A Heat Load Prediction Method for District Heating Systems Based on the AE-GWO-GRU Model

Abstract

Accurate prediction of the heat load in district heating systems is challenging due to various influencing factors, substantial transmission lag in the pipe network, frequent fluctuations, and significant peak-to-valley differences. An autoencoder—grey wolf optimization—gated recurrent unit (AE-GWO-GRU)-based heat load prediction method for district heating systems is proposed, employing techniques such as data augmentation, lag feature extraction, and input feature extraction, which contribute to improvements in the model’s prediction accuracy and heat load control stability. By using the AE approach to augment the data, the issue of the training model’s accuracy being compromised due to a shortage of data is effectively resolved. The study discusses the influencing factors and lag time of heat load, applies the partial autocorrelation function (PACF) principle to downsample the sequence, reduces the interference of lag and instantaneous changes, and improves the stationary characteristics of the heat load time series. To increase prediction accuracy, the GWO algorithm is used to tune the parameters of the GRU prediction model. The prediction error, measured by RMSE and MAPE, dropped from 56.69 and 2.45% to 47.90 and 2.17%, respectively, compared to the single GRU prediction approach. The findings demonstrate greater accuracy and stability in heat load prediction, underscoring the practical value of the proposed method.