Ultrashort-Term Heat Load Prediction for Heat Exchange Stations Based on Dual-Source Data Feature Fusion

Abstract

Accurate load prediction plays a vital role in the field of district heating. However, it is challenging to establish an accurate mathematical model for heat load prediction because of the technical characteristics of heating systems, such as their large-scale, multidimensional, long delay, strong coupling, time-varying, and nonlinear characteristics. This article proposes a heat load prediction model based on feature fusion of dual-source data (DSDF), and this model fully extracts the characteristics of the working condition data of pipe networks and the data of building energy consumption to make high-precision predictions of the ultrashort-term heat load of heat exchange stations. First, to address the problem of incomplete feature coverage of a single-mode data set, the network working condition data set and building energy consumption data set were constructed, and the random forest method was used to select characteristic variables. Second, a neural network prediction model was constructed. The model captures short- and long-term features through the convolution layer and cyclic layer and captures periodic features through the cyclic jump layer. Finally, the scoring idea of the attention mechanism is introduced, and the predicted results of energy consumption data were used to calibrate the predicted results of pipe network working condition data to improve the model’s prediction performance. Comparative experiments with several algorithms showed that the proposed new method performs well in the task of ultrashort-term prediction of heat load.