Abstract: Microgrid clusters can enhance the utilization rate of renewable energy and system reliability, but their traditional dispatching methods have problems such as insufficient flexibility and high computational complexity. To address this issue, this paper proposes a two-layer collaborative optimization dispatching method for microgrid clusters based on an improved SAC algorithm. Firstly, a two-layer collaborative optimization dispatching model for microgrid clusters is constructed, aiming to minimize the operation cost, where the upper layer is responsible for the global power optimization dispatching of the microgrid cluster, and the lower layer is responsible for the local power optimization dispatching of each sub-microgrid. Secondly, an improved SAC algorithm is proposed, which introduces a variational autoencoder (VAE) model into the traditional SAC algorithm to utilize the feature extraction ability of VAE to reduce the dimension of high-dimensional states and extract key features, thereby reducing the computational complexity of the algorithm. Then, a time decay factor is introduced on the basis of the prioritized experience replay mechanism to increase the adoption frequency of later-entered experiences, further improving the learning efficiency of the algorithm. Finally, the improved SAC algorithm and the alternating direction method of multipliers are used to optimize and solve the upper and lower layer models of the microgrid cluster respectively. Through the interaction mechanism of the upper and lower layers, efficient collaborative dispatching and economic operation of the microgrid cluster are achieved. Simulation results show that the proposed method outperforms Scheme 1, Scheme 2 and Scheme 3 in terms of total operation cost, load balance rate and energy utilization rate. To be exactly, the operating costs of the proposed method in this paper are reduced by14.0%, 10.2% and 6.4%, respectively; the load balance rate is increased by 9.3 percentage points, 6.4 percentage points and 3.6 percentage points, respectively; and the energy utilization rate is raised by 10.8 percentage points, 7.6 percentage points and 2.9 percentage points respectively. This result verifies the superiority of the proposed method in the optimal dispatch of microgrid clusters. Meanwhile, this method also shows good robustness in dealing with the prediction errors of wind and solar power generations.

Key words: intelligent grid, microgrid clusters, optimal scheduling, soft actor critic, alternating direction method of multipliers, variational autoencoder