Physics-Informed Adaptive-Weight NBeatsx for Short-Term Wind Power Forecasting

Accurate and physically interpretable wind power forecasting (WPF) is crucial for ensuring the reliable operation of power grid systems. Wind turbines have operational characteristics significantly influenced by complex environmental factors, such as wind speed fluctuations and intermittency, posing challenges for precise wind power modeling. Although deep learning models have become a promising data-driven solution in WPF, their common “closed-box” nature makes it difficult to balance forecast accuracy with the rationality of physical mechanisms. Therefore, based on the neural basis expansion analysis (NBEATSx) network architecture, this article proposes a multistep WPF method, named physics-informed adaptive-weight NBEATSx. This method realizes the deep integration of physical prior knowledge and data-driven models, providing a novel technical path for solving the joint optimization problem of accuracy and interpretability in WPF. The operational constraints of wind turbines, such as cut-in, rated, cut-out wind speeds, and rated power, are explicitly embedded into the network structure. A dynamic trainable weighting mechanism is leveraged for stack outputs, instead of the traditional aggregation strategy of direct summation. The experimental results based on a dataset of a 2-MW wind turbine show that the proposed method is significantly superior to the benchmark models and NBEATSx variants in terms of forecast accuracy and robustness.