A Robust Building Energy Pattern Mining Method and its Application to Demand Forecasting

With the current shift of the Netherlands energy systems to reduce its carbon footprint, appropriate planning is essential. However, obtaining accurate energy predictions become increasingly difficult from coupled weather and user behavior volatility. This work proposes a robust load pattern identification method through clustering whilst assessing the benefit of the attained information on enhancing accuracies of building energy prediction. A robust cross-validation is illustrated from assessing varying distance metrics and clustering algorithms, namely Euclidean and mahalanobis distances with Fuzzy C-Means and Agglomerative Hierarchical clustering. Leveraging nine identified and characterized patterns, a final indirect evaluation is realized with energy demand forecasting. The proposed data mining method provides a better understanding of the interactions between user behavior and future energy needs which can fundamentally impact strategic energy planning such as asset management, and collaborative operations.