Research on High-Precision Load Forecasting Model Based on Improved Metaheuristic Algorithms and Multi-Physics Coupling Feature Optimization

Abstract Accurate load forecasting is imperative for ensuring stable operation of power grids with high-penetration renewable energy. Targeting the dynamic load characteristics of the Zhangjiakou region, this study proposes a random forest framework (RF-AOA) integrating multi-physics feature engineering and an improved metaheuristic algorithm. Key innovations include: Multi-physics Coupling Mechanism Decoding: A four-dimensional temporal decomposition (hourly fluctuation/daily periodicity/weekly trend/monthly evolution) disentangles nonlinear interactions among temperature, wind speed, solar radiation, and humidity, first revealing solar radiation as the dominant driver (feature weight: 2.6) > hourly features (1.6) > temperature (1.3). Kent-Chaos Optimized AOA (KC-AOA): Chaotic initialization (entropy: 0.943) and adaptive exploration-exploitation balancing (23% higher exploration probability) achieve breakthrough hyperparameter optimization for random forests (decision tree depth: 200 layers). Validation using 11,000-hour data from Zhangjiakou demonstrates:RF-AOA attains R²=0.977 (12.3% improvement), MAE=0.057, RMSE=0.117；95% of prediction errors constrained within ±7.5%；37% accuracy boost under radiation mutation scenarios (>50 W/(m²·h))；Critical engineering value: Every 10% reduction in forecast error elevates regional dispatch efficiency by 18% and reduces peak-valley difference by 23%, significantly enhancing renewable energy utilization.