An integrated D-CNN-LSTM approach for short-term heat demand prediction in district heating systems

Abstract

Forecasting short-term heat demand is an integral function of district energy management applications. Although some well-known methods, such as support vector machine and artificial neural networks, can be employed, most of them require additional variables (such as temperature and humidity) in addition to the heat demand itself in order to make an accurate prediction. In this paper, a differencing-convolutional neural network-long short term memory (D-CNN-LSTM) approach is developed to forecast the heat demand half hourly ahead using only the historical heat data. Firstly, features extraction is performed to find a set of model inputs related to the dynamic behavior of heat consumption. This is followed by the design of D-CNN-LSTM to capture different seasonal patterns, in which the differencing aims to convert inputs to be stationary from non-stationary while the CNN-LSTM focuses on accurately predicting the future heat demand. Finally, various experiments are conducted to demonstrate the effectiveness and superiority of the designed method in comparison with existing algorithms.