A probabilistic load forecasting method for multi-energy loads based on inflection point optimization and integrated feature screening

Aiming at the uncertainty and complexity of multi-energy load forecasting in integrated energy systems, this study proposes a hybrid Quantile Regression (QR) model based on Convolutional Neural Network (CNN)-Bidirectional Long Short-Term Memory (BiLSTM)-Time Step Attention. In the feature engineering phase, this model uses a method that combines Mutual Information and LightGBM for comprehensive variable analysis and feature screening to maximize the reduction of redundant features and avoid the omission of key feature variables. In the forecasting model, to address the problem of wide forecast ranges due to increased uncertainty at inflection points, an improved Quantile Regression Loss function is utilized to constrain the forecast intervals through a penalty mechanism, which narrows the forecast range while ensuring coverages. Furthermore, bayesian optimization is utilized to independently optimize the hyperparameters of the cold, heat, and electricity load respectively to match different load characteristics and improve the model training efficiency. The experimental results show that under the precondition of ensuring that the model forecast interval coverage is 100 %, when the Prediction Interval Nominal Confidence is 90 % and 95 %, the load prediction Average Interval Score of this model was reduced by an average of 36.74 %–51.48 % and 44.14 %–55.24 %, respectively.