Multi-agent reinforcement learning based distributed smart grid load balancing scheduling algorithm design

This study addresses the challenges of high-dimensional nonlinear coupling in distributed smart grids, such as voltage instability and load imbalance caused by renewable energy fluctuations. A multi-agent reinforcement learning (MARL) based distributed scheduling algorithm is proposed. It constructs a system with multiple intelligent energy hubs, where each agent manages a specific device class using local observations to avoid the dimensionality issues of centralized approaches. Dynamic conversion factors are introduced to handle renewable energy fluctuations and dynamic electricity prices, forming two complementary optimization strategies. Furthermore, based on the four elements of MARL algorithm, an algorithm is designed to solve the optimal adjustment strategy. Each intelligent body is responsible for a class of devices, and each intelligent body observes the local state to avoid the dimensional disaster and global state space explosion of centralized load scheduling. A multi-objective load balancing scheduling model is constructed for the power grid with the objectives of minimizing network loss, minimizing voltage offset, and achieving optimal load balance. An improved hybrid multi-objective particle swarm optimization algorithm based on R2 index and decomposition strategy is employed to solve the problem. An R2 index is used to screen the particle swarm, the global extremum is randomly selected, the individual extremum is updated using the Penalty-based Boundary Intersection (PBI) decomposition strategy, and elite and Gaussian learning strategies are combined to help particles escape local optima and achieve balanced scheduling. The experiment shows that this method reduces the fluctuation of power grid output and significantly improves network loss and load balancing indicators.