Short-to-Mid-term wind power forecasting is essential for grid dispatching and energy market operations. However, existing methods present challenges in terms of physical consistency, interpretability and computational efficiency. Traditional physical models have coarse resolution and systematic biases, and data-driven approaches lack integration of fundamental aerodynamic principles. This paper therefore proposes a physics-informed back propagation neural network (PIBPNN), which incorporates governing wind turbine constraints, such as Betz's limit, power coefficient boundaries, and mechanical operating laws, as differentiable regularization terms within the loss function. The framework incorporates physics-compliant data pre-processing for anomaly filtering and feature enhancement, a lightweight neural architecture and a hybrid training mechanism. This transforms conventional black-box predictions into physically grounded grey-box outputs. When validated using operational SCADA data from industrial wind turbines, the method is found to significantly outperform state-of-the-art benchmarks in terms of accuracy, robustness, and computational efficiency.

Physics-informed neural network,Wind Power Forecasting,aerodynamic constraint,backpropagation neural network