【目的】准确的多区域负荷预测(MRLF)是确保电网稳定和优化现代电力系统经济调度的关键基础。传统负荷预测方法在捕获多区域负荷之间的动态空间依赖性时面临着挑战。为了解决传统方法无法提取负荷间动态空间特征的问题，本文提出了一种基于解耦和对抗图注意力网络(DAGAT)的MRLF模型。【方法】该模型有效地通过图注意力网络提取了多区域负荷之间的动态空间特征。首先，采用离散小波变换(DWT)将区域负荷序列分解成趋势分量和波动分量。其次，根据趋势分量和波动分量的不同特征设计了一个双通道特征提取模块，用于提取多区域负荷序列的时空特征。此外，引入了一种基于零和博弈的生成对抗网络架构，用于时空预测模型的对抗训练，并将得到的对抗损失与传统的预测损失函数加权结合作为模型的损失函数。最后，基于纽约独立系统运营商的真实多区域负荷数据进行仿真验证。【结果】与主流模型比较分析表明，相较于传统机器学习模型、仅能捕获时间依赖关系的预测模型以及采用静态权重矩阵表达空间相关性的预测模型，本文所提DAGAT模型更具优越性。该模型的优越性源于图注意力网络通过学习动态变化的权重捕获了多区域负荷之间的空间相关性，DWT清晰表示了多区域负荷序列的特征，联合损失进一步优化了模型参数。【结论】DAGAT有效提高了MRLF精度，解决了多区域负荷间动态空间特征无法捕获的问题，在工程实践中具有有效性和优越性。

[Objective] Accurate multi-regional load forecasting (MRLF) is a critical foundation for ensuring grid stability and optimizing the economic dispatch of modern power systems. Traditional load forecasting methods face challenges in capturing the dynamic spatial dependencies among multi-regional loads. To address the problem that traditional methods are unable to extract dynamic spatial features among loads, a MRLF model based on decoupling and adversarial graph attention network (DAGAT) is proposed.[Methods] The model effectively extracted the dynamic spatial features between multi-regional loads through graph attention networks. Firstly, the regional load series was decomposed into trend and fluctuation components using discrete wavelet transform(DWT). Secondly, a two-channel feature extraction module was designed to extract the spatio-temporal features of the multi-regional load series based on the different characteristics of trend and fluctuation components. In addition, a generative adversarial network architecture based on zero-sum game was introduced for adversarial training of spatio-temporal prediction models, the resulting adversarial loss was weighted and combined with a traditional forecasting loss function as the loss function of the model. Finally, simulation validation was performed based on real multi-regional load data from the New York independent system operator. [Results] Comparison and analysis with mainstream models showed that the DAGAT model proposed in this paper was more superior than traditional machine learning models, prediction models that only capture temporal dependencies, and prediction models that express spatial correlations using static weight matrices. The superiority of the model stemmed from the fact that the graph attention network captured the spatial correlation between multi-regional loads by learning dynamically changing weights, and the DWT clearly represented the characteristics of the multi-regional load series, and the joint loss further optimized the model parameters. [Conclusion] DAGAT effectively improves the accuracy of MRLF and solves the problem that dynamic spatial features between multi- regional loads cannot be captured, which is effective and superior in engineering practice.