Interpretable Wind Power Forecasting with Feature and Loss Function Construction Guided by Domain Knowledge

Current wind power forecasting (WPF) methods predominantly rely on data-driven deep learning models, inherently lacking integration of domain knowledge, which limits forecasting accuracy and model interpretability. To address this issue, an interpretable data-knowledge fusion ultra-short-term WPF model is proposed, in which domain knowledge guides the feature construction process and is directly embedded into the design of the loss function. In the feature construction module, a wind speed-power curve is established to generate theoretical power outputs by using historical wind speed as inputs. These theoretical outputs, combined with historical measured data, serve as inputs for the model. In the loss function construction module, a boundary constraint loss is designed by extracting the upper and lower boundaries of wind power output using the alpha shape algorithm and Local Weighted Linear Regression based on wind speed and power data. Notably, the parameters of boundaries are dynamically updated to capture the volatility of wind power. Additionally, an error distribution shape loss is introduced to penalize the deviation of the training error distribution from the normal distribution, using Jensen-Shannon divergence as an indicator. Case studies across 30 wind farms demonstrate that the proposed method guided by interpretable knowledge achieves the best average performance across all time horizons compared to baseline models. The method also shows strong robustness in noise and missing data experiments.