Forecasting multiple attributes considering uncertainties in a coupled energy systems model

Abstract

Time-series prediction has improved enormously with state-of-the-art machine learning. However, it is hard to integrate ML forecasting methods into energy systems models (ESM) because the trained model has to conform to often strict requirements of the ESM, like class structure, computational limits, or restricted input and output. We use the open-source forecasting software FOCAPY to train and compare multiple algorithms from basic benchmarks to comprehensive machine learning models. We also show ways to integrate such production-ready ML-models into ESM. The time series under consideration represent the optimized and aggregated grid interactions of three key actors within the open-source ESM AMIRIS: (a) rooftop photovoltaic systems with battery storage, (b) heat pumps, and (c) electric vehicles. The individual household decisions are obtained using an optimization model for each technology, representative weather regions across Germany, and household types. We present results predicting the aggregate demand for a week ahead in an hourly resolution for one year in Germany for each model.