Abstract: To address the problem that existing short-term power load forecasting methods have difficulty effectively modeling non-linear structures and lack cross-scale and cross-variable interaction capabilities, a short-term power load forecasting model based on Dynamic Convolution Decomposition and Multi-Scale Graph (DCDMSG) was proposed. First, a dynamic convolution decomposition method was adopted to decompose the trend and seasonal components in the load sequence, in order to cope with the complex non-linear structure. Then, the trend component was directly forecasted using a linear layer, and the seasonal component was used to construct a multi-scale sequence using an adaptive multi-scale sequence construction method for deep modeling. Next, to model the complex dependencies within and outside the multi-scale sequence, a multi-scale fusion graph was used to capture dependencies within different scale sequences, a multi-variable correlation graph was utilized to model the correlations between different variables, and multi-scale hybrid jump propagation was employed to aggregate features in the graph. Finally, the seasonal component was forecasted and its prediction result was weighted and fused with the trend component's prediction to obtain the final forecasted value. The performance of DCDMSG is superior to other models on two real-world load datasets. On the Australian dataset, a Mean Absolute Error (MAE) of 0.379 and a Root Mean Square Error (RMSE) of 0.529 were achieved, which represents a 7.11% and 23.55% reduction compared to the sub-optimal model, FourierGNN, respectively. On the Cele dataset, an MAE of 0.437 and an RMSE of 0.708 were achieved, representing a 5.62% and 5.73% reduction compared to the sub-optimal model, PatchTST, respectively.The experimental results verified the superiority of DCDMSG in the short-term power load forecasting task, effectively improving the prediction accuracy through seasonal-trend decomposition and cross-scale and cross-variable modeling.

Key words: short-term power forecasting, dynamic convolution decomposition, multi-scale features, graph neural network, mix-hop propagation

摘要： 针对现有短期电力负荷预测方法难以有效建模非线性结构和缺乏跨尺度与跨变量交互能力的问题，提出基于动态卷积分解和多尺度图的短期电力负荷预测模型（DCDMSG）。DCDMSG首先采用动态卷积分解方法分解负荷序列中的趋势和季节项，以此应对复杂的非线性结构；接着对趋势项直接利用线性层预测，对季节项采用自适应多尺度序列构建方法生成多尺度序列并进行深度建模；其次，为建模多尺度序列内外的复杂依赖关系，模型使用多尺度融合图捕捉不同尺度序列内的依赖关系，利用多变量相关图建模不同变量之间的相关性，并采用多尺度混合跳传播聚合图中的特征；最后，对季节项进行预测并将其与趋势项预测结果加权融合得到最终预测值。DCDMSG在两个真实负荷数据集上的表现均优于其他模型。具体而言，在Australian数据集上，DCDMSG的平均绝对误差（MAE）为0.379，均方根误差（RMSE）为0.529，相比次优模型FourierGNN，MAE和RMSE分别降低了7.11%和23.55%；在Cele数据集上，DCDMSG的MAE为0.437，RMSE为0.708，相比次优模型PatchTST，MAE和RMSE分别降低了5.62%和5.73%。实验结果验证了DCDMSG在短期电力负荷预测任务中的优越性，通过季节趋势分解和跨尺度与跨变量建模，有效提高了预测精度。

关键词: 短期负荷预测, 动态卷积分解, 多尺度特征, 图神经网络, 混合跳传播