Machine learning-based multi-time-scale coupled equivalent modelling of active distribution networks under disturbances

The large-scale integration of renewable energy sources and adjustable loads has significantly increased the complexity and dynamic uncertainty of power systems, making accurate equivalent modelling of active distribution networks particularly challenging under disturbance conditions. Reliable equivalent models are essential for stability analysis, emergency control, and operational optimisation in modern power systems. To address the modelling requirements across multiple time scales and diverse operating scenarios, this paper proposes a multi-time-scale coupled equivalent modelling framework for active distribution networks under disturbance. For long-term modelling, a multi-attribute equivalent modelling approach is developed by integrating deep reinforcement learning with evidential reasoning, where composite models consisting of equivalent ZIP loads connected in parallel with equivalent induction motors are adopted as constituent structure. Furthermore, a multi-attribute reward mechanism is further designed to guide parameter identification while preserving model interpretability under uncertainty. For short-term modelling, a clustering-based load aggregation strategy is employed to construct high-precision equivalent models for adjustable loads. Meanwhile, a Bidirectional Long Short-Term Memory network is introduced to capture fast-varying dynamic behaviours. Case studies conducted demonstrate that MDDQN outperforms conventional DDQN and DQN, while Bidirectional Long Short-Term Memory outperforms conventional Long Short-Term Memory. The results indicate that the proposed approach provides an effective and reliable modelling solution, offering model-based support for the analysis and operation of active distribution networks under disturbance conditions. • A multi-time-scale equivalent modelling framework for active distribution networks under disturbances • Integration of evidential reasoning to address weighting uncertainty and enhance model interpretability • A parameter identification method using Modified Double Deep Q-Network and Bidirectional Long Short-Term Memory.