Integrated Learning and Optimization to Control Load Demand and Wind Generation for Minimizing Ramping Cost in Real-Time Electricity Market

Abstract: We developed a new integrated learning and optimization (ILO) methodology to predict context-aware unknown parameters in economic dispatch (ED), a crucial problem in power systems solved to generate optimal power dispatching decisions to serve consumer load. The ED formulation in the current study consists of load and renewable generation as unknown parameters in its constraints predicted using contextual information (e.g., prior load, temperature). The ILO framework train a neural network (NN) to estimate ED parameters by minimizing an application-specific regret function which is a difference between ground truth and NN-driven decisions favouring better ED decisions. We thoroughly analyze the feasible region of ED formulation to understand the impact of load and renewable learning together on the ED decisions. Corresponding to that we developed a new regret function to capture real-time electricity market operations where differences in predicted and true loads are corrected by ramping generators in real-time but at a higher cost than the market price. The proposed regret function when minimized using ILO framework train the NN to guide the load and renewable predictions to generate ED decisions favouring minimum generator ramping costs. This is unlike conventional sequential learning and optimization (SLO) framework which train NN to accurately estimate load and renewable instead of better ED decisions. The combined training of load and renewable using ILO is a new concept and lead to significantly improved ramping costs when compared with SLO based training of load and renewable and SLO trained load with 100% accurate renewable proving its decision-focused capability.