Advancing energy system optimization via data-centric task-oriented forecasting: An application in PV-battery operation

Forecasting is pivotal for integrating stochastic resources into energy system operations. Regular forecasting models prioritize accuracy, assuming that more accurate forecasts yield superior performance in downstream forecast-based operational tasks. However, recent studies question this assumption and advocate task-oriented learning (TOL), which tailors the training of forecasting models to directly maximize downstream task performance rather than aiming for optimum accuracy. The core challenge of TOL lies in compatibly embedding the model of operational tasks in the training of forecasting models. Existing solutions are model-centric, namely, re-designing the training loss function or the task model formulations. In this paper, we first propose a generic data-centric TOL framework that redirects the focus from model advancement to data enhancement. The proposed framework iterates the training data to maximize downstream task performance, thereby avoiding the incompatibility in connecting forecasting models and tasks. Additionally, a framework based on scenario decomposition is proposed to further improve data-centric TOL performance. Both frameworks are validated through a proof-of-concept study, focusing on the photovoltaic-battery energy storage system (PV-BESS) operation in the day-ahead market. Comparative studies verify that the proposed TOL frameworks substantially enhance the performance of forecasts in a downstream task. The proposed frameworks can be extended to other applications readily.