https://www.mdu.se/

Heat demand prediction is a notable research topic in intelligent energy networks (IENs), due to the rapid growth of heat demand in cities. Given that hourly heat demand data can be considered as a time series data recording and well analyzed by time series seasonal decomposition algorithms, we develop a variant of the recurrent neural network (RNN), namely time frequency-domain memory (TFDM). The TFDM combines fast Fourier transform (FFT) and long short-term memory (LSTM) model to preserve memory of the series in both time and frequency domains, and cascades a residual block to introduce the impact factors (e.g., weathers). In the experiments, we compare the proposed TFDM with various referred methods on a heat demand dataset. The experimental results show that the proposed TFDM has significant performance improvement in the heat demand prediction.