多尺度路由时空注意力的综合能源多元负荷预测

PDF下载HTML

Title:

Integrated energy multiple load forecasting for multiscale routing spatiotemporal attention

作者:

王德文^1,2, 张林飞¹, 苗庆健¹, 李成浩¹, 赵文清^1,3

1. 华北电力大学 计算机系, 河北 保定 071003;
2. 河北省能源电力知识计算重点实验室, 河北 保定 071003;
3. 复杂能源系统智能计算教育部工程研究中心, 河北 保定 071003

Author(s):

WANG Dewen^1,2, ZHANG Linfei¹, MIAO Qingjian¹, LI Chenghao¹, ZHAO Wenqing^1,3

1. Department of Computer, North China Electric Power University, Baoding 071003, China;
2. Hebei Key Laboratory of Knowledge Computing for Energy & Power, Baoding 071003, China;
3. Engineering Research Center of Intelligent Computing for Complex En

关键词:

综合能源; 多元负荷预测; 多尺度; 多核局域分解; 路由时空注意力; 周期性; 趋势性; 耦合性; 相关性

Keywords:

integrated energy; multiple load forecasting; multiscale; multicore local decomposition; routing spatiotemporal attention; periodicity; tendency; coupling; correlation

分类号:

TP399; TM721

DOI:

10.11992/tis.202501003

摘要:

多元负荷预测是保障综合能源系统(integrated energy systems, IES)稳定运行的关键。现有方法缺乏对电、冷、热等多元负荷的深度挖掘与分析，限制了预测性能。为解决此问题，本文深入剖析多元负荷的统计特征、季节–日内耦合性及与天气因素的相关性，进而提出一种基于多尺度路由时空注意力机制的综合能源多元负荷预测模型。该模型通过多核局域分解以捕获多元负荷的多尺度周期与趋势特征；针对多元负荷间的复杂耦合性及负荷与天气的相关性，设计路由时空注意力机制与多尺度编解码器，生成多尺度周期预测结果，并融合循环神经网络的趋势预测结果以输出最终预测值。基于实测数据集的耦合性分析、消融实验及对比实验表明：相较于 LSTM(long short-term memory)、Transformer、CNN-GRU(convolutional neural network gated recurrent unit)、Autoformer、FEDformer 等主流模型，所提模型在不同多元负荷耦合强度下均具备更优的预测精度。

Abstract:

Accurate multi-energy load forecasting is critical for the stable operation of integrated energy systems (IES). Existing methods often fail to capture the complex interactions among electricity, cooling, and heating loads, thereby limiting forecasting effectiveness. To address this challenge, this study first conducted an in-depth analysis of the statistical features of multi-energy loads, their seasonal–intraday coupling patterns, and their correlations with weather factors. Based on these insights, a multiscale spatiotemporal routing attention model was proposed for multi-energy load forecasting in IES. The model incorporates multikernel local decomposition to extract multiscale periodic and trend features, while a routing spatiotemporal attention mechanism, coupled with a multiscale encoder–decoder, is designed to capture inter-load dependencies and load–weather correlations. Periodic forecasts produced by this framework are further combined with trend predictions from recurrent neural networks to generate the final outcomes. Extensive evaluations on real-world datasets, including coupling analysis, ablation studies, and comparative experiments, demonstrate that the proposed model consistently outperforms mainstream methods such as LSTM, Transformer, CNN-GRU, Autoformer, and FEDformer, across varying levels of load coupling strength.