Demand response considering user behaviour differences for load serving entity: A multi-agent deep reinforcement learning approach

Demand response can reshape the load profiles by adjusting the energy price to improve overall performance of power grids. For the load serving entities (LSEs), the optimal demand response shall satisfy operation constraints and gain maximised profits. However, the individual end-consumer will surely value electricity differently, and price adjustments without considering the differences of user behaviours may induce unbalanced dispatch and degrade the overall profits. Given affordable price ranges under which the non-critical loads can properly response to price signals, the demand response can be operated in a manner where non-critical loads at different areas coordinately adjusted to achieve the global dispatch objectives. For this purposes, this paper proposes a cooperative demand response approach for LSE pricing based on a multi-agent deep reinforcement learning approach. The learning approach constructs the distributed pricing agents that cooperatively determine the price signals to be responded by different load clusters. To update the parameters of pricing agents, the deep deterministic policy gradient is derived considering the constraints and different behaviours of load clusters. The effectiveness of the proposed cooperative demand response method is verified based on numerical simulations.