Reducing energy time series for energy system models via self-organizing maps

Abstract

The recent development of renewable energy sources (RES) challenges energy systems and opens many new research questions. Energy System Models (ESM) are important tools to study these problems. However, including RES into ESM strongly increases the model complexity, because one needs to model the fluctuant, weather-dependent electricity production from RES with a high level of granularity. This leads to long execution times. To deal with this issue, our objective is to reduce the input time series of ESM without losing their energy-related key characteristics, such as weather-dependent fluctuations in production or peak demands. This task is challenging, because of the variety and high-dimensionality of the data. We describe a carefully engineered data-processing pipeline to reduce energy time series. We use Self-Organizing Maps, a specific kind of neural network, to select “representative days”. We show that our approach outperforms the existing ones with respect to the quality of ESM results, and leads to a significant reduction of ESM execution times.