Automating Value-Oriented Forecast Model Selection by Meta-learning: Application on a Dispatchable Feeder

Abstract

AbstractTo successfully increase the share of renewable energy sources in the power system and for counteract their fluctuating nature in view of system stability, forecasts are required that suit downstream applications, such as demand side management or management of energy storage systems. However, whilst many forecast models to create these forecasts exist, the selection of the forecast model best suited to the respective downstream application can be challenging. The selection is commonly based on quality measures (such as mean absolute error), but these quality measures do not consider the value of the forecast in the downstream application. Thus, we introduce a meta-learning framework for forecast model selection, which automatically selects the forecast model leading to the forecast with the highest value in the downstream application. More precisely, we use a meta-learning approach that considers the selection task as a classification problem. Furthermore, we empirically evaluate the proposed framework on the downstream application of a smart building’s photovoltaic-battery management problem known as dispatchable feeder on building-level with a data set containing time series from 300 buildings. The results of our evaluation demonstrate that the proposed framework reduces the cost and improves the accuracy compared to existing forecast model selection heuristics. Furthermore, compared to a manual forecast model selection, it requires noticeably less computational effort and leads to comparable results.