<p>Accurate energy forecasting increasingly depends on explicitly modeling occupant behavior while maintaining computational efficiency and statistical reliability. This paper proposes a household-level probabilistic behavior-aware hybrid deep learning framework that integrates latent behavior discovery, probabilistic behavior inference, and a TCN–BiLSTM–CNN based forecasting architecture. Appliance-level energy signals are converted to behavior-sensitive temporal features, which are then clustered to discover latent behavior states, where the number of clusters is determined based on clustering effectiveness, prediction accuracy, and computational behavior efficiency. A random forest-based leakage-free classifier is trained to predict behavior probabilities, which are incorporated into the forecasting model as probabilistic behavior features. This behavior classifier is assessed using accuracy, Brier score, Jensen–Shannon divergence, and expected calibration error, ensuring both predictive accuracy and probabilistic calibration, while the statistical tests ensure robust improvements. The Granger causality test is carried out in a leakage-free fold-wise manner for selecting behavior features. The findings of the experiments based on chronological TimeSeriesSplit reveal that the probabilistic behavior integration improves forecasting accuracy over the behavior-agnostic approaches, reducing the RMSE by 1.62% and 6.99% and improving the R² from 0.6925 to 0.7025 and 0.8822 to 0.8980, respectively. Additional ablation experiments highlight the impact of raw proxy variables, engineered temporal features, hard clustering labels, and probabilistic behavior representations on the forecasting performance. The computational cost analysis highlights the important insights about the fitting time reduction of 3.58% and 13.31% for each case study under identical training settings, supporting practical residential energy management applications.</p>