Improving energy consumption forecasting using DLnet based on periodic modeling

Abstract:
Improving the prediction accuracy of energy consumption in office buildings is necessary to achieve high energy efficiency in smart buildings. The existing forecasting methods rarely analyze the periodic characteristics of energy consumption independently. In this paper, a short-term office building energy consumption prediction model (DLnet) is proposed to address the problem of inefficiency in the utilization of periodic energy consumption data.Firstly, the period component of the energy consumption data is decomposed using STL, and the optimal period of the energy consumption data is searched for by a grid search algorithm, and then the Periodic block is constructed based on the optimal period; Secondly, the Time-series block data is constructed according to the data shape of the Periodic block; then the Time-series block data and the Periodic block data are trained and learned using LSTM; Finally, the prediction results of the Time-series block data and the Periodic block data are fused by linear regression.The four prediction accuracy indicators of the proposed model have been demonstrated to be 7%, 21%, 25% ,and 26% higher than those of the LSTM model.