Interpretable short-term wind power forecasting based on LIME algorithm and multi-source NWP

Short-term wind power forecasts are an essential support for daily grid dispatching and electricity market operations. Their accuracy directly affects system stability and market efficiency. Addressing the issue that single numerical weather prediction (NWP) models struggle to ensure accuracy across various temporal and spatial scenarios, a model based on multi-source NWP inputs is constructed to improve the forecasting accuracy, and considering that the multi-source input model will increase the black-box nature of the forecasting process, an interpretable method is proposed to reveal the influence of each input source on the forecast results. Firstly, a short-term wind power forecasting model based on multi-source NWP wind speed input is constructed by combining the attention mechanism and deep learning network. Then, based on the local interpretable algorithm LIME, a daily trust index is introduced to quantify and analyse the relationship between local inputs and outputs. The results show that the power forecast error obtained from the constructed multi-source input model is reduced by 1.30% on average compared to the single-source input, which also leads to a higher model black-box nature. The proposed interpretable method can quantitatively assess the contribution of each input NWP wind source to the prediction results.