Power load forecasting is a critical aspect of optimizing and operating power systems. However, existing models often exhibit low accuracy in cross-regional forecasting, while traditional federated learning methods face challenges such as high communication overhead and low training efficiency, making them unsuitable for practical applications. To address these issues, this paper proposes an efficient, low-communication framework for cross-regional load forecasting. Specifically, a hybrid model based on Temporal Convolutional Networks (TCN), Long Short-Term Memory Networks (LSTM), and Attention Mechanism (TCN-LSTM-Attention) is designed. This model is combined with the proposed Decentralized Aggregation Learning (DAL) strategy, optimizing the global model through local training of sub-models from different regions and parameter aggregation. Furthermore, a dynamic learning rate halving and parameter reset mechanism is introduced to accelerate model convergence. Experimental results demonstrate that the proposed method significantly improves forecasting accuracy, reduces communication overhead, and enhances training efficiency, fully validating its effectiveness and practicality in cross-regional power load forecasting tasks.