Abstract: To address the challenges of power supply security posed by regional droughts and other special circumstances, and to achieve accurate load forecasting and refined management for distribution areas, this study focuses on 96-point high-frequency power data from 14,703 user sessions from 842 users in three power supply stations of a county during drought periods. A hybrid forecasting method integrating user behavior analysis and advanced prediction models is proposed. Firstly, data cleaning and normalization are performed, and the K-Means clustering algorithm is employed to identify six typical electricity consumption patterns. Subsequently, the proportion of users with different consumption patterns in each feeder is then used as a macro-feature, combined with exogenous variables such as weather and holidays to construct a feature set. Finally, three models, which are 3-day moving average (3-dayMA), SARIMAX, and XGBoost, are built and compared. After hyperparameter optimization using Optuna and rolling forecast evaluation, the SARIMAX model demonstrates the best performance (RMSE is approximately equal to 0.22), followed by XGBoost (RMSE is approximately equal to 0.25), both outperforming 3-dayMA (RMSE is approximately equal to 0.28). The study confirms that transforming microscopic user electricity consumption behavior into macro-features can effectively improve forecasting accuracy, providing data support and methodological references for grid load regulation and orderly electricity consumption strategies.