Multi-step-ahead prediction of thermal load in regional energy system using deep learning method

Abstract

Accurate and reliable multi-step-ahead thermal load prediction is generally considered as the basis of model predictive control and multi-energy dispatch of building energy systems. Conventional approaches to future load time series prediction rely either on iterative one-step-ahead predictors or direct predictors, which ignores the temporal dependency between successive loads in time-delay building energy systems. This study, therefore, proposes a temporal attention encoder-decoder network (TA-EDN) model to improve the accuracy of multi-step-ahead thermal load prediction with the following three functional modules: long short-term memory (LSTM) network, which is to process the intrinsic temporal relationships among input and output variables, encoder-decoder network (EDN), which is to realize the multi-input multi-output modeling, and attention mechanism, which is to improve the ability of processing variables with long sequences. An actual regional energy system is selected to perform the 24-hour-ahead prediction using the proposed model as a validation experiment. The results suggest that the TA-EDN model significantly improves the prediction accuracy of the future thermal loads time series, achieving a mean absolute percentage error of 7.4%, compared to that of 9.1% of EDN model, 12.4% of LSTM-IS (a model combining LSTM with iterative strategy) and 12.7% of LSTM-DS (a model combining LSTM with direct strategy). In addition, compared with the ideal benchmark of multi-step-ahead prediction, the proposed TA-EDN model has room for improvement in the prediction of low load or large fluctuation load.