A complete consumer behaviour learning model for real-time demand response implementation in smart grid

Accurate and optimal implementation of Demand Response (DR) programs essentially requires knowledge of occupants’ behavioral aspects regarding power usage. Maintaining consumers’ comfort has become an imperative component along with cost reduction; there is utmost need to understand their power consumption trends completely. In this paper, a complete solution regarding consumer behavior learning has been presented for designing efficient demand response algorithms. Firstly, appliance-level power forecasting has been performed using deep learning ensemble model: CNN-LSTM and XG-boost; Secondly, dynamic itemset counting (DIC), a variant of the Apriori algorithm, has been utilized to generate association rules which determine appliance-appliance association and discovery. In this way, all the aspects of the dynamic and non-stationary nature of appliances’ power time series have been addressed for DR implementation. Using two benchmark datasets, it has been demonstrated that the proposed approach exhibits enhanced performance in comparison to other competitive models in terms of RMSE and MAE.