A spatial-temporal layer-wise relevance propagation method for improving interpretability and prediction accuracy of LSTM building energy prediction

At present, data-driven methods have achieved satisfactory results in building energy consumption pre-diction, especially deep learning models such as long short-term memory (LSTM). However, complex deep learning models struggle to achieve acceptable interpretability, preventing building professionals from understanding the models and reducing their trust in them. Aiming at the problem of poor model interpretability, this study developed a feature-level spatial-temporal layer-wise relevance propagation (ST-LRP) method with a firm explanation by improving layer-wise relevance propagation (LRP). This method quantitatively obtains the correlation of multiple input feature data to the energy consumption prediction results in both spatial and temporal dimensions. Through the size of the correlation value, the characteristics of both spatial and temporal dimensions can be screened out and explained in combina-tion with expert knowledge. The energy consumption data of a building from the open-source data set building data genome project 2 is verified in the prediction horizon of 1 similar to 24 h. Results show that ST-LRP can effectively identify spatial-temporal correlations between input data and energy consumption predictions. In the spatial dimension, ST-LRP can identify the correlation values of time-based features such as hour and week day, which are second only to building energy consumption, and the coefficient of variation of the root mean squared error (CV-RMSE) can be reduced by 7.17 % when the least correlated spatial dimension feature is removed, while in the temporal dimension, the CV-RMSE can be reduced by 0.87 % when the least correlated temporal dimension feature is removed. (c) 2022 Elsevier B.V. All rights reserved.