Prediction of heating energy consumption with operation pattern variables for non-residential buildings using LSTM networks

Abstract The precise prediction of building energy consumption plays an important role in balancing the demand and supply of energy. In general, regular operation patterns occur for non-residential buildings, and the reflection of such operation patterns to models can play a positive role in predicting building energy consumption. In this study, three long short-term memory (LSTM) models were created, and the results were compared to examine the effect of applying the operation pattern data of a non-residential building on the prediction performance of the LSTM model. Each model exhibited different results according to the type of the input data used: (i) case 1 – building environmental data; (ii) case 2 – building environmental data and outdoor environmental data; and (iii) case 3 – building environmental data, outdoor environmental data, and operation pattern data. The building exhibited regular schedules by time and cycle according to the occupancy characteristics. When the prediction model utilized additional variables related to building operation as input data, it exhibited better performance than the other cases. The prediction performances in best case were 17.6% for CVRMSE and 0.6% for MBE. Applying insights into the change points at which the energy consumption pattern of the building changes was reflected to the model resulted in a higher pattern similarity than the other cases.