A heuristic dataset reduction for green computing of photovoltaic power generation prediction

A heuristic dataset reduction for green computing of photovoltaic power generation prediction Ana Paula Aravena-Cifuentes 1 •J. David Nun ˜ez-Gonzalez 1 •Manuel Gran ˜a 1 Accepted: 25 May 2025 / Published online: 16 June 2025 ÓThe Author(s) 2025 Abstract Artificial Intelligence (AI) has become increasingly integrated into everyday life, with the general population progressively relying on it for even routine tasks. As AI models grow in complexity, precision, and computational power, their energy consumption rises exponentially, raising serious concerns about the sustainability of their widespread adoption. The field of green learning aims to mitigate these concerns by developing energy-efficient AI solutions. In this work, we propose a method to preserve the accuracy of photovoltaic (PV) power generation forecasting while reducing the environmental impact through dataset size reduction. Our approach employs a heuristic strategy that iteratively reduces the training dataset until a cutoff point is reached, balancing predictive accuracy and environmental efficiency. Experimental results using publicly available PV generation datasets demonstrate that the proposed data reduction method decreases training time by up to 17.13%, with only a 1.47% decline in prediction accuracy. These findings highlight the potential of the method to substantially reduce the carbon footprint of AI applications with minimal performance degradation. Keywords Green Computing Photovoltaic power generation prediction Dataset reduction 1 Introduction Photovoltaic (PV) solar energy is a renewable, inexhaustible, and environmentally friendly energy source derived from the conversion of sunlight into electricity using photoelectric technology (Haegel and Kurtz 2022). Among its many advantages are its minimal environmental impact, the ability to store surplus energy in batteries, and its straightforward installation in both industrial and residential settings. Furthermore, the production of PV panels has a significantly smaller carbon footprint compared to other energy generation technologies. As a result, recent decades have witnessed a notable increase in the deployment of renewable energy sources, particularly solar photovoltaic systems, as part of the global effort to ensure sustainable electricity production (Aravena-Cifuentes et al. 2023). Solar PV power generation, however, is inherently dependent on various environmental factors-especially dynamic weather conditions-which influence the intensity, duration, and angle of solar radiation. These variables introduce substantial challenges to the accurate prediction of solar power output, a task currently addressed using statistical forecasting techniques and Artificial Intelligence (AI)-based models (Abdelsattar et al. 2024). Reliable forecasting is crucial to harmonize energy production with demand, reduce dependence on fossil fuels, and ensure efficient integration of solar energy into the electrical grid (Chau et al., 2028). In particular, short-term forecasting is essential for enhancing grid stability and reliability (Almonacid et al. 2014). Over the past decade, the field of AI-and particularly machine learning (ML)-has made remarkable strides in improving forecasting accuracy. Much of this progress has been driven by the availability of increasingly large datasets, which support the development of deep neural network architectures with high predictive power (Jay Kuo et al. &J. David Nun ˜ez-Gonzalez [email protected] Ana Paula Aravena-Cifuentes [email protected] Manuel Gran ˜a [email protected] 1 computational intelligence group, university of basque country, otaola av 29, 20600 Eibar, gipuzkoa, Spain 123 Natural Computing (2025) 24:637–649 https://doi.org/10.1007/s11047-025-10029-6(0123456789().,-volV)(0123456789().,-volV) 2023). However, a growing concern, often overlooked, is the substantial carbon footprint associated with the training of these models. Deep learning (DL) models, due to their inherent complexity and reliance on massive datasets, require extensive computational resources, thus undermining broader sustainability goals (Schwartz et al. 2020). In response, the field of green computing-also referred to as sustainable computing-has emerged as a critical area of research within ML (Raja 2021; Paul 2023). As training and deploying ML models become increasingly resourceintensive, researchers and industry practitioners are seeking methods to minimize energy consumption and environmental impact. Strategies under exploration include hardware optimization, energy-efficient training techniques, and the incorporation of renewable energy sources into computational infrastructure. A key research direction within this domain is the development of energy-efficient model training techniques. Innovations such as model pruning, quantization, and knowledge distillation aim to reduce computational complexity without significantly compromising performance. The focus of this work is the reduction of training dataset size while maintaining the predictive efficiency of the ML model. This strategy aims to reduce training time and, consequently, the carbon footprint associated with ML workflows (Ougiaroglou et al. 2023). Data reduction can be achieved through several techniques, including feature selection and extraction, discretization, instance selection, and data generation (Ramı ´rez and Sergio 2014). The method proposed in this paper centers on reducing the training dataset by preserving key statistical characteristics-specifically, the mean and standard deviation-of the original data. It is important to emphasize that the primary goal of this work is not to enhance predictive performance, but rather to contribute foundational ideas in green natural and artificial computation. The remainder of this paper is organized as follows: Section 2 reviews the state of the art. Section 3 presents the experimental data and the proposed dataset reduction method. Section 4 describes the experimental design. Section 5 reports the results. Section 6 concludes the paper. 2 Related works 2.1 Green computing There is a growing body of scientific research dedicated to exploring the sustainable use of information technologies, particularly in light of their increasing ubiquity in daily life. Among these technologies, artificial intelligence (AI) tools stand out due to their widespread adoption and significant computational demands. A methodological framework for estimating the carbon footprint of computational tasks in a standardized and reliable manner is presented by Lannelongue et al. (2021). The authors propose a set of metrics designed to contextualize greenhouse gas (GHG) emissions across diverse computational domains. Their approach integrates energy consumption data from various sources-including processors, memory, and facility overhead-and accounts for geographical factors. The methodology was validated using use cases from particle physics simulations, weather forecasting models, and natural language processing (NLP) systems. A key contribution of this work is its balance between estimation accuracy and practical applicability, making it a valuable tool for assessing the environmental cost of computational workloads. In another notable contribution, Sheryl et al. (2023) propose a heuristic-based green computing approach combined with an optimized routing protocol for energyaware management in secure systems. The energy requirements are estimated using a first-order radio model, which quantifies the energy needed to transmit a packet of m bits. Leveraging this data, the authors suggest that it is possible to define energy-saving standards through the use of a Management Information Base (MIB) model tailored to energy management. Their approach not only improves energy efficiency but also enhances security in energysensitive applications. Additionally, green computing principles have been applied in the context of the Internet of Things (IoT), as explored by Jaiswal et al. (2021). Given the massive connectivity and energy constraints of IoT networks, energy harvesting (EH) has emerged as a key strategy to prolong network lifetime. However, conventional EH techniques often result in significant energy consumption at the sensor level. To address this limitation, the authors propose a Time-Switching Simultaneous Wireless Information and Power Transfer (T-SWIPT) protocol. This technique enhances the energy efficiency of sensor-enabled IoT networks by allowing concurrent data and energy transmission, thereby reducing energy overhead and extending device operability. 2.2 Dataset size reduction Early research on dataset size reduction is exemplified by the doctoral dissertation of Lozano and Marı ´a(2007), which addresses the construction and condensation of training sets from given data. In addition to formulating strategies for dataset reduction, the work compares several classification techniques. The author categorizes condensation methods into two main groups: non-adaptive and adaptive techniques. The former includes approaches such as the Nearest Centroid Neighbour (NCN) Rule, MaxNCN, 638 A. P. Aravena-Cifuentes et al. 123 and Reconsistent. The latter group encompasses NCNbased Adaptive Condensation Algorithms and Gaussianbased Adaptive Condensation Algorithms. The study concludes that MaxNCN, Hart, and Reconsistent yielded the most promising results in terms of classification performance and data reduction efficiency. To further reduce the training set size, Chouvatut et al. (2015) introduce an enhancement to the well-known graphbased Optimum-Path Forest (OPF) classifier. Their approach leverages the Segmented Least Squares Algorithm (SLSA) to estimate the structure of the classification tree more efficiently. This methodology achieves a reduction in training set size ranging from 7% to 21%, while maintaining classification accuracy with only marginal losses between 0.2% and 0.5% across the evaluated datasets. More recently, Ougiaroglou et al. (2023) presents advancements aimed at accelerating the performance of the k-Nearest Neighbour (k-NN) classifier. Specifically, the study introduces two novel prototype generation algorithms for multi-label datasets. The first is a variant of the Reduction by Homogeneous Clustering (RHC) method, while the second is a modified version of the RSP3 algorithm-originally designed for single-label tasks and notable for being parameter-free. Both techniques contribute to reducing computational complexity without compromising classification effectiveness. In this context, our contribution proposes a heuristic method for dataset reduction that preserves the statistical representativeness of the original data. Unlike traditional dimensionality reduction techniques, which typically operate on attribute space, our approach operate on instance space. This methodology represents a novel perspective in the pursuit of sustainable machine learning. It is worth noting that our previous studies (Aravena-Cifuentes et al. 2023,2024) did not incorporate dataset reduction strategies, which are introduced and evaluated in the present work. 3 Data and algorithm 3.1 Data This study leverages localized photovoltaic (PV) energy production data published by Williams and Wagner (2019), preceding the synthesis provided by Pasion et al. (2020). The dataset comprises power output measurements from twelve U.S. Department of Defense (DoD) solar installations distributed across diverse climatic regions within the United States, covering the period between 2017 and 2018 (see Figure 1). The dataset contains 21,046 instances and 17 attributes, encompassing solar power output along with a variety of geographical, temporal, and meteorological features. The sampling frequency ranges from 15 min to several hours, depending on the installation. All experiments were conducted on a machine equipped with an Intel(R) Core(TM) i5-10300 H CPU @ 2.50GHz, 8.00 GB RAM (7.78 GB usable), running a 64-bit operating system. Data preprocessing and modeling were implemented in Python within the Anaconda distribution, using the Spyder development environment. During the preprocessing phase, categorical values in the Season attribute were encoded numerically to enable model ingestion. A prior study by Aravena-Cifuentes et al. (2024) extended the benchmark introduced in AravenaCifuentes et al. (2023), which initially focused on the Travis Air Force Base dataset-identified as the highestperforming location in terms of PV output. The extended analysis incorporated data from additional locations: Malmstrom, MNANG, Hill Weber, and Camp Murray. Figures 2to 6present scatter plots of the Polypower variable for each selected location. Input features used for model training include temporal (time, month, season), geographical (latitude, altitude), and meteorological variables (humidity, ambient temperature, wind speed, visibility, barometric pressure, and cloud cover). The apparent dispersion in the scatter plots is attributed to the high sampling resolution (minute-level intervals), compounded by irregular data acquisition schedules across sites. Some locations exhibit substantial data gaps, while others are densely populated with high-frequency measurements. Fig. 1 Heat map of power generation (Polypower variable) per location and month A heuristic dataset reduction for green computing of... 639 123 Fig. 4 Polypower samples from Malmstrom location Fig. 3 Polypower samples from Hill Weber location Fig. 5 Polypower samples from MNANG location Fig. 2 Polypower samples from Camp Murray location 640 A. P. Aravena-Cifuentes et al. 123 3.2 Data reduction algorithm The objective of this work is to minimize the amount of training data required by a Random Forest (RF) regression model while preserving the underlying statistical structure of the dataset, maintaining predictive accuracy above a predefined threshold, and reducing the environmental impact associated with computational resource usage. This study builds upon prior work by Aravena-Cifuentes et al. (2023), which proposed a baseline RF model trained using a standard five-fold cross-validation strategy in 5 seeds (25 runs in total). The same protocol is followed here to reestablish a benchmark prior to data reduction. Subsequently, the model is retrained on a reduced dataset, generated by the novel heuristic described in Algorithm 1. This data reduction heuristic operates by leveraging a stochastic sampling mechanism inspired by principles of diversity-preserving selection. Specifically, one data instance (i.e., row) is randomly selected for each feature column-ensuring that no instance index is reused-so that each selected value comes from a distinct row. These values are then merged to construct a new synthetic instance. The rows from which the selected values were drawn are subsequently removed from the dataset, and the synthetic instance is appended, yielding a compressed dataset. This process achieves a compression ratio equal to the number of variables (e.g., 12:1 for a 12-variable dataset), and maintains the original feature dimensionality. Given the stochastic nature of the reduction process, an iterative refinement scheme is applied. Each reduced dataset candidate is evaluated using a fitness function that quantifies statistical similarity to the original dataset. This evaluation, detailed in Algorithm 2, computes and compares summary statistics-namely mean, median, and standard deviation-between the original dataset and the candidate reduced dataset. A reduced dataset is considered valid if the deviations of all selected statistics remain Fig. 6 Polypower samples from Travis location Algorithm 1 Heuristic strategy for data reduction A heuristic dataset reduction for green computing of... 641 123 within user-defined thresholds. This enables a flexible calibration of the trade-off between data compression and statistical fidelity. 4 Experimental design In this section, we present the experimental setup and results used to evaluate the performance of the proposed data reduction strategy. The methodology established in the benchmark study by Aravena-Cifuentes et al. (2023) serves both as a reference framework and as a comparative baseline for assessing the new approach. That study underscored the necessity of tailoring predictive models to specific geographic locations, and emphasized the dual importance of achieving high predictive accuracy while minimizing the environmental impact associated with computational resource consumption. To ensure robustness, each iteration of the proposed data reduction algorithm is subjected to a 5-fold cross-validation protocol. Furthermore, to mitigate the effects of stochastic variability, the algorithm is executed under five different random seeds, resulting in a total of 25 runs per iteration. The initial iteration (Iteration 0) uses the complete training dataset without any reduction, serving as a baseline. Subsequent iterations progressively reduce the training dataset by applying the heuristic reduction function described in Algorithm 1. Two experimental configurations are examined in this study, hereafter referred to as Method 1 and Method 2. In Method 1, the entire training dataset is subjected to the reduction process. In contrast, Method 2 restricts the reduction process to instances corresponding to the months of June through September. This temporal segmentation is motivated by the distribution of the original dataset, which spans from 09 June 2017 to 03 October 2018, and exhibits higher data density during the summer months. Finally, objective evaluation metrics such as mean square error (MSE), mean absolute error (MAE), and coefficient of determination 1(R2) are calculated to assess the performance of the models at each fold, seed, and generation. The coefficient of determination R2is calculated using Equation 1. R2¼1SSres SStot ð1Þ where SSres: Sum of squared differences between predicted and actual values SStot: Total sum of squared differences between actual values and their means. Additionally, the mean absolute error (MAE) and the mean square error (MSE) are calculated using Eqs. 2and 3, respectivel:. MAE ¼1 nX n i¼1 jypred ytruej;ð2Þ MSE ¼1 nX n i¼1 ðypred ytrueÞ2 ;ð3Þ where nis the number of data instances, ypred are predicted values and ytrue are the observed values 5 Results This section presents the results obtained through the implementation of the proposed data reduction technique for solar power generation prediction. The analysis illustrates the evolution of model performance throughout the optimisation process. Algorithm 2 Fitness function 642 A. P. Aravena-Cifuentes et al. 123 Figures 7,8and 9compare the performance of Method 1 and Method 2 in terms of R-squared (R2) against the size of the reduced datasets, measured by the number of training instances, for the Travis location. The original training dataset contained 2197 instances. A reduction to 2000 instances (approximately 10%) in both methods, regardless of the threshold value, did not lead to notable changes in the evaluation metrics, as shown in Figure 9. The best results were obtained with a tolerance threshold of 18% for Method 1 and 20% for Method 2. In particular, Method 2 achieved a substantial reduction of 17.13% in the training dataset size with only a 1.5% decrease in the R2 value. Method 1, by comparison, resulted in an 8.3% reduction in dataset size with a minimal 0.78% drop in the average coefficient of determination. Method 1 showed that increasing the tolerance threshold in the selection function allowed more aggressive data reduction, leading to fewer training instances. However, this was accompanied by a degradation in predictive performance, demonstrating the trade-off between dataset compactness and model accuracy. In contrast, Method 2 exhibited a different pattern. Once the threshold exceeded 19%, no further instances were removed, as the reduction mechanism is constrained to a fixed temporal window within the training data. As a result, a plateau was reached where additional threshold increments no longer affected the training set size. This behaviour is clearly illustrated in Figure 9. The Camp Murray location, which already has a few instances, specifically 890. We stopped the process after only 2 iterations because reducing the size of the dataset gets worse. Figures 10 and 11 present the results for the Hill Weber location. In this case, both data reduction methods were applied, achieving equivalent levels of reduction; however, Method 2 yielded slightly better performance. The R2 values remain comparable between the two methods for dataset sizes down to approximately 1200 instances. The results for the Malmstrom location are depicted in Figures 12 and 13. Method 1 demonstrates good performance for reductions in the range of 10–15%; specifically, the removal of 200 instances has a negligible impact on R2. However, beyond this range, performance metrics begin to deteriorate, as is particularly evident around the 18% reduction mark. Finally, Figures 14 and 15 display the results for the MNANG location. The behaviour observed is similar to that of the Camp Murray location, though with improved overall performance. Given the original dataset of 600 instances, acceptable predictive accuracy is maintained with reductions of up to 20%. The Travis dataset, when analyzed using Method 1 (see Table 1), shows relatively stable results with MAE values around 1.95 for the different window sizes (10, 15, 17, 18), with slight fluctuations in the standard deviation. The MSE values also remain close, hovering around 9, which suggests consistent prediction errors. The R2values are strong, consistently above 0.79, indicating a good fit of the model to the data, with small variability. Overall, Method 1 seems to perform consistently across the different parameters with strong prediction performance and little variation. In contrast to Method 1, Method 2 (see Table 2for the Travis dataset shows slightly higher MAE values, ranging from 1.96 to 2.07, which indicates a marginally less accurate model compared to Method 1. The MSE values are similarly higher than those in Method 1, peaking at 9.96. However, R2values remain relatively stable, with slight fluctuations around 0.79. While the MAE and MSE values suggest slightly worse performance, the R2values indicate that the model still provides a good fit, albeit with a small increase in error margins. 5.1 Camp murray The results for Camp Murray (see Table 3) show consistent and stable performance across various configurations. The Mean Absolute Error (MAE) fluctuates slightly between 2.274 and 2.278, with the standard deviation remaining close to 0.18, indicating minor variation in the model’s predictive accuracy. This consistency suggests that the model performs reliably in terms of absolute prediction error. For the Mean Squared Error (MSE), values range from 12.250 to 12.272, again showing stable prediction errors across the different configurations. The standard deviation of MSE is also quite low, indicating little fluctuation in the squared errors between runs. The R2values are consistently around 0.741, with a very small standard deviation of 0.04. This indicates that the model is able to explain about 74% of the variance in the data, and the predictive power is stable across different runs. The performance of Method 1 at Hill Weber (see Table 4) shows a clear trend of degradation as the parameter value increases. MAE rises steadily from 2.492 at 10 to 2.802 at 25, indicating a growing average error in predictions. Likewise, MSE increases from 13.931 to 16.860, showing a significant rise in squared errors. The R 2 score drops from 0.700 to 0.647, reflecting a reduced ability of the model to explain the variance in the data. The standard deviations also increase slightly, particularly for MSE and R2, suggesting that the model’s reliability decreases as the parameter grows. A heuristic dataset reduction for green computing of... 643 123 Method 2 (see Table 5) shows a more stable behavior across configurations, particularly when compared to Method 1. MAE remains relatively low and fluctuates less dramatically, ranging from 2.492 to 2.662. Similarly, MSE increases moderately, from 13.924 to 15.277, which is a slower growth than in Method 1. Interestingly, the R 2 values remain close to 0.690 through most of the range and only decrease slightly toward the end, indicating a more consistent explanatory power. The standard deviations remain small across all metrics, pointing to higher model robustness. The model’s performance at Malmstrom (see Table 6) demonstrates considerable variability and a clear tendency toward deterioration as the parameter value increases. The MAE grows from 2.654 at parameter 10 to 3.183 at 25, reflecting a significant rise in average prediction error. Fig. 7 Travis location. Performance results under the data reduction processed followed by Method 1 (left) and Method 2 (left). The evolution of the graphs is from right to left, i.e. the decreasing number of instances in the dataset Fig. 8 Travis location. Performance results under the data reduction processed followed by Method 1 (left) and Method 2 (right): A quadratic approximation was used to filter out the noise in the results 644 A. P. Aravena-Cifuentes et al. 123 Similarly, the MSE follows an upward trajectory, increasing from 16.427 to 21.130, indicating larger squared deviations. The R 2 score consistently decreases from 0.671 to 0.574, showing a weakening explanatory power of the model. Interestingly, parameter 18 marks a temporary recovery in performance, with metrics nearly matching those at parameter 10. However, this appears to be an outlier, as results continue worsening afterward. Standard deviations also increase, particularly for MAE and MSE, highlighting growing inconsistency and reduced model stability at higher parameter values. Overall, the model performs best at the lower parameter values, especially at 10 and 18. The MNANG model (see Table 7) shows consistent yet moderate predictive performance across all parameter configurations. The MAE values remain within a narrow range, from 2.832 to 2.940, reflecting relatively stable absolute error. MSE values increase gradually from 18.406 to 19.512, indicating a slight rise in the average squared error as model complexity grows. Correspondingly, the R2 metric declines from 0.689 to 0.669, suggesting a small reduction in the proportion of variance explained. Fig. 9 Travis location. Performance results under Method 1 (left) and Method 2 (right): Close-up considering the minimum number of instances achieved with Method 2 Fig. 10 Hill Weber location. Performance results under Method 1 (left) and Method 2 (right) A heuristic dataset reduction for green computing of... 645 123