Electrical utilities store large volumes of meter readings but lack tools that translate raw consumption into actionable insight. We present an integrated, explainable frame work that combines appliance-level energy disaggregation, type aware anomaly detection, and next-day demand forecasting from aggregate meter data, and generates explainable recommenda tions for utility personnel. Using two years of 15-minute AMPds2 household data (70,080 bins; 23 sub-metered appliances aggre gated to one meter), we compare deep and statistical models on each branch. For non-intrusive load monitoring (NILM) the plain CNN and a CNN-LSTM hybrid are statistically indistinguishable across four seeds (p = 0.125), so we report no decisive winner. Our central finding concerns anomaly detection: feeding raw contextual columns alongside consumption (naive concatena tion) is null or negative across all five anomaly types, but a residual-based formulation—subtracting a conditional expecta tion E[signal | context] fitted on training folds only—significantly improves detection on the two context-conditioned types, raising per-type F1 from 0.0288 to 0.0546 (duration, IsolationForest, ∆=+0.026, 95%CI [0.019,0.033]) and from 0.0014 to 0.0540 (combination, ∆=+0.053, CI [0.044,0.062]), while control types show no such benefit. For forecasting, LSTM, LightGBM, and XGBoost are again a statistical toss-up (best MAE margin 1.86%). Explainability is provided via SHAP-based expectation vs-actual attribution, and a rule-based engine turns detections and forecasts into actionable, non-automated recommendations surfaced on a PostgreSQL-backed Node.js/React dashboard. A three-arm human evaluation of the explanations was designed and built but not executed with participants and is reported as future work only.