A branch-adaptive framework for multi-timescale wind power forecasting: Optimization via hybrid evolutionary algorithm and physics-informed modeling

High-precision multi-timescale wind power forecasting is critical for grid security. However, existing unified architectures often struggle to simultaneously capture high-frequency transients in real-time horizons and long-term dependencies in day-ahead horizons, leading to feature interference and accuracy degradation. To address this challenge, this paper proposes a branch-adaptive framework optimized via a hybrid evolutionary algorithm and physics-informed modeling. Distinct from conventional static models, the proposed method constructs a seasonally adaptive three-branch architecture that decouples feature extraction by employing lightweight networks for real-time volatility and deep networks for day-ahead trends. Furthermore, physical constraints, including the Betz limit and turbine power curves, are explicitly embedded into feature engineering to ensure physical consistency, while a multi-stage evolutionary strategy is designed to prevent premature convergence in high-dimensional hyperparameter spaces. Following independent branch optimization, the framework achieves collaborative forecasting across timescales through global unified optimization and feature-layer fusion. Experiments on a 200 MW wind farm demonstrate superior performance. Real-time forecasting achieves an R 2 of 0.9971 in summer with RMSE reduced by up to 42.7% compared to benchmarks; meanwhile, day-ahead forecasting effectively suppresses long-horizon error accumulation, reducing RMSE by 17.6% against conventional baselines. This method provides a novel approach for addressing multi-timescale forecasting problems in non-stationary time series and significantly advances high-penetration wind power integration.