交通运输系统工程与信息

交通运输系统工程与信息 ›› 2026, Vol. 26 ›› Issue (5): 283-296.DOI: 10.16097/j.cnki.1009-6744.2026.05.024

• 系统工程理论与方法 • 上一篇    下一篇

面向容量压力识别的城市电动汽车充电需求预测

姚胜永*a,b ，赵晨a,b ，孙伟萍a,b   

  1. 石家庄铁道大学，a.交通运输学院；b.河北省交通安全与控制重点实验室，石家庄050000
  • 收稿日期:2026-06-15 修回日期:2026-07-25 接受日期:2026-08-20 出版日期:2026-10-25 发布日期:2026-10-21
  • 作者简介:姚胜永（1978—），男，河北保定人，教授，博士。
  • 基金资助:
    河北省科学技术厅创新能力提升计划软科学研究专项 (2535763D)。

Urban Electric Vehicle Charging Demand Prediction for Capacity Pressure Identification

YAO Shengyong*a,b, ZHAO Chena,b, SUN Weipinga,b   

  1. a. School of Transportation; b. Hebei Key Laboratory of Traffic Safety and Control, Shijiazhuang Tiedao University, Shijiazhuang 050000, China
  • Received:2026-06-15 Revised:2026-07-25 Accepted:2026-08-20 Online:2026-10-25 Published:2026-10-21
  • Supported by:
    Soft Science Research Project of Innovation Capacity Enhancement Program of Hebei Provincial Department of Science and Technology (2535763D)。

PDF

摘要： 针对城市电动汽车充电需求短时波动性强和区域差异显著，且连续需求预测结果难以直接表征公共充电设施运行压力的问题，本文构建面向容量压力识别的短时充电需求预测与压力转化框架。该框架首先融合区域充电量、设施占用率、充电时长、电价、服务价格和天气因素等多源特征，利用时间卷积网络提取局部时间特征，引入双向长短期记忆网络刻画时序依赖关系，并通过注意力机制强化关键历史时刻对下一时段需求的影响；其次，以区域训练期历史充电需求第95分位值作为相对容量基准，构建容量压力指数，将连续需求预测结果转化为不同等级的容量压力状态；最后，结合压力等级的时空分布特征，识别重点高压力区域及时段。基于公开数据集开展实证分析，结果表明，所构建模型的拟合优度为0.9682，均方根误差、平均绝对误差和加权平均绝对百分比误差分别为152.7197kWh、41.2489kWh和16.6336%；在第95分位容量基准下，高压力及以上样本占比为22.6380%，其中，超负荷风险样本占比为8.7514%。研究实现了由短时充电需求预测向区域容量压力状态识别的转化，可为公共充电设施运行监测、重点区域筛选及差异化资源配置提供量化依据。

关键词: 城市交通, 充电需求预测, 深度学习, 电动汽车, 容量压力识别

Abstract: To address the problems that the charging demand in urban electric vehicle features the strong short-term fluctuations and significant regional differences, as well as the continuous demand forecasts cannot directly characterize the operational pressure of public charging facilities, this paper develops a short-term charging demand forecasting and pressure transformation framework for capacity-pressure identification. First, the proposed framework integrates multi-source features, including regional charging volume, facility occupancy rate, charging duration, electricity price, service price, and weather factors, and employs a TCN-BiLSTM-Attention model for short-term charging demand forecasting. Specifically, the TCN extracts local temporal features, the BiLSTM captures temporal dependencies, and the attention mechanism emphasizes key historical time steps for next-period demand prediction. Second, taking the 95th percentile of historical regional charging demand during the training period as a relative capacity benchmark, a capacity-pressure index is constructed to transform continuous demand forecasts into capacity pressure states at different levels. Finally, key high-pressure regions and periods are identified according to the spatiotemporal distribution characteristics of the pressure levels. Empirical analysis based on a public dataset shows that the developed model achieves a goodness-of-fit value of 0.968 2, with a root mean square error, mean absolute error, and weighted mean absolute percentage error of 152.719 7 kWh, 41.248 9 kWh, and 16.633 6%, respectively. Under the 95 th-percentile capacity benchmark, the proportion of samples classified as high pressure or above account for 22.638 0% of all samples, among which the proportion of overload-risk samples account for 8.751 4%. The proposed framework enables the transformation from short-term charging demand forecasting to regional capacity-pressure state identification and provides a quantitative basis for the operational monitoring of public charging facilities, the screening of key regions, and differentiated resource allocation.

Key words: urban transportation, charging demand prediction, deep learning, electric vehicle, capacity pressure identification

中图分类号: