Reinforcement Learning Based Demand Response Approach Incorporating Thermostatically Controlled Loads and Battery Loads

Currently, most demand response (DR) algorithms are static, which don't allow for dynamic demand side management based on current states of load. To solve above questions, we proposed a deep reinforcement learning (DRL) based algorithm of demand response, mainly using Actor-Critic (AC) algorithm. Firstly, we model the non-responsive load, thermostatically controlled load and battery load in buildings. Then, we establish a deep reinforcement learning environment based on Markov reward process according to the load model of buildings. Then we train and update the parameter of actor and critic network with a dataset containing 36 buildings in Hubei province. Finally, we validate the effectiveness of the DR algorithm on a sub-dataset containing 9 buildings, resulting in a significant reduction in load volatility and load power ramp in the region. Electricity consumption in these 9 buildings was reduced by 6%. In general, the proposed DR algorithm has the ability to help power systems to achieve significant power-smoothing effectiveness to increase the operational efficiency of the main grid.