Distributed Pareto Reinforcement Learning for Multi-objective Smart Generation Control of Multi-area Interconnected Power Systems

Multiple objectives of smart generation control (SGC) are considered. The paper proposes a distributed Pareto reinforcement learning (DPRL) based on game theory to address the multi-objective control problem (MOCP) of SGC of multi-area interconnected power systems (MAIPSs). The proposed DPRL consists of the framework of reinforcement learning and multiple Q matrices, which can provide multiple outputs for MOCPs. As a control algorithm based on the Markov decision process, the stability of the proposed approach can be guaranteed with the ability to provide dynamic control strategies online. With the Pareto concept of optimization algorithms introduced into a control algorithm, the proposed approach can obtain multiple comprehensive objectives. The case studies under the two-area and practical four-area power systems show that: compared with four reinforcement learning methods and a proportional-integral controller, the proposed DPRL can achieve the highest control performance with multiple objectives in MAIPSs. The case studies of SGC under two MAIPSs verify the feasibility and effectiveness of the DPRL for MOCPs.