A data-driven interval forecasting model for building energy prediction using attention-based LSTM and fuzzy information granulation

Abstract

Quantifying uncertainties in the prediction of building energy consumption is critical to building energy management systems. In this study, a deep-learning-based interval forecasting model is developed by combining fuzzy information granulation, attention mechanism, and long short-term memory (LSTM) network for predicting building energy consumption and to present future uncertainties in the form of intervals. In particular, fuzzy information granulation theory is used for interval estimation, and attention-based LSTM is deployed in combination with automatic hyperparameter optimizer to provide interval prediction. A case study based on real building dataset is carried out for validation of the proposed model. Our study demonstrates that granulation window size is one of the most significant parameters that determines the quality of prediction intervals (PIs). The decrease of granulation window size would result in decrease in PI coverage probability. Moreover, the proposed interval forecasting model using attention-based LSTM is compared with the model using conventional LSTM. It is shown that the attention-based LSTM provides better interval forecasting performance than the conventional LSTM in terms of PI coverage probability, indicating that the attention mechanism has significant advantages to improve interval forecasting performance by increasing the efficiency with which the model uses information.