A multi dataset validation model for hybrid feature selection in wind energy maximum power point tracking systems.

Wind energy Maximum Power Point Tracking (MPPT) systems face computational challenges due to high-dimensional sensor data, frequently exceeding 400-500 monitoring variables in modern turbine installations. This study presents a two-stage hybrid Feature Selection (FS) method that balances prediction accuracy with computational efficiency for wind energy MPPT applications. The methodology combines statistical filtering based on mutual information (MI) for initial dimensionality reduction with an Adaptive Multi-Objective Binary Harmony Search (AMO-BHS) optimization algorithm. The algorithm incorporates adaptive parameter control and multi-objective optimization to generate multiple deployment options representing different accuracy-efficiency trade-offs. Validation across 3 distinct datasets-Kelmarsh Wind Farm (464 features, 280,743 records), laboratory experimental data (5 features, 1,503 records), and VV Wind Farms (87 features, 101,644 records)-validates robust performance across utility-scale, laboratory, and tropical environments. The proposed method achieves dimensionality reduction of 76.9-87.5%, reducing feature sets from 464 to 58-97 (Kelmarsh) and from 87 to 8-21 (VV). Feature subsets are validated through supervised learning tasks predicting active power output for utility-scale turbines (Kelmarsh, VV) and rotor speed for laboratory systems (RAC) using Random Forest (RF) regression. The approach achieves RMSEs of 0.2567 kW for Kelmarsh active power prediction, 0.2467 kW for VV active power, and 0.1234 RPM for RAC rotor speed tracking, representing 9.4-14.7% improvements in accuracy over full-feature baselines (0.2834 kW, 0.2891 kW, 0.1456 RPM). Compared to standard FS methods, this method reduces RMSE by 20.6-28.6% over MI ranking alone and 25.7-34.7% over LASSO regularization. Statistical validation confirms reliability with p-values less than 0.001, effect sizes of 0.76-0.89, and a coefficient of variation below 8% across cross-validation. These improvements translate to tangible operational benefits: dimensionality reduction from 464 to 58 features enables real-time MPPT control on resource-constrained SCADA hardware while maintaining accuracy, reducing computational overhead by 87.5% without sacrificing predictive performance. The compact feature subsets decrease sensor dependency and maintenance costs, improve model resilience to sensor failures, and enable faster model updates during operational deployment. For utility-scale wind farms, even modest RMSE improvements (9.4-14.7%) correspond to significant energy-generation optimization-a 10% increase in prediction accuracy can translate into a 2-3% annual increase in energy production through more accurate MPPT control, representing a substantial income impact across multi-turbine installations. The two-stage network provides computationally efficient solutions suitable for operational MPPT systems with 10 min control cycles, delivering multiple optimal FS that enable practitioners to select deployment options based on their specific accuracy requirements and computational constraints.