Real-Time Price Forecasting Combined Deep Reinforcement Learning for Predictive Home Energy Management System

Owing to the consistent increase in residential energy consumption, novel intelligent energy management strategies that efficiently schedule household appliance consumptions must be developed to reduce energy costs. However, developing such strategies is challenging owing to the uncertainties in future real-time prices (RTPs), various characteristics of user behavior, and appliance demand. Thus, this study proposes a new predictive home energy management system (PHEMS) based on an efficient RTP forecasting model combined deep reinforcement learning (DRL) algorithm to provide optimal energy consumption recommendations for end-users. First, to overcome the risk associated with highly complex RTPs, a new bidirectional temporal convolution-based Transformer is employed to obtain the predicted interval (PI) of RTPs and provide considerable potential for the DRL agent to learn the optimal control policy. Subsequently, a novel dynamic action–based twin delayed deep deterministic policy gradient combined with the PI of RTPs is proposed to solve the home energy management problem and adaptively adjust the operating time and power of household appliances efficiently. Experimental results demonstrate the superiority of the proposed PHEMS in terms of high RTP prediction accuracy $R^{2}$ of 0.94, effectively balanced positive and negative biases, enhanced learning performance, and significant reductions in electricity cost to $ {\$} $ 3.11 and dissatisfaction cost to 1.45, thereby outperforming other benchmark algorithms for end-users.