Operational risk assessment of power distribution systems: A physics-informed multi-task learning approach

• The operational risk assessment method of distribution systems is formulated considering the flexibility characteristics of EVs and TCLs. • A physics-informed multi-task learning neural network approach is proposed for online, multi-dimensional, and feasible operational risk assessment. • Numerical results show that the proposed approach can achieve more rapid and accurate assessment of operational risks compared to purely data-driven models. Recently, the uncertainty of high proportion renewable energy and the operation of multiple flexible resources bring significant challenges to the power distribution system. Consequently, rapid operational risk assessment has become a critical problem, and data-driven methods provide a potential solution. However, it cannot satisfy constraints and predict multiple indices concurrently. To address this issue, this paper proposes a physics-informed (PI) multi-task learning (MTL) approach for rapid assessment of operational risks in the distribution system. Firstly, this paper models the operation of renewable and flexibility resources, and establishes an operational risk assessment index system to evaluate the model performance. Secondly, this paper introduces Long Short-term Memory (LSTM) with strong capability of extracting sequence and temporal dependency data to capture temporal dependencies in distribution system operations. Meanwhile, this approach employs multi-task learning paradigm for multiple related risk indices parallel prediction, thereby enhancing the prediction accuracy and generalization performance of the model. Furthermore, physics-informed neural network (PINN) imposes operational constraints as penalty terms in the loss function of data-driven models, ensuring it follow physical laws for more accurate prediction. Finally, the feasibility and effectiveness of the proposed approach is validated by case studies. The results demonstrate that the proposed method can realize the rapid operational risk assessment of the distribution system within less time. The prediction results of risk indices all meet the physical laws, and the prediction accuracy is improved by about 10%, compared with the purely data-driven model.