Utilization of Hierarchical Agglomerative Clustering Algorithm to Find Representative Days for the Optimization of Electricity Generation Input Data Dimension in Solar Energy Systems

Abstract

Over the recent decades, the energy system decarbonization has played an essential role in the greenhouse gas emissions reduction required to limit the climate change impacts. Accordingly, renewable energy resources such as solar power plants have become more critical. Hence, integrating more solar plants into the power system generation mix makes it undeniable to model their temporal and spatial variability properly. A high temporal resolution is ideal for capturing renewables variability in an energy system model. However, computational restrictions pose design and implementation-related constraints and make it infeasible or computationally expensive in practice. Many of the current models only include a limited number of representative time slices that aggregate periods with similar input data profile patterns to reduce the time resolution of energy models, which in turn increases the computational tractability. The proper selection of the time slices to consider in a model is vital to downscale the time dimension while resulting in a minimum error on the model outputs. However, available methods are limited in applying to the input data with many time segments, which is a disadvantage of models with high shares of renewable energy. This project presents a computational efficient time slice clustering approach applicable to hourly solar generation input data for multiple locations. This method determines representative days (instead of all days in a year) to be utilized in the energy system modeling procedure by applying the hierarchical agglomerative clustering (HAC) method into the input data profile. It is indicated that four representative days in every thirty days and twelve representatives in every ninety days suffice across the entire input dataset to keep the error within an acceptable range. The input dataset comprises real-world electricity generation values for three solar power plants (1 MW installed capacity each) located in three spots on Vancouver Island, including Victoria, Nanaimo, and Port Hardy. The proposed algorithm has been evaluated using monthly and seasonal data segments. The best candidates with a minimum sum of squared errors have been introduced as their cluster’s representative days in every scenario. Finally, the effectiveness of our proposed ML approach has been demonstrated using the dendrogram, and also the importance of properly clustering representative days for solar power generation units is emphasized by comparing our proposed HAC approach with the downsampling method and the utilization of the CH index as a clustering quality measure.