What is it about?
Microgrids need forecasts of wind power, solar generation, and electricity demand to decide when to generate, store, buy, sell, or shift energy. Yet a forecast with a smaller statistical error does not always lead to a cheaper operating decision. This study develops an economic feedback-guided forecasting and scheduling framework that trains forecasts according to how useful they are for downstream rolling operation. A modified encoder-only Informer produces day-ahead forecasts of wind power, photovoltaic generation, and load, which are supplied to a fixed rolling model predictive control scheduler. The operating costs of the actions actually implemented by the scheduler are then fed back to fine-tune the forecasting model. Because the mixed-integer scheduler is not directly differentiable, simultaneous perturbation stochastic approximation is used to estimate how forecasting-model changes affect economic outcomes. Importantly, the MPC decision structure, constraints, and solver remain unchanged. The framework also coordinates batteries, electric vehicles, a microturbine, grid exchange, and time-shiftable demand, and is evaluated using measured microgrid data from northern China.
Featured Image
Photo by Michael Förtsch on Unsplash
Why is it important?
Most practical forecasting-and-scheduling systems optimize forecasting accuracy first and operational decisions second, even though forecasting errors at different times can have very different economic consequences. Errors during high-price periods or near operational constraints may change charging, generation, or grid-purchase decisions and cause disproportionately large costs. This work addresses that mismatch by letting the economic consequences of actual rolling decisions influence how the forecasting model is trained, while preserving the existing MPC scheduler. On a representative day, the proposed framework reduced implemented-action operating cost by 6.79% compared with the conventional baseline while maintaining comparable forecasting accuracy. In a separate 15-day ex post evaluation under the same observed wind–PV–load conditions, it achieved lower settled costs on 13 of 15 days and reduced average daily cost from ¥7035.07 to ¥6946.53. These results show that the value of an energy forecast should be judged not only by prediction error, but also by the quality of the operational decisions it enables.
Perspectives
What I find particularly valuable about this work is the shift from asking only “How accurate is the forecast?” to asking “How useful is the forecast for the decision that follows?” From my perspective, this is especially relevant for real energy-management systems, where forecasts are rarely an end in themselves and their practical value depends on how they influence constrained operational decisions. The framework also provides a pragmatic way to introduce decision-aware learning without rebuilding an established mixed-integer MPC scheduler, which may make this idea easier to integrate with existing energy-management architectures. At the same time, the results reveal an important trade-off: lower operating cost can be accompanied by a smaller grid-purchase margin under some disturbances. Future work should therefore combine economic feedback with more detailed flexible-resource models and robust, chance-constrained, or reserve-aware scheduling so that economic value and operational security are improved together.
Trustworthy AI & Resilient CPES | IEEE PES Representative to the Systems Council | SysC Distinguished Lecturer | Clarivate HCR Yang Li
Northeast Electric Power University
Read the Original
This page is a summary of: Economic feedback-guided forecasting and scheduling for microgrids using rolling model predictive control, Applied Energy, January 2027, Elsevier,
DOI: 10.1016/j.apenergy.2026.128778.
You can read the full text:
Contributors
The following have contributed to this page
