Consensus-based time-series clustering approach to short-term load forecasting for residential electricity demand

• Published in: Energy and buildings 2023-11, Vol.299, p.113550, Article 113550
• Main Authors: Dab, Khansa, Henao, Nilson, Nagarsheth, Shaival, Dubé, Yves, Sansregret, Simon, Agbossou, Kodjo
• Format: Article
• Language: English
• Subjects: Aggregated load forecast; Consensus clustering; Gaussian process; Residential load patterns; Short-term load forecasting; Time-series
• ISSN: 0378-7788
• DOI: 10.1016/j.enbuild.2023.113550

Load forecasting could play a crucial role in energy management and control of buildings in residential neighborhoods. In these areas, electricity demand is influenced by different phenomena accounting for climate conditions and comfort preferences. The uncertain nature of these circumstances results in power profiles with diverse patterns. Under this condition, overall load prediction is suggested by utilizing Cluster-based Aggregate Forecasting (CBAF). Accordingly, this paper proposes a unified approach to such a practice. The proposed scheme employs an unsupervised machine-learning algorithm to develop a time-series clustering scheme that performs the classification task through the k-medoids-based clustering incorporating the Dynamic Time Warping (DTW) algorithm. Subsequently, a consensus is achieved among the resultant clusters where the Jaccard similarity index adjudges the similarity measurement. The Additive Gaussian Process (AGP), a powerful non-parametric forecasting technique, is exploited to predict aggregated load at each cluster level. With low complexity and high scalability, AGP is particularly utilized to provide effective forecasting. Numerical simulations on synthetic as well as real datasets have been carried out to illustrate the effectiveness of the proposed methodology. Additionally, two comparative studies are carried out with forecasts without clustering and with the benchmark non-parametric models employing a cluster-based technique. The proposed method demonstrates significant improvement in forecasting accuracy for both datasets by reducing the error metrics and achieving 7% improvement in the coefficient of determination (R2) value as compared to the aggregated load forecast achieved without clustering. The comparative study demonstrates that the proposed method with AGP can forecast the total residential load more accurately than other benchmark models with an improvement of 26% and 21% in R2, respectively, for both datasets.