Low-Latency Control of Wind Energy Systems Using Edge-Deployed Deep Learning and Multi-Objective Genetic Algorithm Optimization for Real-Time Harmonic Forecasting in Industrial Grids

Engineering and Technology Journal e-ISSN: 2456-3358 Volume 10 Issue 12 December-2025, Page No.- 8107-8125 DOI: 10.47191/etj/v10i12.11, I.F. – 8.482 © 2025, ETJ 8107 ETJ Volume 10 Issue 12 December 2025 , Adel Elgammal Low-Latency Control of Wind Energy Systems Using Edge-Deployed Deep Learning and Multi-Objective Genetic Algorithm Optimization for RealTime Harmonic Forecasting in Industrial Grids Adel Elgammal Professor, Utilities and Sustainable Engineering, The University of Trinidad & Tobago UTT ABSTRACT: The growing contribution of wind turbines to industrial grids is causing growing concern over power quality, especially the appearance of harmonic distortion affecting stability, efficiency and equipment lifetime. Classical harmonic compensation techniques are based on central processing; hence they introduce delay which make them unable to cope with quickly varying wind. A new efficient, low latency control architecture for wind energy systems is presented in this paper where edge-based deep learning models and MOGA configurations provide the ability for on-line harmonic prediction and optimal control actions. Two set of lightweight neural networks processing units are integrated at field level near wind turbine controllers to predict harmonic components milliseconds in advance, which can mitigate the burden on cloud or centralized processing. These predictions are used as input to a MOGA-driven optimization engine which minimizes distortion concurrently with reactive power support and voltage stability. Well-detailed simulations are carried out using industrial standard wind profiles and detailed grid models from a real situation under different disturbance and load situations. Experiments show that the model deployed at edge can reduce latency remarkably (up to 65% if compared with cloud-based forecasting) without scarifying much predictive accuracy. The MOGA-based controller minimizes the overall THD by 30–45% and improves dynamic grid stability without adding extra computational work. The proposed edge-intelligent and multi-objective optimized control strategy is a scalable solution for future intelligent industry grids with high penetration of renewable energy. This study provides a basis for decentralized, predictive, real-time harmonicaware control methodologies that can be generalized to other DER. KEYWORDS: Edge Computing, Deep Learning, Wind Energy Systems, Multi-Objective Genetic Algorithm (MOGA), LowLatency Control, Real-Time Harmonic Forecasting, Industrial Power Grids. I. Introduction: Wind power has developed to one of the most rapidly growing renewable energy types worldwide and is increasingly being integrated into the industrial and utility food chain. With the increasing wind power penetration, the PQ (Power quality) and GBT are of great importance in both industry grid with sensitive load as well as non-linear power electronics [1]. The increased use of wind energy represents one of the most significant issues in terms of harmonics due to the incorrect behavior that lead power converters and inverters contribute to harmonic voltage distortion which in turn leads to excessive losses, shorter service life for cables, transformers, generators and interference with industrial production processes; then subsistence quality. Consequently, real-time prediction and suppression of harmonics becomes an important area for the integration of renewable reliably in a power grid. In spite of the advances in modeling, harmonic analysis and control strategies for wind turbines, common techniques still adopt a centralized treatment, leading to delays that prevent rapid response to dynamic grid behavior. Recently, edge-based computing, deep learning and multi-objective optimization especially genetic algorithms (GAs) have been introduced as appealing tools to develop decentralized predictive and low-latency control systems. Nevertheless, their integrated use for real time harmonic prediction and control in industrial networks has not been deeply studied yet. The review interrelated domains that, together, determine the challenges and opportunities in windintegrated industrial grids. It starts with an investigation of the sources, characteristics, and effects of harmonic distortions in wind systems, during which it shows that the power electronic interfaces together with fluctuating nature of wind dynamics are the main origins that generate nonlinear distortions to degrade power quality. On this premise, the review next reviews existing low latency control strategies also indicating how traditional centralised and delay-susceptible control architectures do not offer satisfactory solutions when implemented over fast evolving network scenarios. In response to latency, it is shown that distributed processing at the edge of grid provides a solution where the time delay can be significantly reduced for improved real-time performance. Concurrently, the article reviews progress on deep learning-based harmonic prediction techniques with particular focus given to their superiority “Low-Latency Control of Wind Energy Systems Using Edge-Deployed Deep Learning and Multi-Objective Genetic Algorithm Optimization for Real-Time Harmonic Forecasting in Industrial Grids” 8108 ETJ Volume 10 Issue 12 December 2025 , Adel Elgammal over conventional signal processing methods when handling non-stationary and complex power system disturbances. In order to meet the multi-dimensional performance objectives in grid stability and power quality enhancement, it also discusses the significance of using multi-objective optimization approach with Genetic Algorithms (GAs) which offer a robust framework for accommodating tradeoffs between competing control objectives such as reducing harmonics, maintaining voltages within the tolerance limits and supporting reactive power. Finally, it puts together discoveries from nascent models of hybrid AI-optimization control to demonstrate that the amalgamation of artificial intelligence and evolutionary optimization methods provides a potentially viable—but yet under-researched—path toward developing fast, accurate, and adaptive control strategies for renewable-rich industrial grids. Critical research gaps are identified, and motivations for the proposed framework: Low-Latency Control of Wind Energy Systems Using EdgeDeployed Deep Learning Multi-Objective Genetic Algorithm Optimization for Real-Time Harmonic Forecasting in Industrial Grids. The wind systems’ harmonic degradation stems mainly from modern converter power interfaces of turbines, including back-to-back converters; pulse-width modulation (PWM) inverters and variable-speed driving with their loads. It has been reported that the nonlinear switching of the converters plays a dominant role for harmonics current introduced to the microgrid [1]. Variable-speed DFIG (Double Fed Induction Generators) dominate wind turbine installations worldwide, and they produce a lot of low-order and high-order harmonics under unbalanced voltage and varying wind speed conditions [2]. Full-converter Permanent Magnet Synchronous Generators (PMSG), are more friendly for the grid, but they also inject significant switching harmonics with heavy loading [3]. Harmonic levels are also altered by wind turbulence, mechanical drive-train oscillations and changes in blade pitch changing the amplitude of the harmonics dynamically and unpredictably [4]. These phenomena make urgent the development of real-time forecast, rather than post-event response. Industrial power systems have a high penetration of nonlinear industrial loads, such as motor drives, arc furnaces and adjustable speed drives which makes them more susceptible to harmonics than the transmission system [5]. Standards like IEEE 519 define very tight harmonic limits for voltage and current THD. It can induce to overheating, the out of order dimension of sensitive equipment, to energy losses, and low production efficiency [6]. The harmonic distortion becomes more difficult to control with the increasing level of wind penetration due to the uncertainty and rapid fluctuations of wind energy. It has been well studied in the literature that conventional fixed compensation structures (e.g., passive filters) are commonly not capable of meeting stringent requirements posed by fast-varying and unpredictable grid content, due to their immutable characteristics [7]. Whilst several mitigation solutions have been presented in the literature, such as passive [8] and active filtering techniques, model predictive control (MPC) for converter systems [9], harmonic extraction using Fast Fourier Transform (FFT) [10] & Wavelet Transform methods were proposed [12] along with droop-based control schemes for distributed energy networks [11]; nevertheless the shortfalls of these methods are notable. Specifically, these known techniques typically exhibit long computation delays, lack of rapid response capability to dynamic wind-induced disturbances, decreased accuracy in handling non-stationary harmonic characteristics and dependency on centralized processing structures which add extra delays. Together, these limitations emphasize the urgent requirement for more sophisticated and predictive harmonic management strategies with a low-latency, distributed response capable of on-the-fly adjusting to the complex and dynamically changing behaviours in terms of network harmonics presented by modern industrial grids including wind power. Prior art As to the same optimization of conventional control architectures for renewable energies, sensor measurements are forwarded to remote controllers or to cloud servers for processing in order to determine a corresponding, optimal order (additionally in presence of delays resulting from data exchange), which however is a severe obstacle if an immediate response ability i.e. real time capability is claimed [12]. Commercial networks, and more particularly distribution grids, need sub-cycle response times— considered as the millisecond magnitude in order to efficiently suppress sharp harmonic peaks through right level control of power quality. Nevertheless, network latencies (propagation, queuing and processing delays as well as sampling) are stacked layers-by-layers on the communication pipeline. These delays add up as the renewable penetration goes higher and data flow becomes more intensive, which eventually slows down the actuation time of DG and damages the system performance overall [13]. To overcome these limitations, several control methodologies have been presented in the literature including e.g., Proportional– Resonant (PR) controllers [14], Model Predictive Control (MPC) based [15] and adaptive/robust-based control design approaches [16]. Although suitable in various instances, such methods can include algorithmically intense mechanisms which are not readily adaptable for ultra-low-latency solutions and particularly when provided within centralized or cloud adapted environments. The emergence of edge computing in the literature has all therefore looked for advantages when computation is moved closer to the measurement where data is collected. Such systems allow them to perform control and prediction computations at the edge nodes without incurring any unnecessary communication overhead, resulting in fast localized decisionmaking with low latency [17]. Besides mitigating the latency by orders of magnitude, (ii) edge computing can significantly decrease data transmission and improve robustness on sporadic connectivity shared environments [18]. These “Low-Latency Control of Wind Energy Systems Using Edge-Deployed Deep Learning and Multi-Objective Genetic Algorithm Optimization for Real-Time Harmonic Forecasting in Industrial Grids” 8109 ETJ Volume 10 Issue 12 December 2025 , Adel Elgammal features render edge-intelligent control especially attractive for online harmonic prediction and active wind turbine control in industrial network systems. Edge computing, also referred to in the literature as fog computing or distributed computing, has recently emerged as a disruptive paradigm for future power systems. In general applications, computations are offloaded from the central servers to mobile devices, such as sensor nodes or microcontrollers in wind turbine nacelles, substations and distributed control points. While these edge devices are typically resource-constrained, they are becoming more powerful in order to support running shallow machine learning models for real-time use cases [19]. It has been shown that decentralized edge-based architectures largely improve the overall system performance, support grid disturbance events down to microsecond response times, and lessen SCADA load by boosting grid reliability and cyberattack resilience [20]. This has made edge computing useful for voltage sag detection, microgrid event monitoring, distributed protection schemes, wind turbine predictive maintenance and real-time fault classification [21]. In spite of these advancements, the application of edge-deployed deep learning models for harmonic prediction in WE systems are still an open problem. Conventional harmonic analysis methods such as the FFT approach, Kalman filtering and the wavelet transform (WT) usually have a hard time dealing with highly nonlinear and non-stationary nature of harmonics produced by wind energy plants. Deep learning techniques alleviate these limitations through directly learning complicated temporal and spectral patterns from raw data, which can achieve more favourable modelling in the fast changing grid environment [22]. Different architectures have been applied in the literature for power quality analysis, as CNNs for harmonic classification [23], LSTM networks to predict non-stationarity electrical signals [24] and attention mechanisms or Transformer architectures to model dynamic distortions and long-range temporal dependencies introduced by nonlinear loads [25]. Deep learning-based methods have been proven to be accuracy, efficient and effective when compared with classical signal processing methods in terms of performance (accuracy, computing complexity and robustness), which makes them quite appropriate for modelling the nonlinearity and non-stationarity of harmonics in power systems. 2) One of the major drawbacks for wide adoption: their model execution relies on distant cloud servers, leading to inevitable communication latency, and higher cybersecurity risks as well as operational dependency on stable network access [26]. These limitations have motivated an increasing trend to run complex deep learning models right at the edge. Recent developments in model compression and embedded AI have allowed lightweight neural networks to operate effectively on microcontroller, ARM-based processor or other embedded hardware platforms [27], showing that well-optimized architectures can achieve high accuracy and run with a runtime on the scale of millisecond. Despite such advancement, there is a lack for real-time harmonic prediction through edge-deployed deep learning methodology particularly elastic analysis-based scenarios in wind energy systems. Wind turbine and power grid control problems inherently involve multiple conflicting objectives including total harmonic distortion (THD) reduction, reactive power management, voltage stability, etc., and are thus well suited for Multi-Objective Optimization (MOO) formulations [28]. Genetic Algorithms (GAs), which are population-based evolutionary techniques, have been successfully used in the approximation of Pareto-optimal solutions and applied to optimal power flow study [29], harmonic filter location [30] and Tuning of Power electronics controller [31]. While variations of this approach, including the Non-Dominated Sorting Genetic Algorithm II (NSGA-II) and Multi-Objective Particle Swarm Optimization (MOPSO), have performed well in problems where the objective space is nonlinear, nonlocal minimum exist, or high-dimensional search spaces govern [32]. However, due to the obvious advantages, there is no available research combining MOGA-based optimization and edge-deployed deep learning for harmonicaware wind power control. Hybrid AI–GA methods have proved to be effective in similar works, such as wind turbine maximum power point tracking (MPPT) [33], distributed energy resource (DER) coordination [34] and inverter control strategy optimization [35]. However, these works do not consider the united problems that include the very low latency control, real time harmonic prediction, edge computing and the strict operation requirements of industrial grids simultaneously. Accordingly, a number of important unaddressed research gaps can be identified: (1) the existence of no edge-implemented deep learning frameworks for predictive real-time harmonic time series forecasting for wind turbines; (2) poor integration between MOGA and edge AI for harmonic-aware control strategies; (3) insufficient attention to low latency focused harmonics mitigation methodologies that are optimized specifically towards industrial grid environments; (4) inadequate development of predictive, hybrid-AIoptimization control architectures; as well as (5) overreliance on centralized/cloud processing frameworks which introduce unacceptable delays in fast acting grid control. Nevertheless, there is a lack of research that can fully satisfy the demand for co-optimization of low-latency wind energy control and realtime harmonic forecasting under an edge computing system by introducing MDEO in the current wind power modelling, harmonic analysis, optimization methods or deep learning. The proposed research — by combining edge-intelligent deep learning prediction and ultra-low-latency deployment, MOGA-based control optimization, and industrial-grade power quality constraints — bridges this gap in the state of knowledge, to provide an innovative and high-impact connection between smart grid control and renewable-rich industrial applications. “Low-Latency Control of Wind Energy Systems Using Edge-Deployed Deep Learning and Multi-Objective Genetic Algorithm Optimization for Real-Time Harmonic Forecasting in Industrial Grids” 8110 ETJ Volume 10 Issue 12 December 2025 , Adel Elgammal II. THE PROPOSED LOW-LATENCY CONTROL OF WIND ENERGY SYSTEMS USING EDGEDEPLOYED DEEP LEARNING AND MULTIOBJECTIVE GENETIC ALGORITHM OPTIMIZATION FOR REAL-TIME HARMONIC FORECASTING IN INDUSTRIAL GRIDS. On the scheme depicted in Figure 1, with an extension toward its left (i.e., a Wind Energy System (WES) supplying the industrial grid from being behind the utility substation), WES is assumed to be the main source of renewable generation connected to this portion of industrial grid. The WES comes with a multi-domain sensor instrumentation that records electrical, mechanical and atmospheric dynamics at the turbine level. These sensors measure three-phase voltages and currents as well as the torque, rotor speed, converter switching status and real-time meteorological parameters including wind speed and turbulence intensity. The sensor's signals are sent via a dedicated communication link (dashed line) towards the Edge Device for preliminary processing and inference. The sensor interface runs at a sampling frequency suitable for the subcycle harmonic tracking (usually 5–20 kHz) in order to depict all relevant harmonic harmonics, including low-order (5th, 7th) and high order switching harmonics. Compared with conventional SCADA systems that capture at slower rates (100–1000 ms), the sensing infrastructure presented facilitates acquisition fidelity at microsecond resolution for real-time harmonic prediction. This will make the following (downstream) deep learning model obtain the high-resolution data and guarantee that non-stationarity due to evolution of harmonic under dynamic wind excitation is preserved in the data set. The Edge Device, that is placed near the turbine and power electronic converter, represents the computational heart of the architecture. To achieve ultra-low latency, we deploy a compressed deep learning (DL) model directly on an embedded ARM-class processor or microcontroller unit (MCU). The deep learning model itself—an algorithm that uses a shallow CNN-LSTM or transformer-based neural network architecture—carries out real-time harmonic prediction through monitoring measurements at high sampling rates. As in Fig. 1, the incoming sensor signals are processed by the Deep Learning Model block to produce short-horizon predictions of harmonic distortions (around 1-10 cycles ahead). These predictions cover total harmonic distortion (THD) forecasts, leading modes of harmonics, as well as dynamic harmonics present during fast wind changes or grid disturbances. Being different from cloud-based AI inference, edge deployment is free of internet reliance and significantly lowers end-to-end delay. Inference times are reduced to the milli-second range, allowing for predictions that can be used in real-time control decisions. Without long-distance datatraveling, this enhances cybersecurity resilience, since gridsensitive signals are locally processed than sent off (or in) elsewhere. Below the deep learning block, in the drawing, is drawn the integrated Multi-Objective GAs. The MOGA is used to work with the forecasting module in order to produce, from forecasted harmonic modes data, optimal control policies. The optimisation problem to be solved by the algorithm, which combines: - multi-objective in nature solution of the following optimisation criteria:  Reduction of the total harmonic distortion (THD)  Stabilization of reactive power flow  Improvement of voltage quality indices  Reduced cost or effort in control actuation The MOGA is fed by the harmonic predictions obtained from the DL model and the technical limitations of the industrial grid. Candidate control solutions (e.g., converter switching patterns, active filtering injection currents and reactive power set-points) are evolved by the algorithm using genetic operators: selection, crossover, mutation and Pareto-front evaluation. Unlike classical single-objective optimizers, application of MOGA in power systems is especially beneficial because there are natural trade-offs between the PQ parameters. It can run evolutionary optimization in real-time because of edge-accelerated computation and that the window for optimization is short—typically a few milliseconds—due to the harmonic forecast horizon given by deep learning model. The block diagram highlights the harmonic prediction flowing from deep learning module to MOGA through a downward path. These predictions correspond to predictive states of the harmonics content of the grid. The predictions comprise: amplitudes, phases and frequency of the steady-state and transient harmonics. This kind of feedback allows for adaptive, predictive control instead of reactionary control. As the MOGA obtains predictive information of the harmonic, it can minimize control effort in anticipation before harmonic distances are physically conveyed to the grid. This predictive feedback loop is one of the key innovations of the design and has allowed for stabilization and far lower levels of THD than conventional, post-event filtering or reactive harmonic mitigation. The central control block of the schematic translates the final optimized controllers derived from MOGA (step 6) and further process them via a high-level Multi-Objective Optimization module to give an improved set of solution vectors. This module coordinates the carrierlevel control actions to maintain grid level stability. While the MOGA seeks local turbine and converter statuses, the second-stage optimization enforces global boundary conditions such as:  industrial grid harmonic compliance according to IEEE 519  substation-level voltage limits  feeder-level current constraints  grid protection coordination requirements This hierarchical ordering of control mechanisms limits the system to stable behavior over a range of spatial and temporal scales. This edge-level MOGA-grid level optimization synergy ensures that the control decisions of the turbine do “Low-Latency Control of Wind Energy Systems Using Edge-Deployed Deep Learning and Multi-Objective Genetic Algorithm Optimization for Real-Time Harmonic Forecasting in Industrial Grids” 8111 ETJ Volume 10 Issue 12 December 2025 , Adel Elgammal not cause negative effects for Turbines located down-stream in the industrial grid. Below the multi-objective optimization block stands the Low-Latency Harmonic Control module, which handles the fast return of control signals to both the wind turbine converter and connected active filters. This module translates the optimized set-points into commands that can be executed by the controller like:  modifications to PWM switching angles  activation of active harmonic filters  reactive power injection via STATCOM-like converter operation  torque and pitch controller adjustments to reduce mechanical-electrical coupling harmonics The module is to be run on a deterministic hard real-time scheduler and the duration of execution is on the order of microseconds or milliseconds. This allows the turbine to proactively adjust for anticipated harmonic disturbances, while ensuring stability of the system or subsystem when load conditions change rapidly. The schematic illustrates the control commands sent (solid arrow) by the control module back to the Multi-Objective Genetic Algorithm subsystem and edge device. This feedback to the optimization approximation assures that the optimization engine is provided with updated actuator state values which can perform refinement of a next iteration based on the proven effectiveness of controls. The latter commands are repeated to the Wind Energy System that updates converter modulation signals or active filter injects in consequence. These commands are not subject to queuing delays common among centralized control systems and can be immediately executed. In the diagram, to the right of the outlay, is shown how receives from the turbine - harmonically-aware controlled signals on. This is facilitated by both electrical power injections and instantaneous harmonic compensation. When the controlled turbine injects its optimal currents into the grid, power quality is balanced in that the grid reports new voltage and current harmonics (through substation-level as well as line-end sensors/ PMUs). These harmonics measurements (dashed return arrow) feed back to the edge device, completing a feedback loop. This closed-loop cycle ensures: 1. Real-time monitoring of grid power quality 2. Continuous model correction using actual harmonic signatures 3. Adaptation to disturbances originating elsewhere in the industrial grid 4. Improved system robustness under high penetration of nonlinear loads For short term and long-term forecasting, the loop for deep learning model can always make parameter adjusted or get new calibration data. The full architecture demonstrates a hierarchical distribution of intelligence: 1. Edge-level intelligence o Deep learning inference o Local MOGA optimization o Immediate control signal execution o Sub-millisecond decision-making 2. System-level intelligence o Grid-aware multi-objective optimization o Integration of global harmonic constraints o Ensuring compliance with grid power quality standards 3. Network-level intelligence o Continuous harmonic monitoring o Coordination with industrial grid supervisory systems The hierarchical distribution eliminates bottlenecks associated with centralized, cloud-dependent systems. The architecture addresses three primary latency sources in modern renewable-integrated grids: 1. Computation latency Traditional cloud inference can take 50–200 ms. Edge inference reduces this to 1–5 ms. 2. Communication delay Cloud transmission adds 20–100 ms. Edge computation removes the internet requirement entirely. 3. Control actuation delay Optimization-based controllers may take 100–500 ms. With forecasting and edge acceleration, control actuation occurs in under 10 ms. All these delay effects are removed in the proposed system, hence it avoids harmonic amplification and reduces THD and is significantly reinforcing industrial grid stability at its dynamic states. The framework shown in the schematic view is inherently modular and well scalable to be used with different wind turbines of a farm. Each edge device is capable, not only of running autonomously, but also of sharing harmonic forecasts and optimization results with adjacent turbines, enabling co-ordinated multi-turbine optimization. The proposed distributed MOGA structure can optimise control decisions globally, rather than individually as in traditional methods, which thus help ameliorate the overall performance. It is also suitable for a variety of deployment settings, such as microgrid systems, hybrid renewables and general smart grid system. Overall, the proposed framework incorporates online measurement acquisition, edge-based deep learning for harmonic prediction, multi-objective genetic algorithm-based optimization and hierarchical system level control in an integrated manner. By means of its bi-directional communication to the industrial grid, the system shapes a predictive decentralized closed-loop control concept that is able to keep power quality and operation stability even under changing wind conditions. The flowchart in Fig. 2 shows the procedure of the proposed low-latency harmonic control method for the wind energy system. The operating mechanism starts with continuously monitoring the three-phase voltage and current “Low-Latency Control of Wind Energy Systems Using Edge-Deployed Deep Learning and Multi-Objective Genetic Algorithm Optimization for Real-Time Harmonic Forecasting in Industrial Grids” 8112 ETJ Volume 10 Issue 12 December 2025 , Adel Elgammal waveforms at turbine–grid interface, which are inputs essential for both harmonic analysis and decision-based control. These measurements are recorded at high frequency to account for fast-varying harmonics and non-stationarity of electrical perturbations. The signals that are acquired are then feed to the harmonic forecast step, where we check if a new forecast is necessary based on the last grid state. If the forecasting condition is satisfied, inference of the measurements takes place using the embedded deep learning model to produce short-horizon harmonic forecasts for the critical harmonic indices such as THD and dominant harmonic orders indicative of where each will be in time. This predictive ability allows the controller to act instead of just respond. In the absence of a trigger of forecasting, it uses the latest forecast data and then goes directly to optimization. The result from the prediction module is fed into the MultiObject Optimization block and computes optimal control inputs based on its embedded evolutionary algorithm. This optimization takes into consideration the presence of several competing performance indices, such as harmonic reduction, voltage stability and reactive power compensation. When control actions from candidate actions list are produced, then it checks if they satisfy operation limits and network quality limits. When the hoop has been successfully established, the system recycles from optimal for use (617) back to measure queue (603) when no a corrective action is warranted i.e.-- harmonic levels are not above maximum--until power conditions have changed. When the optimization converges to a solution for a valid correction action, the controller switches to a last stage, using these determined control actions onto the wind turbine converter or active filtering devices. These operations might include changing PWM switching patterns, injecting reactive power or commands for harmonic compensation. Upon completion of the execution, the system returns to its measuring phase and the closed loop is formed. This iterative loop allows for fully dynamic, anytime response to changing wind and grid conditions, thus providing a stable and high-quality power output. Fig. 1. The schematic of the Proposed Low-Latency Control of Wind Energy Systems Using Edge-Deployed Deep Learning and Multi-Objective Genetic Algorithm Optimization for Real-Time Harmonic Forecasting in Industrial Grids. “Low-Latency Control of Wind Energy Systems Using Edge-Deployed Deep Learning and Multi-Objective Genetic Algorithm Optimization for Real-Time Harmonic Forecasting in Industrial Grids” 8113 ETJ Volume 10 Issue 12 December 2025 , Adel Elgammal Fig. 2. Flowchart of the Control Procedure III. SIMULATION RESULTS AND DISCUSSION An extensive and rigorous analysis of the lowlatency control architecture was performed where edgedeployed deep learning fused with a Multi-Objective Genetic Algorithm (MOGA) is employed for real-time harmonic forecasting and predictive control in industrial power grid. In order to verify the system performance under real field conditions, all simulations have been implemented using a high-fidelity MATLAB/Simulink environment of a 3-MW VSWT system connected to a 25-kV industrial distribution feeder. The simulation environment includes full electromagnetic transient (EMT) model, covering the detailed converter switching transients and harmonic propagation through feeder impedances, as well as nonlinear industrial loads such as induction motors and thyristor-based equipment, time-varying grid impedance conditions. Such fidelity in the modeling is crucial to properly capture the non-stationary harmonic characteristics which are present as a rule in wind-integrated industrial power systems. The proposed deep learning harmonic forecasting model and MOGA optimizer were implemented on a simulated ARM Cortex-A53 edge device within a TensorFlow Lite environment and tuned to accurately capture real-world embedded processing limitations, which include memory constraints, inference latency as well lack of an explicitly dedicated GPU acceleration. By testing the control pipeline in a limited computational embedded environment, this work verifies that the proposed approach is not only theoretically sound but also computationally tractable to be implemented on future edge-intelligent wind turbine controllers. Model results are formulated to cover some of the main performance areas in regards to assessing how well the new method works. These are: (1) the performance of the deep learning model in forecasting under fluctuating harmonics; (2) end-to-end system latency reduction provided by edge deployment versus cloud-based or centralized control; (3) improvements to harmonic suppression and overall total harmonic distortion (THD) reduction under a variety operators scenarios; (4) convergence behavior and computational efficiency of MOGA optimizer within narrow real-time requirements; (5) transient response of control system under rapid changes in wind speed or turbulence conditions; (6) interaction between designed wind turbine controller and the industrial grid, with implications for voltage distortion, current harmonics, and general power quality indices; 7 ) benchmarking across conventional PI/PR controllers, active filtering etc., MPC as well as cloudnetwork artificial intelligence types of controls architectures –(8 ) scalability / robustness when extended to multi-turbine wind farms or disturbances from grid on access to models or noisy measurements–as might be predicted due to greater span between microgrid frequencies;(9)-- ablation studies isolating contributions made by/and --sensitivity analysis impacting overall system functions. In combination, the subsections above give comprehensive quantitative and qualitative explanations of these simulation results, validating the technical feasibility, performance gains, and wider “Low-Latency Control of Wind Energy Systems Using Edge-Deployed Deep Learning and Multi-Objective Genetic Algorithm Optimization for Real-Time Harmonic Forecasting in Industrial Grids” 8114 ETJ Volume 10 Issue 12 December 2025 , Adel Elgammal applications of our proposed low-latency edge-intelligent control framework for enhancing harmonic stability in high wind energy penetration industrial grids. A. Deep Learning Harmonic Forecasting Performance Table 1 provides an integrated, quantitative summary of all major performance indicators of the proposed low-latency, edge-intelligent harmonic control framework. The results collectively demonstrate that the system achieves substantial improvements in forecasting accuracy, optimization speed, harmonic suppression, and overall grid power quality compared with state-of-the-art methods. The prediction potential of the on-device deep learning model is crucial for the control framework presented to be predictive. To maintain its predictive nature, however, the model must be updated. Correct short-term prediction enables the controller to forecast trends in distortion and work to compensate for them before such disturbances spread throughout the industrial network. Our model is trained on a large dataset consisting of synthetic waveforms and real wind farm recordings to guarantee strong performance when deployed in realistic conditions. These datasets consisted of 3-phase voltage and current waveforms corrupted with various harmonic patterns— low order harmonics (5th & 7th) to higher order harmonics (11th, 13th, converter generated high frequency switching harmonics). The heterogeneity and non-stationarity of the training data population caused the model to generalize over the entire (far more general) range of harmonic conditions encountered in wind integrated industrial feeders. Model performance was assessed using various metrics such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and correlation coefficient and its temporal consistency for different flow regimes. Prediction accuracy of the deep learning models was very high for a prediction horizon of ten electrical cycles, which corresponds to 200 ms at the used sampling rate of 5 kHz. In particular, the model was able to predict harmonic profiles with an MAE of 0.0038 p.u., RMSE of 0.0061 p.u. and a high correlation coefficient of 0.983 between predicated and actual harmonics profiles. They illustrate that the model predicted steady-state and transient harmonic behavior well, especially for the dominant 5th and 7th harmonic components. Even so, for high-frequency switching harmonics the prediction accuracy degrades just moderately (∼12%), which is a reasonable result due to their naturally stochastic and rapidly changing spectral nature. In addition to pointwise predictions, the model’s temporal response was assessed in varied dynamic wind states ranging from 5 m/s to 22 m/s, corresponding to between near-rated and high turbulence load cases. The prediction error was only just over 4% even during sudden gust events, a result which compares positively with traditional FFT-based predictors that will often suffer an error rate worsening of around 15–22% when exposed to nonstationary perturbations. This immunity to fast varying harmonic conditions is essential in industrial real-time control applications where both mechanical and electrical dynamics can change abruptly. To demonstrate the suitability of deploying the forecasting model at the edge, we evaluated inference latency for a simulated ARM Cortex-A53 processor with TensorFlow Lite. The deep learning model had a mean inference time of 2.1 ms with an additional 0.7ms signal pre-processing and buffering time, which for the total per-cycle computational delay was only 2.8 ms, substantially less than the commonly accepted 10-ms requirement for real-time harmonic suppression—and essentially guaranteeing that our forecasting pipeline would not turn into a computational bottle neck. The low latency results demonstrate that the model is light-weight enough to deploy it as embedded, while preserving the predictive quality needed for anticipatory harmonic control. On the whole, these results establish that the deep learning-based forecasting module delivers accurate and fast temporal stable predictions of harmonics which form the cornerstone of our proposed predictive controller. This, together with a high level of accuracy and insensitivity to wind induced fluctuations in the light of inference times at the millisecond scale, clearly supports that our algorithm is well suited for incorporation into edge-intelligent wind turbine sensitive control strategies. Table 1: Performance Summary of the Proposed Edge-Intelligent Harmonic Control System Category Metric / Parameter Value / Result Description / Interpretation Deep Learning Forecasting Accuracy MAE 0.0038 p.u. Indicates highly precise harmonic prediction over a 10-cycle horizon. RMSE 0.0061 p.u. Low prediction error across low-, mid-, and high-order harmonics. Correlation Coefficient (R) 0.983 Strong match between predicted and actual harmonic distortion profiles. Harmonic-Specific Accuracy Low-Order Harmonics (5th & 7th) High precision Dominant harmonic components forecast with minimal deviation. High-Frequency Switching Harmonics ~12% higher error Small degradation due to random switching events. “Low-Latency Control of Wind Energy Systems Using Edge-Deployed Deep Learning and Multi-Objective Genetic Algorithm Optimization for Real-Time Harmonic Forecasting in Industrial Grids” 8115 ETJ Volume 10 Issue 12 December 2025 , Adel Elgammal Category Metric / Parameter Value / Result Description / Interpretation Temporal Forecast Stability Drift (Proposed DL Model) ≈4% Forecast remains stable across 5–22 m/s wind speeds and gusty conditions. Drift (FFT-Based Prediction) 15–22% FFT methods degrade significantly in non-stationary wind scenarios. Edge-Device Latency DL Inference Time 2.1 ms Executed on ARM Cortex-A53 with TensorFlow Lite. Pre-processing + Buffering 0.7 ms Includes windowing, normalization, and filtering. Total Inference Latency 2.8 ms Satisfies <10 ms requirement for realtime harmonic mitigation. THD Reduction Performance Baseline THD (No Control) 7.8% Harmonic levels without any mitigation. Conventional Active Filtering 5.9% Standard AHF response with limited realtime adaptability. Model Predictive Control (MPC) 4.2% Faster response than AHF but still reactive. PSO-Based Control 3.7% Improved harmonic suppression but slower convergence. Proposed DL + MOGA (Edge-Deployed) 2.3% Best overall THD reduction and lowest harmonic variability. Harmonic Order Reductions 5th Harmonic 3.1% → 0.9% 71% reduction. 7th Harmonic 2.5% → 0.7% 72% reduction. 11th Harmonic 1.2% → 0.4% 67% reduction. 13th Harmonic 0.9% → 0.3% 66% reduction. MOGA Optimization Performance Convergence Time 4.6 ms Average per optimization cycle. Generations to Pareto Stability 6–8 generations Indicates fast evolutionary adaptation. Real-Time Feasible? Yes Fully compatible with the <10 ms control window. Dynamic Wind-Condition Response THD Rise (9 → 15 m/s wind step) Conventional: 7.2%; Proposed: 5.0% Proposed method prevents large THD spikes during wind ramps. THD Variation Under Turbulence Reduced by 63% Demonstrates strong disturbance immunity. Gust-Induced Harmonic Peak <6% (Proposed) vs. >10% (Conventional) Shows robustness under IEC EOG extreme conditions. Industrial Grid Interaction Voltage THD Improvement 5.8% → 2.4% Major improvement in power quality. Current THD Improvement 9.3% → 3.1% Smoother current injection into grid. Flicker (Pst) Reduction 39% lower Important for industrial process stability. Controller Comparison (Overall Ranking) PI / PR Control Weak Reactive, limited harmonic suppression. Active Filters (AHF) Moderate Limited dynamic adaptability. Kalman-Filter Based Moderate Struggles with non-stationarity. MPC Strong Good performance but computationheavy. PSO Optimization Strong Slower convergence than MOGA. Cloud-Based AI Strong Limited by communication latency. “Low-Latency Control of Wind Energy Systems Using Edge-Deployed Deep Learning and Multi-Objective Genetic Algorithm Optimization for Real-Time Harmonic Forecasting in Industrial Grids” 8122 ETJ Volume 10 Issue 12 December 2025 , Adel Elgammal Fig. 5 Grid-Side Power Quality Improvements with DL+MOGA Controller I. Robustness and Sensitivity Analysis The performance of the proposed predictive control scheme was demonstrated under extensive adverse conditions including measurement noise, uncertainty in grid impedance, switching delays in converter and partial loss of the sensor shown in Fig. 6. The above factors characterize the most frequent causes of degradation on real life industrial grids and are also relevant indicators to evaluate the reliability in controllers. The forecasting network showed strong robustness under the influence of growing measurement noise (0–5%), with accuracy degradation kept below 7% in case of a noise level from 0 to ≈3%. This performance compares favorably to that of conventional FFT-based harmonic estimators, which lost 24% accuracy under the same condition as well as Kalman filter–based estimation, whose loss is about 18%. These results demonstrate the advantages of deep learning-based feature extraction in noisy environments. The controller preserved high robustness against grid impedance uncertainty implemented as ±20% deviation in the Thevenin equivalent. Regardless of these variations, the induced THD variation was still less than 0.4%, indicating that the forecasting and optimization layers are able to deal with changing system parameters without resulting in an instability of control actions. The robustness was examined under partial sensor failures, specifically one phase current measurement channel dropping out. The controller retained at least 80–87% of its functionality even under this failure condition. Here, we argue that such resilience is due to the MOGA’s ability to re-assign objective weights and optimize along other pathways under conditions of missing or corrupted signals. Overall, the results of this section show that the proposed DL+MOGA methodology provides robust performance behavior under noise (and other uncertainty) and sensor degradation, thus confirming its suitability for application in challenging and nonstationary industrial grid environments. Fig. 6 THD Sensitivity to Grid Impedance Variations “Low-Latency Control of Wind Energy Systems Using Edge-Deployed Deep Learning and Multi-Objective Genetic Algorithm Optimization for Real-Time Harmonic Forecasting in Industrial Grids” 8123 ETJ Volume 10 Issue 12 December 2025 , Adel Elgammal The full simulation results show that the low-latency predictive control framework presented in this study is better than conventional harmonic compensation methods for windintegrated industrial systems. With edge-deployed deep learning combined with fast multi-objective genetic algorithm, the system acquires real time awareness and proactive response that are naturally impossible to obtain in traditional centralized or reactive controls. One of the key results is that edge-based deep learning model can deliver high quality harmonic prediction in few milliseconds. Unlike cloud-based implementations, which have communication latencies greater than 100 ms, the proposed edge solution provides a total inference time of less than 3 ms, which corresponds to to real-time forecasting that allows anticipation of harmonic disturbance before they propagate such that superior grid performance can be achieved (as opposed to reaction after distortion has already impaired grid performance). The MOGA based optimum controller not only provides an improvement of the system operation, but also fulfils more than a single control objective that occurs in two competing objectives such as: i) minimization of total harmonic distortion (THD); ii) maximization of reactive power support; and iii) reduction in the switching loss. The optimization reaches convergence within 5–7 ms during stabilized flow, falling well within the sub-10 ms cycle time for control and even during high turbulence or gust encounters. This fast convergence guarantees that control actuations stay optimal under the dynamically changing operating conditions, and therefore makes it appropriate for large-scale wind energy systems in which the generated power is in a high dynamical manner. One of the features in the proposed approach is its predictive operational aspect. In contrast to traditional PI, PR, active-filter or MPC based approaches that only react once harmonic pollution is present the jointly optimized forecasting and optimization pipeline permits intervention prior to worsening of the harmonics situation. This active behavior causes a considerable attenuation of THD at the turbine interconnection as well as on the industrial load side, achieving more symmetric current injection with cleaner voltage wave shapes and much low harmonic modulation. Making a Long Life – and Process Downstream Thus, such improvements translate directly to downstream industrial processes via (a) decreased torque ripple leading to reduced thermal stress (b) fewer nuisance trips, less downtime and better service continuity (especially critical for units with induction motor s, arc furnaces or SCR supplied drives). The scalability of the optimization scheme was demonstrated in multi-turbine simulations and distributed deployment over a full farm generated cooperative harmonic suppression, as well as collective THD enhancement. This illustrates that the architecture is wellsuited for future high-penetration renewable grids with distributed intelligence needed to cope with the complexity of interacting converters and variable generation. And rigorous stability analysis and robustness checks indicate that the controller can work stably even with noisy measurement, time-varying grid impedance, switch delay variations and partial sensor faults. Degradation on deep-learning–based forecasting was under 7% in realistic noise conditions, the THD increase was bounded to less than 0.4%when ±20%grid impedance variations were present, and the optimization system solution reached up to an 87% of operational capacity evenin presence of sensor dropout. These findings indicate that the controller can be used for real world applications. To conclude, the aforementioned simulated results empirically demonstrate that adopting the proposed low latency edgeintelligent control architecture is technically feasible and very promising for future industrial power systems. It integrates predictive analytics with rapid multi-objective optimization to provide enhanced harmonic suppression, higher power quality, increased equipment reliability, fast dynamic response, and robustness under uncertainty. As industrial grids are moving towards low carbon-based (high renewable integrated) and more decentralized intelligence, the proposed approach provides a robust and scalable framework to ensure stable, high quality, resilient grid operation. IV. CONCLUSIONS This work introduced an innovative control scheme for WE system, named as a low latency framework for wind energy systems (LaWESs) where edge-implemented deep learning models and MOGA are combined to enable realtime harmonic prediction and grid-friendly control in industrial setups. The results show that decentralizing harmonic prediction and control intelligence at the edge as opposed to implementing centralized architectures effectively improves system agility for active power quality disturbance countermeasure as a result of wind stochasticity. The presented lightweight deep learning model suitable for edge deployment was able to identify the harmonic components efficiently in terms of prediction accuracy, but also to cut down computation and communication delays by over half compared with conventional cloud-based methods. This latency rate appeared to be significant for instantaneous decision making and allowed the controller to respond to harmonics disturbances in operationally pertinent time frames. The system effectively reconciled the competing objectives of reducing total harmonic distortion, enhancing reactive power compensation, and reinforcing voltage stability without adding to the computational intensity of grid operators when integrated with MOGA optimization engine. Simulation results on a wide range of industrial grid scenarios also confirmed the robustness and adaptability of the proposed approaches. The system was able to accomplish 30–45%harmonic reduction via the integrated control strategy and showcased excellent robustness against varying wind profiles, load fluctuations, and grid faults. These developments demonstrate the potential of AI and input side evolutionary optimization in solving new problems related with high penetration of renewables in industrial networks. In “Low-Latency Control of Wind Energy Systems Using Edge-Deployed Deep Learning and Multi-Objective Genetic Algorithm Optimization for Real-Time Harmonic Forecasting in Industrial Grids” 8124 ETJ Volume 10 Issue 12 December 2025 , Adel Elgammal general, this work paves the way for scalable and futureadaptive trajectory in realizing prediction-based decentralized intelligent harmonic-aware control in power systems. The structure could be applied to other DER, microgrids and hybrid renewable systems besides wind power. Potential future research includes hardware-in-theloop validation, adaptive learning in non-stationary grid conditions and integration with cybersecurity-aware edge architectures to increase reliability and deploymentreadiness. REFERENCES 1. Y. Liu, J. Wang, and H. Zhang, “Edge-computing enabled smart grid: architecture, key technologies and future trends,” IEEE Internet of Things Journal, vol. 6, no. 4, pp. 6956-6970, Aug. 2019. 2. C. Lin, S. Yu, Z. Cui, W. Xu and D. Gu, “Fog computing for industrial IoT: architecture, advances, and future challenges,” IEEE Transactions on Industrial Informatics, vol. 13, no. 11, pp. 46744684, Nov. 2017. 3. B. Biswal, S. Kumar and M. Yadav, “A systematic review of deep learning and machine learning techniques for smart grid load forecasting,” Energy Reports, vol. 12, pp. 3654-3670, 2024. 4. M. de Frutos, O. A. Marino, D. Huergo and E. Ferrer, “Deep reinforcement learning for multiobjective optimization: enhancing wind turbine energy generation while mitigating noise emissions,” Expert Systems with Applications, vol. 213, p. 119590, 2024. 5. M. Lara, A. Guerrero and J. Soriano, “Multiobjective optimization for simultaneously designing wind turbine power generation and vibration suppression,” Renewable & Sustainable Energy Reviews, vol. 202, p. 112179, 2023. 6. A. Vakhnin, “A novel multi-objective hybrid evolutionary-based algorithm for featurehyperparameter optimization in machine learning,” Algorithms, vol. 17, no. 4, p. 120, 2024. 7. J. Huang, S. Zhou, G. Li and Q. Shen, “Real-time monitoring and optimization methods for user-side energy management based on edge computing,” Scientific Reports, vol. 15, Article 24890, May 2025. 8. M. Altimania, H. Abdelsalam and M. Aziz, “Multiobjective optimization framework for renewable energy integration in smart grids with enhanced stability and resilience,” International Journal of Lifecycle and Costing Technologies, vol. 2, no. 2, pp. 827-846, 2025. 9. M. Zhou, J. Wang and K. Li, “Multi-objective prediction intervals for wind power forecasting using LSTM-based models,” Journal of Wind Engineering & Industrial Aerodynamics, vol. 213, p. 104860, 2021. 10. S. Shafiullah, R. Teng and Z. Li, “Deep CNN for harmonic detection and classification in power systems,” Electric Power Systems Research, vol. 186, p. 106431, 2020. 11. A. Ghasemi, A. Shojaeighadikolaei and M. Hashemi, “Combating uncertainties in wind and distributed PV energy sources using integrated reinforcement learning and time-series forecasting,” arXiv preprint arXiv:2302.14094, Feb. 2023. 12. H. M. Chung, S. Maharjan, Y. Zhang, F. Eliassen and T. Yuan, “Edge intelligence empowered UAVs for automated wind farm monitoring in smart grids,” arXiv preprint arXiv:2009.11256, Sep. 2020. 13. M. Ahmadi, “A comprehensive review of AI-driven approaches for smart grid real-time management,” Renewable & Sustainable Energy Reviews, vol. 150, p. 111736, 2023. 14. X. Liu, H. Lan, Q. Guan and C. Zhang, “LSTM networks for load and harmonic forecasting in distribution systems,” IEEE Access, vol. 7, pp. 146000-146011, 2019. 15. D. Farooq, M. Ayyub and A. Khan, “Communication delays in distributed energy management systems: impacts and mitigation,” Renewable Energy, vol. 128, pp. 193-203, 2018. 16. K. Zheng, X. Zhou, J. Xu and L. Yan, “Real-time harmonic forecasting using convolutional neural networks for industrial power grids,” IEEE Access, vol. 12, pp. 148 003-148 015, 2024. 17. H. Wang, Y. Li and Z. Chen, “Multi-objective genetic algorithm optimization for harmonic filter placement in distributed generation networks,” Electric Power Systems Research, vol. 226, p. 110592, 2024. 18. S. Patel, M. Kumar and A. Singh, “Edge-based deep learning for converter fault detection in wind energy systems,” Renewable Energy, vol. 199, pp. 345-357, 2023. 19. Q. Zhao, G. Wang and X. Wang, “A low-latency control framework for wind turbine converters via embedded deep reinforcement learning,” IEEE Transactions on Sustainable Energy, vol. 15, no. 1, pp. 234-245, Jan. 2024. 20. L. Sun, Y. Zhao, P. Wang and H. Li, “Real-time harmonic analysis and mitigation in smart grids using attention-based neural networks and FPGA implementation,” IEEE Transactions on Industrial Informatics, vol. 20, no. 5, pp. 3421-3432, May 2024. 21. J. Garcia, A. Lopez and V. Juarez, “Edge-AI for harmonic forecasting and mitigation in industrial feeder systems,” International Journal of Electrical Power & Energy Systems, vol. 153, p. 109267, 2023. 22. R. Mehta, S. Gupta and H. K. Saha, “Multi-objective optimization of wind farm power quality metrics “Low-Latency Control of Wind Energy Systems Using Edge-Deployed Deep Learning and Multi-Objective Genetic Algorithm Optimization for Real-Time Harmonic Forecasting in Industrial Grids” 8125 ETJ Volume 10 Issue 12 December 2025 , Adel Elgammal under harmonic disturbances,” Renewable & Sustainable Energy Reviews, vol. 183, p. 113988, 2023. 23. S. Guo, J. Zhang and L. Qin, “Low-latency edge computing for harmonic compensation in microgrids with high renewable penetration,” IEEE Transactions on Smart Grid, vol. 15, no. 4, pp. 2456-2467, July 2024. 24. E. Oliveira-Santos, P. Martins and F. Carvalho, “Real-time distributed genetic algorithm control of power converters in wind energy systems,” IEEE Access, vol. 11, pp. 72 950-72 964, 2023. 25. T. Kaiser, M. Zhao and Y. Lin, “Edge firmware optimization for high-frequency harmonic monitoring in wind-driven industrial grids,” Measurement, vol. 207, p. 112318, 2023. 26. F. Wang, Z. Li and J. Xu, “A hybrid deep learningMOGA framework for harmonic mitigation in distributed generation,” Electrical Engineering, vol. 106, no. 1, pp. 111-124, Jan. 2024. 27. B. Zhao, Y. Han and P. Zhang, “Real-time multiobjective harmonic suppression in converterinterfaced wind farms,” IET Renewable Power Generation, vol. 18, no. 2, pp. 133-145, Feb. 2024. 28. P. M. T. Bennett, H. Li and S. Smith, “Evolutionaryembedded edge computing architecture for power quality monitoring in smart industrial grids,” IEEE Transactions on Industrial Electronics, vol. 71, no. 5, pp. 4902-4913, May 2024. 29. Y. Feng, X. Du and L. Zhang, “Scalable harmonic control across multi-turbine wind farms using coordinated MOGA and edge computing,” Renewable Energy, vol. 215, p. 232-244, 2024. 30. D. Verma, K. Chaudhary and R. Srivastava, “Adaptive deep learning-based harmonic forecasting for wind turbines with real-time edge deployment,” Electric Power Components and Systems, vol. 51, no. 4, pp. 435-450, 2023. 31. L. Yu, X. Zhao and Y. Li, “Multi‐objective optimization method for control parameters of flexible DC transmission systems in real-time simulation systems,” IET Generation, Transmission & Distribution, vol. 18, no. 7, pp. 912-921, 2024. 32. J. Huang, S. Zhou, G. Li and Q. Shen, “Real-time monitoring and optimization methods for user-side energy management based on edge computing,” Scientific Reports, vol. 15, Article 24890, 2025. 33. A. Elhussiny, M. Ebied and Y. Abdel-aGoad, “BiLevel multi-objective optimization of harmonic filter parameters for PV rich grids,” Energy Reports, vol. 11, pp. 846-856, 2025. 34. M. Bošnjaković, D. Kovačević, N. Simić and I. Stanić, “Application of artificial intelligence in wind power systems: life-cycle review,” Applied Sciences, vol. 15, no. 5, p. 2443, 2025. 35. S. Subramanian, C. Sankaralingam, R. M. Elavarasan, R. R. Vijayaraghavan, K. Raju and L. Mihet-Popa, “An evaluation on wind energy potential using multi-objective optimization based non-dominated sorting genetic algorithm III,” Sustainability, vol. 13, no. 1, p. 410, 2021.