Multi-task learning for solving OPF in an evolving environment

Optimal Power Flow (OPF) aims to minimize operating cost subject to AC network constraints, but its nonconvex, nonlinear nature makes the problem NP-hard, which in practice motivates solving simplified models or using conventional nonlinear solvers; both approaches typically fall short of real-time, high-quality AC-feasible solutions. By leveraging universal approximation ability with fast inference, modern machine-learning methods have become a promising direction for OPF surrogates. Despite the promising results achieved by existing methods, they typically focus on static scenarios without dataset distribution shifts. In contrast, practical power systems operate in an evolving environment where scenarios can change dynamically (e.g., generator outages), causing existing methods to fail to provide accurate OPF solutions. By proposing a multi-task learning framework, we aim to expand the application region of data-driven OPF methods. Specifically, to address the weak correlation across different scenarios, we model multiple scenarios simultaneously through shared network parameters in the proposed framework, with a hypernetwork facilitating knowledge transfer. Meanwhile, to address dataset imbalance under dynamic scenarios, we introduce an error-focused up-sampling method that resamples data exhibiting large deviations from the pre-trained model’s predictions. Furthermore, to balance the training process, we adopt an adaptive-weight algorithm that assigns trainable weights to each sample, updated alongside the hypernetwork weights in a dual ascent manner. Simulation results based on the 14-bus system and 118-bus system show that the proposed framework can provide optimality and feasibility-enhanced OPF solutions in near real time compared to existing methods with or without a post-processing step.