Multi-Time-Scale Heating Load Prediction Using Attention-Enhanced LSTM with Improved Adam Optimization

Abstract

This paper proposes an extended short-term memory network (LSTM) heating load prediction model that integrates an attention mechanism and an improved adaptive moment estimation (Adam) algorithm. The model dynamically focuses on key influencing factors such as outdoor temperature and user behavior through the attention mechanism, and combines the improved Adam algorithm to optimize the parameter update process. The experiment uses the heating data of a city for 12 consecutive months, divides the training set and the test set into 7:3, and compares them with traditional LSTM, ordinary Adam-optimized LSTM, SVM, Transformer, TCN, and CNN-LSTM hybrid models. The results show that the root mean square error (RMSE) of the improved algorithm on the test set is 10.23, which is 31.6% lower than that of the traditional LSTM; the mean absolute error (MAE) is 8.12, which is 29.4% lower; the mean absolute percentage error (MAPE) is 7.2%, which is 25.8% lower. At the same time, in short-term (1–24 hours), medium-term (1–7 days), and long-term (1–30 days) prediction tasks, the predicted values closely follow the observed load curve, and the generalization ability is significantly enhanced.