This paper proposes a deep learning framework for bus load time series forecasting. The data comes from the hourly operation records of 48 bus nodes in the regional power system, which form a total of 132,480 valid observations. The input is composed of active load, reactive load, time index, node number and rolling statistical characteristics. The model combines local fluctuation convolution extraction, long-term dependence modeling, attention enhancement in key periods and residual correction output to compress the peak section deviation and stabilize the single-step prediction. The framework is trained under unified hyperparameter search and compared with SVR, random forest, GRU, and standard LSTM. Experimental results show that the proposed model has R² of 0.9969, MAE of 18.43 MW, and MAPE of 0.86%, which is superior to the comparison models in terms of overall error compression and peak tracking in critical periods. Visual analysis further confirms that the proposed method maintains good response consistency in the pre-peak lifting, peak maintenance and post-peak falling stages, indicating its applicability in scheduling support and bus load management.
