Investigating the “Whole-Life Performance” of Representative Profile Extraction for Microgrid Planning

Numerous innovations emerge for decarbonization in energy systems. Practitioners and policymakers look for reliable methodology to evaluate their actual contributions unbiasedly. Usually, such evaluation starts from extracting a collection of profiles, which represents the actual application scenarios and certainly influences the ultimate results. However, such a fundamental task has seemingly long been treated casually. Limited literature on this topic rarely extends their attention beyond clustering methods. In this paper, we present mainly three innovations. First, we make a systematic comparison between clustering-based and sampling-based methods both theoretically and empirically. Particularly, we introduce importance sampling with the hope to lower estimation variances. Second, we highlight the importance of evaluating extraction methods on specific downstream tasks, and demonstrate with a case study of optimal battery sizing for a microgrid with real-world data. Lastly, we propose extensions for multi-modal data extraction. Specifically, we include electric vehicle (EV) sessions in representative profiles, which allows downstream flexible charging scheduling. In our numerical study, we find that sampling-based methods preserve more within-profile (WPV) and cross-profile variances (CPV) of the raw dataset. However, for planning battery systems in grid-connected microgrids, conventional K-means clustering K = 10 gives satisfactory estimates. Whether, and how to include EVs in representative profiles has a significant impact, especially in high EV penetration scenarios.