Interpretable multivariate wind speed forecasting using sliding masked window-based decomposition and deep autoregressive networks

As an increasingly important renewable energy source, accurate wind energy forecasting provides crucial decision support for the efficient scheduling and operation of wind farms. This study introduces an innovative short-term multivariate interpretable method for predicting wind speeds aimed at enhancing both the accuracy and interpretability of forecasts. The proposed model employs a novel feature selection method based on comprehensive relative importance analysis (CRIA) to filter the original meteorological variables and uses dynamic mode allocation (DMA) to determine each selected feature’s number of modes to be decomposed for a proposed sliding masked window-based decomposition, which is designed to avoid information leakage. The IVY algorithm is then used to optimize the hyperparameters of the deep autoregressive network (DeepAR) network, and a novel perturbation-based traceable post-hoc feature ranking (TPFR) mechanism is proposed to provide explainable results. In this paper, by comparison with multiple advanced feature screening algorithms such as XGBoost and Lasso, as well as multiple advanced prediction models such as TFT and Transformer, the proposed CRIA-MVMD-IVY-DeepAR-TPFR model has improved the prediction accuracy by at least 20% in terms of the MAPE value, comprehensively demonstrating the superiority of the model proposed in this paper. In addition, the interpretable results of this model and the experimental results of wind power conversion provide rich perspectives for the analysis of relevant data and offer analytical insights for the decision-making process of real wind power plants.