models. The main advantage of this approach is that the models are selected based on their performance on the moments near to the present, which can increase the quality of the prediction in the nearby days. However, the models would be less generalist when applied to periods of time that are far away from the period used for validation, so a more frequent model retraining could be required to keep up-to-date models. 2.3 Forecasting techniques The machine learning techniques that are applied for performing deterministic forecasting are multi-layer perceptron regressor (MLPR), Random Forest Regressor (RFR), Autoregressive Integrated Moving Average (ARIMA), and Naive. For probabilistic forecasting, the techniques Quantile Random Forest (QRF) and the Random Forest Regressor (RFR) considering a normal distribution are applied for obtaining quantile distributions. The distributions are then processed to obtain the required prediction intervals using pairs of quantiles, in a similar way than it is done by Parejo et al. (2024). As Khosravi et al. (2013) say, a prediction interval (PI) is composed of an upper and lower bound. The probability that a variable is between these bounds can be described as (1 – α). 3. RESULTS AND DISCUSSION This section exposes the metrics of the forecasting obtained following the aforementioned procedure. The deterministic forecasting is evaluated using RMSE, as done by Parejo et al. (2021), the quantiles by means of Pinball, and the prediction intervals by means of Winkler, as done by Parejo et al. (2024). 3.1 Deterministic forecasting The deterministic forecasting was generated applying RFR and MLPR techniques with different combinations of inputs and selecting the best ones with the validation period. ARIMA and Naive models have also been applied. The results can be seen in Table 1. Table 1. Metrics for deterministic forecasting Variable Average Daily RMSE (kW) Naïve ARIMA ML techniques, Naive, and ARIMA Mean Std Mean Std Mean Std ge2 41.10 29.86 38.13 26.06 26.89 16.36 lc 34.33 31.85 22.54 16.18 18.79 11.38 pv 8.30 6.06 8.65 5.27 6.80 5.88 It can be observed that the inclusion of ML techniques improves the forecasting quality. As an example, the deterministic forecasting that has been obtained for the variable lc for the five days from 12th September 2018 to 16th September 2018 is represented in the Figure 1. 3.2 Forecasting of quantiles The prediction of quantiles from 1 to 99 has been performed using QRF, and also RFR considering a normal distribution with a standard deviation of 10%, 20%, and 30%. The results can be found in Table 2. Table 2. Metrics for forecasting of quantiles Variable Mean Pinball (kW) ge2 9.99 lc 5.54 pv 1.31 The quantile forecasting (from quantile 1 to 99) that has been obtained for the variable lc for five days is represented in the Figure 2. 3.3 Prediction intervals under two approaches The prediction of intervals through ML techniques can be done in two different ways regarding the selection of models. The first proposed way is by choosing the model that provides the quantiles with the best Pinball, and then using these quantiles for generating all the desired intervals. The second way is evaluating all the available models according to the Winkler (a metric used to evaluate intervals) during the validation phase, and then choose the best model for each of the desired intervals. The first one is simpler, as a single model is used to obtain all the required intervals of a variable. The second one could require different models for each type of interval (according to their associated probability). For comparing the two approaches, the metric "Improvement" is defined, which measures the percentage of improvement (i.e., reduction of the value of the Winkler metric) achieved when the second approach is applied instead of the first one. The global results of the comparison of the two approaches can be seen in Table 3. Table 3. Metrics for interval prediction and comparison between the two proposed approaches Interval probability Variable Winkler Score (kW) Improvement (%) First approach Second approach 98% ge2 121.28 68.90 43.19 lc 73.55 63.76 13.31 pv 17.40 9.27 46.75 94% ge2 105.60 64.88 38.56 lc 62.26 56.00 10.06 pv 15.32 9.04 40.98 90% ge2 96.84 63.49 34.44 lc 56.34 51.08 9.34 pv 14.06 9.05 35.66 80% ge2 84.65 65.42 22.72 lc 48.30 46.17 4.41 pv 12.14 9.40 22.60 70% ge2 79.75 70.51 11.58 lc 45.07 44.77 0.66 pv 11.25 10.18 9.54 60% ge2 80.42 79.55 1.08 lc 45.24 44.54 1.55 pv 11.01 10.98 0.20 Average for the tested intervals ge2 94.76 68.79 27.40 lc 55.13 51.05 7.39 pv 13.53 9.65 28.65 Following the second approach (as it showed a lesser value of the error metrics), the interval forecasting that has been obtained for the variable lc for the five days from 12th September 2018 to 16th September 2018 is represented in the Figure 3. Figure 1. Plot of the deterministic forecasting of the variable lc for five days. Figure 2. Plot of the quantile forecasting (quantiles 1 to 99) of the variable lc for five days. 154 Antonio Parejo et al. / IFAC PapersOnLine 58-2 (2024) 150–155 a) b) c) The comparison between the two approaches for selecting the model to create the prediction intervals shows that the second approach produces a lesser error according to the Winkler metric. The improvement is positive for all the tested intervals and variables. The average value of the improvement for all the intervals in each of the variables has been a 27.40% of improvement of ge2, a 7.39% for lc, and a 28.65% for pv. The relatively low improvement that is achieved in the variable lc could be explained because it is easier to predict (as it is the load of some of the main buildings) that the other two variables. Specially, ge2 (the global demand of the campus) includes the whole consumption of the buildings, and also other elements such as the microturbines and PV generation, which makes more difficult to predict its value (especially due to the power variability caused by the microturbines), and therefore the margin of possible improvement in the prediction is higher. Therefore, from the point of view of forecast quality, it is better to follow the second approach. It means that the selection of models should be done independently for each probability interval of each variable. As a result, it could happen that several models must be trained and stored for a same variable, instead of using a single one, and therefore it means that the computational requirements would be higher. While the higher quality of the obtained forecasting could compensate the extra computational requirements, it should be carefully chosen how many different techniques and intervals are tested to keep the computational load under an affordable level. The evaluation of the computational load for the proposed forecasting procedure constitutes a future task of this research line. 4. CONCLUSIONS AND FUTURE STEPS Forecasting of variables is key for performing the energy management of energy systems and achieve a successful integration of renewable generation sources and electric mobility solutions. The forecasting procedure that has been proposed and applied includes diverse types of probabilistic (and stochastic) models, which are quantile distributions (from 1 to 99) and six intervals of probability (98, 94, 90, 80, 70, and 60%) in a case study centered on the SPM of Savona Campus. To evaluate the performance of the forecasting methods in this case study, the system has been retrained once a month. The last 15 days available were used to validate the models and create the rankings, while the rest of the historical data were firstly used for training the models. The results showed that for most of the forecasted variables (three power measurements of the microgrid), the proposed combination of methods achieved better forecasting results than the naïve method and ARIMA method, which were used as references. Regarding the evaluation of different kinds of probabilistic models, it has been seen that the decision of including independent model rankings with their respective evaluation metrics can produce a global increasing of the quality of forecasting. Specially, this improvement has been observed in the prediction intervals with higher associated probability. Therefore, this approach is more convenient than simply evaluating the global quantile distribution (using pinball) of all types of models and using the best of these to obtain the rest of the probabilistic models (which would be computationally simpler but would achieve a worse quality of forecasting). Considering the obtained results, the next research steps would be applying the forecasting techniques and procedures in a case study where the optimization of the energy management of the campus and its electric mobility charging stations is done by means of stochastic methods (for them to include the probabilistic forecasting information). In this way, it will be possible to evaluate how the quality of the probabilistic forecasting affect to the optimization, and which is the computational load required for their calculation. Figure 3. Plot of the prediction intervals of the variable lc for five days. a) interval 98%; b) interval 90%; c) interval 80%. Antonio Parejo et al. / IFAC PapersOnLine 58-2 (2024) 150–155 155 a) b) c) The comparison between the two approaches for selecting the model to create the prediction intervals shows that the second approach produces a lesser error according to the Winkler metric. The improvement is positive for all the tested intervals and variables. The average value of the improvement for all the intervals in each of the variables has been a 27.40% of improvement of ge2, a 7.39% for lc, and a 28.65% for pv. The relatively low improvement that is achieved in the variable lc could be explained because it is easier to predict (as it is the load of some of the main buildings) that the other two variables. Specially, ge2 (the global demand of the campus) includes the whole consumption of the buildings, and also other elements such as the microturbines and PV generation, which makes more difficult to predict its value (especially due to the power variability caused by the microturbines), and therefore the margin of possible improvement in the prediction is higher. Therefore, from the point of view of forecast quality, it is better to follow the second approach. It means that the selection of models should be done independently for each probability interval of each variable. As a result, it could happen that several models must be trained and stored for a same variable, instead of using a single one, and therefore it means that the computational requirements would be higher. While the higher quality of the obtained forecasting could compensate the extra computational requirements, it should be carefully chosen how many different techniques and intervals are tested to keep the computational load under an affordable level. The evaluation of the computational load for the proposed forecasting procedure constitutes a future task of this research line. 4. CONCLUSIONS AND FUTURE STEPS Forecasting of variables is key for performing the energy management of energy systems and achieve a successful integration of renewable generation sources and electric mobility solutions. The forecasting procedure that has been proposed and applied includes diverse types of probabilistic (and stochastic) models, which are quantile distributions (from 1 to 99) and six intervals of probability (98, 94, 90, 80, 70, and 60%) in a case study centered on the SPM of Savona Campus. To evaluate the performance of the forecasting methods in this case study, the system has been retrained once a month. The last 15 days available were used to validate the models and create the rankings, while the rest of the historical data were firstly used for training the models. The results showed that for most of the forecasted variables (three power measurements of the microgrid), the proposed combination of methods achieved better forecasting results than the naïve method and ARIMA method, which were used as references. Regarding the evaluation of different kinds of probabilistic models, it has been seen that the decision of including independent model rankings with their respective evaluation metrics can produce a global increasing of the quality of forecasting. Specially, this improvement has been observed in the prediction intervals with higher associated probability. Therefore, this approach is more convenient than simply evaluating the global quantile distribution (using pinball) of all types of models and using the best of these to obtain the rest of the probabilistic models (which would be computationally simpler but would achieve a worse quality of forecasting). Considering the obtained results, the next research steps would be applying the forecasting techniques and procedures in a case study where the optimization of the energy management of the campus and its electric mobility charging stations is done by means of stochastic methods (for them to include the probabilistic forecasting information). In this way, it will be possible to evaluate how the quality of the probabilistic forecasting affect to the optimization, and which is the computational load required for their calculation. Figure 3. Plot of the prediction intervals of the variable lc for five days. a) interval 98%; b) interval 90%; c) interval 80%. FUNDING AND ACKNOWLEDGEMENTS The publication is part of the project TED2021-129702B-I00, funded by MCIN/AEI/10.13039/501100011033 and the European Union “NextGenerationEU”/PRTR. The authors want to acknowledge the University of Genoa, especially the researchers at the Savona Campus, for their help and support that have made it possible to carry out the present study. REFERENCES Aicardi, M., Ferro, G., Minciardi, R., & Robba, M. (2023). Stochastic Traffic assignment with Electric Vehicles: a convex optimization approach. IFACPapersOnLine, 56(2), 2. https://doi.org/10.1016/j.ifacol.2023.10.139 Al-Rajab, M., & Loucif, S. (2024). Sustainable EnergySense: a predictive machine learning framework for optimizing residential electricity consumption. Discover Sustainability, 5(1), 1. https://doi.org/10.1007/s43621- 024-00243-0 Delfino, F., Ferro, G., Robba, M., & Rossi, M. (2019). An Energy Management Platform for the Optimal Control of Active and Reactive Powers in Sustainable Microgrids. IEEE Transactions Industry Applications, 55(6), 6. https://doi.org/10.1109/tia.2019.2913532 Domínguez-Cid, S., Larios, D. F., Barbancho, J., Salvador, A. G., Quintana-Ortí, E. S., & León, C. (2023, July). TEFNEN: Transformer for Energy Forecasting in Natural Environment. 2023 3rd International Conference Electrical, Computer, Communications Mechatronics Engineering(ICECCME). https://doi.org/10.1109/iceccme57830.2023.10253223 Ennassiri, Y., Ferro, G., Bellotti, D., Magistri, L., & Robba, M. (2023). Optimal scheduling and real-time control of a microgrid with an electrolyzer and a fuel cell systems using a reference governor approach. Sustainable Energy, Grids Networks, 36, 101218. https://doi.org/10.1016/j.segan.2023.101218 Ferro, G., Minciardi, R., Parodi, L., & Robba, M. (2023). Optimal Location and Line Assignment for Electric Bus Charging Stations. IEEE Systems Journal, 17(2), 2. https://doi.org/10.1109/jsyst.2023.3240364 Franc, A. L., Carpentier, P., Chancelier, J.-P., & Lara, M. L. (2021). EMSx: a numerical benchmark for energy management systems. Energy Systems. https://doi.org/10.1007/s12667-020-00417-5 Fresia, M., & Bracco, S. (2023). Electric Vehicle Fleet Management for a Prosumer Building with Renewable Generation. Energies, 16(20), 20. https://doi.org/10.3390/en16207213 Gouveia, C., Moreira, J., Moreira, C. L., & Lopes, J. A. P. (2013). Coordinating Storage and Demand Response for Microgrid Emergency Operation. IEEE Transactions Smart Grid, 4(4), 4. https://doi.org/10.1109/tsg.2013.2257895 Harmouch, F. Z., Krami, N., & Hmina, N. (2018). A multiagent based decentralized energy management system for power exchange minimization in microgrid cluster. Sustainable Cities Society, 40, 416–427. https://doi.org/10.1016/j.scs.2018.04.001 Hong, T., & Fan, S. (2016). Probabilistic electric load forecasting: A tutorial review. International Journal Forecasting, 32 (3), 3. https://doi.org/https://doi.org/10.1016/j.ijforecast.2015.11. 011 Khosravi, A., Nahavandi, S., & Creighton, D. (2013). Quantifying uncertainties of neural network-based electricity price forecasts. Applied Energy, 112, 120–129. https://doi.org/10.1016/j.apenergy.2013.05.075 Ma, J., & Ma, X. (2018). A review of forecasting algorithms and energy management strategies for microgrids. Systems Science & Control Engineering, 6(1), 1. https://doi.org/10.1080/21642583.2018.1480979 Mazhari, S. M., Safari, N., Chung, C. Y., & Kamwa, I. (2019). A Quantile Regression-Based Approach for Online Probabilistic Prediction of Unstable Groups of Coherent Generators in Power Systems. IEEE Transactions Power Systems, 34 (3), 3. https://doi.org/10.1109/tpwrs.2018.2888831 Nowotarski, J., & Weron, R. (2018). Recent advances in electricity price forecasting: A review of probabilistic forecasting. Renewable Sustainable Energy Reviews, 81, 1548–1568. https://doi.org/10.1016/j.rser.2017.05.234 Parejo, A., Bracco, S., Personal, E., Larios, D. F., Delfino, F., & León, C. (2021). Short-Term Power Forecasting Framework for Microgrids Using Combined Baseline and Regression Models. Applied Sciences, 11(14), 14. https://doi.org/10.3390/app11146420 Parejo, A., García, S., Personal, E., Guerrero, J. I., Carrasco, A., & León, C. (2024). Probabilistic Forecasting Framework Oriented to Distribution Networks and Microgrids. IEEE Transactions Automation ScienceEngineering, 1–13. https://doi.org/10.1109/tase.2024.3361651 Piazza, G., Bracco, S., Delfino, F., Somma, M. D., & Graditi, G. (2023). Impact of electric mobility on the design of renewable energy collective self-consumers. Sustainable Energy, Grids Networks, 33, 100963. https://doi.org/10.1016/j.segan.2022.100963 Valderrama, D. F., Ferro, G., Parodi, L., & Robba, M. (2023, July). A Multilevel Optimization Model for a Distribution Power Grid with the Active Participation of Electric Vehicles Via Aggregators. 2023 9th International Conference Control, Decision Information Technologies (CoDIT). https://doi.org/10.1109/codit58514.2023.10284285 Yang, D. (2020). Reconciling solar forecasts: Probabilistic forecast reconciliation in a nonparametric framework. Solar Energy, 210, 49–58. https://doi.org/10.1016/j.solener.2020.03.095 Zhou, B., Wang, H., Xie, Y., Li, G., Yang, D., & Hu, B. (2024). Regional short-term load forecasting method based on power load characteristics of different industries. Sustainable Energy, Grids Networks, 38, 101336. https://doi.org/10.1016/j.segan.2024.101336