Unsupervised learning-based operational regime discovery and weather sensitivity analysis of a residential photovoltaic battery system for energy management

Behind-the-meter photovoltaic battery systems exhibit heterogeneous daily operating behaviors shaped by weather conditions, household demand, storage control, and grid interaction. Identifying these operational regimes is essential for improving forecasting, system control, and policy design, particularly where labeled operational data are unavailable. This study proposes an unsupervised framework to identify recurring daily operational regimes in a residential photovoltaic battery system and to quantify regime-specific weather sensitivities. Fourteen months of high-resolution operational data from a residential photovoltaic battery system in Yangon, Myanmar, are analyzed. Daily features capturing photovoltaic generation, electricity consumption, grid exchange, battery charging and discharging, state of charge, and self-consumption are derived from intraday profiles. Isolation Forest outlier detection removes anomalous operating days, yielding 411 representative days for analysis. K-means clustering identifies four operational regimes (Silhouette = 0.216), validated using Gaussian Mixture Models, hierarchical clustering, and density-based methods. Principal component analysis confirms consistent and interpretable operational structure and interpretability of the identified regimes. Cluster-wise standardized regression models link daily photovoltaic energy production to meteorological drivers. Solar irradiance is the dominant explanatory variable across all regimes, with effect sizes increasing from grid-dependent to high PV self-sufficient modes. Temperature and relative humidity exhibit regime-dependent secondary effects, while wind speed plays a limited role. Seasonal analysis reveals systematic shifts in regime prevalence driven by climatic conditions. The proposed framework provides an interpretable approach applicable to tropical monsoon PV-battery systems for diagnosing real-world photovoltaic battery system behavior and yields actionable regime labels that support regime-aware forecasting and adaptive energy management in distributed energy systems.