基于改进Autoformer的电力负荷预测  

An Electricity Load Forecasting Based on Improved Autoformer

作　　者：李明[1] 石超山 谭云飞 文贵豪 罗勇航 LI Ming;SHI Chao-shan;TAN Yun-fei;WEN Gui-hao;LUO Yong-hang(School of Computer and Information Science,Chongqing Normal University,Chongqing 401331,China)

机构地区：[1]重庆师范大学计算机与信息科学学院,重庆401331

出　　处：《计算机技术与发展》2025年第4期107-112,共6页Computer Technology and Development

基　　金：国家自然科学基金(61877051,61170192)。

摘　　要：针对基于Transformer的预测模型具有较高复杂度且仅关注时间步之间依赖性而忽略跨变量依赖性的问题,提出了一种基于Patch-CDConv-Autoformer的电力负荷预测方法。首先,对输入的序列数据进行可逆实例归一化处理,以提高数据的平稳性。然后,将序列数据分块编码并投影到向量空间中。接着,将分块后的序列数据输入到Autoformer的编码器中,以捕获各时间周期之间的依赖关系。之后,通过CDConv模块对编码器输出的时间依赖关系进行二次建模,并对跨变量之间的关系进行建模。最后,对全连接层输出的预测结果进行逆实例归一化,以还原数据的原始分布,从而获得最终的预测结果。该方法不仅进一步降低了复杂度,还提高了预测精度。在三个公共电力数据集上的实验中,该方法在短期(预测步长≤48)预测任务中的均方误差(MSE)平均降低了40.84%,在长期(预测步长≥192)任务中平均降低了25.72%。与序列预测领域的先进模型相比,该方法在大多数预测任务中取得了更高的精度。To address the issues of high complexity and the focus on temporal step dependencies without considering cross-variable dependencies in Transformer-based prediction models,we propose an electricity load forecasting method based on Patch-CDConv-Autoformer.Initially,the input sequence data undergoes reversible instance normalization to enhance data stationarity.The sequence is then partitioned into patches,embedded,and projected into a vector space.Subsequently,the patch-wise sequence data is fed into the Autoformer encoder to capture dependencies across different time periods.The CDConv module is then employed to perform secondary modeling on the temporal dependencies output by the encoder and to model the relationships between variables.Finally,the predictions output by the fully connected layer are subjected to inverse instance normalization to revert to the original data distribution,thereby obtaining the final forecasting results.The proposed method not only reduces complexity but also enhances prediction accuracy.In experiments on three public power datasets,the mean square error(MSE)of the proposed method decreased by an average of 40.84%in short-term(horizon≤48)prediction tasks and 25.72%in long-term(horizon≥192)tasks.Compared to advanced models in the sequence prediction domain,the proposed method achieves higher accuracy in most forecasting tasks.

关 键 词：电力负荷预测 TRANSFORMER 分块编码 Autoformer CDConv 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TM715[自动化与计算机技术—控制科学与工程]