Heat Load Prediction for District Heating Systems with Temporal Convolutional Network and CatBoost

Abstract

Accurate heat load prediction is essential for heat production and refined management of district heating systems (DHSs). More advanced technology can often achieve more accurate forecasts. This paper suggests using temporal convolutional network (TCN) and categorical boosting (CatBoost) for heat load prediction. To test the performance of TCN and CatBoost in heat load prediction missions, two additional benchmark models, the decision tree model (DT) and the statistically based multiple linear regression (MLR), are built for comparison. A DHS in Tianjin, China, is used as the study case. Two historical operational characters (day-ahead heat load and hour-ahead heat load) and four meteorological characters (outdoor temperature, relative humidity, wind scale, and air quality index) are selected as input features for the models. The prediction results of every model on the test set are displayed and discussed. The experimental findings indicate that the prediction results of TCN and CatBoost are more accurate than the traditional prediction models, while the modeling process of CatBoost is simpler. Overall, TCN and CatBoost are potential heat load prediction methods.