Temporal aggregation of time series to identify typical hourly electricity system states: A systematic assessment of relevant cluster algorithms

Comprehensive numerical models are pivotal to analyze the decarbonization of electricity systems. However, increasing system complexity and limited computational resources impose restrictions to model-based analyses. One way to reduce computational burden is to use a minimum, yet representative, set of system states for model simulation. These states characterize fluctuating renewable generation and variable demand for electricity prevailing at a certain point in time. A review of possible time series aggregation techniques identifies cluster algorithms as most adequate, with k-Means and the Ward algorithm predominating. However, throughout the surveyed literature, the line of reasoning for the selection of these algorithms remains unclear. To support the electricity system modeling community in selecting an algorithm, this paper devises a systematic multi-stage evaluation approach to compare a large variety of cluster analysis configurations, differing in algorithm, cluster representation, and number of clusters. Results show that electricity demand and renewable energy generation time series can be compressed to below one percent while sustaining global characteristics of the original data. Two potent cluster configurations are identified, confirming k-Means and WARD as being prevalent. Beyond electricity market data, the methodology can be applied to various types of fundamental time-dependent input data.(c) 2022 Elsevier Ltd. All rights reserved.

Automated summary

Comprehensive numerical models are crucial for analyzing the decarbonization of electricity systems. However, the increasing system complexity and limited computational resources pose limitations on model-based analyses. To reduce computational burden, one approach is to use a minimum yet representative set of system states for model simulation, which characterize fluctuating renewable generation and variable demand for electricity at a specific point in time. This paper presents a systematic multi-stage evaluation approach to compare different cluster analysis configurations in order to support the selection of a cluster algorithm. Results show that electricity demand and renewable energy generation time series can be compressed to less than 1% while maintaining global characteristics of the original data. Two effective cluster configurations, utilizing k-Means and WARD algorithms, are identified. The methodology can also be applied to various types of time-dependent input data.