Sequence-to-sequence deep learning model for building energy consumption prediction with dynamic simulation modeling

Abstract

This study used a more detailed dataset approach to address the limitations of existing building energy prediction methods and to predict building energy demand more accurately. The EnergyPlus dynamic simulation program closely modeled building envelope performance, zone division, and heating, ventilation, and air conditioning systems. Building energy simulation used actual weather data and generated occupancy data. The occupancy, lighting, and equipment schedules for each zone were generated in 5-min intervals using the Lawrence Berkeley National Laboratory occupancy simulator. Summer electric power consumption based on Internet of Things information from the testbed building validated the model. When applying generated occupancy, lighting, and equipment schedules, the mean bias error of simulation similarly improved to 4.73%, and the coefficient of variation of the root-mean-squared error (Cv(RMSE]) improved to 12.26%. Subsequently, a demand prediction model was constructed as a sequence-to-sequence (seq2seq) model using long short-term memory (LSTM) cells in recurrent neural network (RNN) algorithms, then its accuracy was evaluated. In the seq2seq model, the learning performance based on the EnergyPlus data exhibited an RMSE of 4.48% and a weighted average percentage error of 3.07%. As a result of applying prediction methods while changing climate scenarios, prediction performance also satisfied statistically significant levels. The occupancy information and solar radiation were determined to exert the greatest influence on the building energy demand prediction.