Abstract

Decision making is generally modeled as sequential forecast-decision steps with no feedback, following an open-loop approach. For instance, in the electricity sector, system operators use the forecast-decision approach followed by ad hoc rules to determine reserve requirements and biased net load forecasts to guard the system against renewable generation and demand uncertainty. Such procedures lack technical formalism to minimize operating and reliability costs. We present a new closed-loop framework, named application-driven learning, in which the best forecasting model is defined according to a given application cost function. We consider applications in which the decision-making process is driven by two-stage optimization schemes fed by multivariate point forecasts. We present our estimation method as a bilevel optimization problem and prove convergence to the best estimator regarding the expected application cost. We propose two solution methods: an exact method based on the KKT conditions of the second-level problems and a scalable heuristic suitable for decomposition. Thus, we offer an alternative scientifically grounded approach to current ad hoc procedures implemented in industry practices. We test the proposed methodology with real data and large-scale systems with thousands of buses. Results show that the proposed methodology is scalable and consistently performs better than the standard open-loop approach.