Exploring the Potential of Down-Sampling to Reduce Temporal Resolution in Energy System Optimisation Models

Abstract

Energy system optimisation models (ESO) are increasingly important for managing renewable energy source (RES) integration. However, due to the intermittency of RES, the inclusion of storage and the increasing interconnectedness of energy sectors the complexity of these models increases and makes them computationally intensive. This study aims to assess the effectiveness of alternative down-sampling (DS) methods in reducing temporal resolution in ESO models while approximating accuracy on hourly resolution. We compare this performance to benchmark DS that uses the default average or energy value for aggregation. Basic statistics and hybrid DS methods were applied in an operational model and two variations of an investment model within the context of a Northwestern European case study. The methods were evaluated based on their ability to reduce solve time and approximate model accuracy in terms of total system costs, investment decisions, flows and net position behaviour. The results show that all DS methods significantly reduced solve time across all model configurations, with reductions ranging from 44-75 % on 2-hourly resolution, 62-88 % on 3-hourly resolution and 81-92 % on 4-hourly resolution. Using the first, last, or midpoint values when aggregating, consistently outperformed the benchmark showing lower differences with the hourly reference for total system costs and investment decisions. This performance can be linked to their ability to represent energy value and ramp distributions at a lower resolution, which showed to be consistently better than the benchmark. Whereas using the maximum and minimum methods when aggregating, showed higher errors in energy value and ramp distribution representation and accordingly lower model performance in the evaluated metrics than the benchmark. The hybrid method attempts to balance energy value and ramp distribution representation by combining the benchmark average and maximum when aggregating. The hybrid method generally showed promising results in the different model configurations when applied to the energy demand profile and wind profiles, especially in total system costs and investment decisions. These findings suggest that alternative DS methods have the potential to offer a more effective solution than the benchmark that uses default averaging to reduce model complexity while approximating accuracy on hourly resolution. Moreover, they contribute to understanding the impact of DS on model performance. Future work could explore their application in larger-scale models, across other optimisation-based domains, in existing methods that use DS or applications that rely on time series such as time series forecasting, trend and anomaly detection.