Abstract: [Objectives] With the widespread adoption of distributed new energy sources and flexible loads, the complexity of power interactions between the active distribution network (ADN) and the backbone grid has increased significantly. Traditional load models usually rely on given structures and parameters, which cannot adapt to the time-varying and stochastic nature of loads. Therefore, a dynamic equivalence modeling method based on hierarchical double deep Q network (HDDQN) for ADN is proposed. [Methods] This paper first performs preliminary modeling of ADN through multiple dynamic and static load models, then abstracts the corresponding task into a two-stage, three-layer Markov decision process, and finally combines it with the HDDQN algorithm to achieve online identification of parameters. The HDDQN algorithm employs gated recurrent units to perform feature extraction on the input discrete time series data, and introduces a competitive network and a prioritized empirical replay mechanism to optimize the traditional reinforcement learning algorithm is optimized. [Results] The proposed method can generate strategies in only 2.8 s after being well-trained, with its average active and reactive power calculation accuracy 2.64 times and 9.10 times that of the traditional methods, respectively. [Conclusions]The proposed HDDQN algorithm can identify the equivalent model parameters online, and the convergence performance and equivalence accuracy are significantly better than that of traditional methods, which is of great significance to improve the efficiency and accuracy of real-time modelling of ADN.

Key words: active distribution network (ADN), new energy, distributed generation, flexible load, equivalent modeling, parameter identification, deep reinforcement learning, hierarchical decision