Forecasting very short-term power load with hybrid interpretable deep models

Very Short-Term power Load Forecasting (VSTLF) plays a key role in electricity planning and operational transactions. However, with the growing complexity and uncertainty of today’s electricity requirements, e.g. nonlinear correlations across electricity factors, unpredictable wavy trends and sudden changes, existing approaches struggle in forecasting performance and generalizing to production practice. Based on a public dataset and its baseline models, this study identifies the most significant attributes of power load forecasting upon data preprocessing and correlation analysis and then trains several end-to-end regression models based on the Bidirectional Long Short-Term Memory (BiLSTM) network and attention mechanisms. Experiment results demonstrate that the hybrid model based on Convolutional Neural Network (CNN) and BiLSTM outperforms several state-of-the-art solutions. Finally, we quantify the feature impacts using Shapley Additive Explanations (SHAP), which effectively improves our model’s interpretability.