Developing a hybrid time-series artificial intelligence model to forecast energy use in buildings

Abstract

Abstract The development of a reliable energy use prediction model is still difficult due to the inherent complex pattern of energy use data. There are few studies developing a prediction model for the one-day-ahead energy use prediction in buildings and optimizing the hyperparameters of a prediction model is necessary. This study aimed to propose a hybrid artificial intelligence model for forecasting one-day ahead time-series energy consumption in buildings. The proposed model was developed based on the integration of the Seasonal Autoregressive integrated Moving average, the Firefly-inspired Optimization algorithm, and the support vector Regression (SAMFOR). A large dataset of energy consumption in 30-min intervals, temporal data, and weather data from six real-world buildings in Vietnam was used to train and test the model. Sensitivity analyses were performed to identify appropriate model inputs. Comparison results show that the SAMFOR model was more effective than the others such as the seasonal autoregressive integrated moving average (SARIMA) and support vector regression (SVR), SARIMA-SVR, and random forests (RF) models. Evaluation results on real-world building depicted that the proposed SAMFOR model achieved the highest accuracy with the root-mean-square error (RMSE) of 1.77 kWh in, mean absolute percentage error (MAPE) of 9.56%, and correlation coefficient (R) of 0.914. The comparison results confirmed that the SAMFOR model was effective for forecasting one-day-ahead energy consumption. The study contributes to (1) the knowledge domain by proposing the hybrid SAMFOR model for forecasting energy consumption in buildings; and (2) the state of practice by providing building managers or users with a powerful tool for analyzing and improving building energy performance.