Collaborative, parallel Monte Carlo Tree Search for autonomous electricity demand management

Abstract

Balancing electricity supply and consumption is critical for the stable performance of an electricity Grid. Demand Side Management (DSM) refers to shifting consumers' energy usage to off-peaks as much as possible to avoid more electricity demand than available supply during peak times. Artificial intelligent planning algorithms have been applied to enabling electric devices to reschedule their operation to off-peak. One such algorithm is Monte Carlo Tree Search (MCTS), which takes advantage of tree search and random sampling on decision space in order to fine an optimal domain decision. In particular for DSM, MCTS has been used to control both smart meters and also electrical devices, In both applications, MCTS acts as a centralized consumption planner choosing the optimum approach for all devices in the space. Centralized computation thread limits these approaches in terms of flexibility and scalability. as applied in domains outside DSM, an alternative, decentralized MCTS algorithm, called Parallel MCTS (P-MCTS), allows every agent to run its independent MCTS thread to have its own solution. In thios paper, frist we studied the feasibility of applying P-MCTS in DSM to make demand planning more flexible and scalable. P-MCTS has been evaluated by running two different scenarios one with 6 electrical vehicles (EVs) and the other with 90 EVs which roughly results in 30% peak load shifting for P-MCTS. To improve the results of P-MCTS, we propose a new decentralized collaborative approach, called Collaborative P-MCTS (CP-MCTS), which exploits and extends P-MCTS to enable each electrical device to actively affect the planning process but also to improve the final decision using collective knowledge obtained during the collaboration. Additionally, in comparison with P-MCTS, CP-MCTS obtained better results, including more peak load shufting and smoother load curve due to 30% lower Peak to Average Ratio (PAR).