Abstract

The increasing availability of high-frequency electricity consumption data has created new opportunities for understanding complex energy behaviors beyond dominant global patterns. However, conventional statistical and deep learning models primarily focus on average trends and often overlook weakly expressed local structures embedded within time series. This study proposes an exploratory framework for identifying latent micro-patterns in multivariate electrical consumption data using a combination of entropy-based feature scoring, temporal residual analysis, anomaly detection, and density-based clustering. The methodology is applied to the individual household electric power consumption dataset from the University of California, Irvine repository . A dark value score combining entropy and rarity is introduced to identify variables potentially containing hidden informational structures. Residuals extracted from Long Short-Term Memory predictions are analyzed using Isolation Forest and density-based clustering techniques to reveal statistically coherent micro-signals. Results show that these residual micro-structures are not random noise but exhibit identifiable temporal and statistical organization. Although clusters remain weakly separated, their distributions differ significantly from normal observations, supporting the existence of localized operational regimes invisible to dominant predictive models. The proposed framework provides an interpretable complementary perspective to traditional forecasting approaches and offers potential applications in anomaly detection, behavioral analysis, and intelligent energy monitoring.