Year: 2025
Type: conference-paper
Abstract: Residential photovoltaics have seen significant uptake in recent years.With battery systems becoming more affordable, the optimal operation of a photovoltaic-battery system can bring significant savings to households.Optimization of such a system requires us to make accurate forecasts of the photovoltaic power generation, in order to properly know when to charge and discharge the battery, given how much energy was generated.Forecasting models have become more and more accurate, due to algorithmic advances and data availability, but forecast accuracy is typically measured in generic statistical metrics, which might not align well with the goals of the downstream application.Therefore, we propose an integrated approach to residential PV forecasting, in which the optimization algorithm is solved online during the training of the forecaster.Our results show that applying this methodology decreases the cost when applying the forecast to the overall energy cost minimization of a photovoltaic-battery system.These findings show that using generic statistical metrics for forecast evaluation will not necessarily yield the optimal forecast for any given scenario and that coupling the forecast of a parameter with the optimization problem itself adds value for the user, providing a relative cost decrease of c. 9% over a two-stage approach.
Institution: KU Leuven
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4411345738