Electric load forecasting is a core foundation for power grid planning and operation. However,traditional methods rely on training models with data from a single region,resulting in a significant decline in generalization capability when applied to cross-regional forecasting. To address this issue,this paper proposes a hybrid model that integrates a Temporal Convolutional Network (TCN),Long Short-Term Memory (LSTM),and an attention mechanism (TCN-LSTMs-Attention),combined with a Decentralized Aggregation Learning (DAL) strategy. In this framework,multiple sub-models are jointly trained to obtain a global model capable of cross-regional forecasting by sequentially training with data from different regions on the same server. Moreover,the proposed method incorporates a dynamic learning rate halving and parameter resetting mechanism to further accelerate model convergence. Experiments based on the dataset from the 9th China Electrical Engineering Cup Competition demonstrate that,compared to models trained independently on regional data,the proposed method improves Mean Squared Error (MSE),Mean Absolute Error (MAE),Root Mean Squared Error (RMSE),and Mean Absolute Percentage Error (MAPE) by 43.0%,29.5%,24.4%,and 35.4%,respectively,in cross-regional forecasting tasks. These results validate the model's robustness and engineering practicality in high-heterogeneity load scenarios.