A Two-Stage Approach for Topology Change-Aware Data-Driven OPF

Abstract

Data-driven OPF has been widely studied recently to satisfy the real-time requirements of applications like economic dispatch, security analysis, etc. However, traditional data-driven models are typically trained for a specific system topology. When the system topology changes, the models must either be retrained (which demands a substantial amount of training data) or fine-tuned (which necessitates the selection of an appropriate pre-trained model). To this end, we propose a two-stage approach for topology change-aware data-driven OPF. It consists of: 1) generating data-driven models using a topology transfer framework; and 2) ensembling well-trained models. In Stage 1, GPR is employed to capture the nonlinear correlation between the new and predicted OPF data. The new data is obtained by solving the OPF problem using traditional optimization solvers under the new topology; the predicted data is obtained by inputting the same power demand into the data-driven OPF model trained on one of the historical datasets. This framework allows us to obtain sample-efficient topology transfer models. In Stage 2, a dynamic ensemble learning strategy is developed, where the weights and the topology transfer models that need to be ensembled are dynamically determined. This strategy allows us to avoid obtaining biased OPF solutions from sub-models. Numerical experiments on the modified IEEE 14- and TAS 97-bus test systems demonstrate that the proposed approach can obtain optimality-enhanced and equality function-satisfied OPF solutions as compared to other data-driven approaches.