A hyperparameter-space clustering methodology of residential electricity loads

Clustering residential electricity consumption patterns is a crucial step toward scalable and interpretable energy forecasting. Traditionally, clustering relies on direct analysis of load time series, which may obscure deeper behavioral or structural similarities between households. This study introduces a novel approach that shifts the focus from using raw consumption data to using the hyperparameters of stacked hour-ahead deep learning forecasting models—specifically based on LSTM, Bi-LSTM, and GRU architectures. By optimizing models independently for each household and clustering based on the resulting hyperparameter configurations, we reveal functional similarities in forecasting behavior that are not always evident in the original data. This method enables the formation of new, interpretable consumer segments, which can support the development of tailored forecasting models per cluster. Utilizing over three years of data from 200 households located in London, we evaluate the proposed hyperparameter-based clustering against traditional time-series clustering, applying standard metrics and explainable AI techniques (SHAP and LIME) to interpret the results. Findings demonstrate a strong alignment between the mean and median consumption patterns of clusters for both approaches, validating the effectiveness of the proposed method. Moreover, the analysis highlights that the feature transformation layers play the most critical role in shaping cluster formation, underscoring its significance in capturing underlying consumption behaviors. Beyond its analytical value, the method also offers a privacy-preserving advantage by requiring only the exchange of model hyperparameters rather than raw consumption data.