近年来，随着电动汽车产业的发展完善和人民环保意识的增强，我国电动汽车保有量不断增加。不过，目前阶段电动汽车仍具有续航里程不高、单次充电时间较长等缺点，这些缺点已严重影响了用户的使用体验、购买意愿和电动汽车的推广。而代客“加电”服务被认为是减少用户等待时长问题的一种有效手段。而为了更有效的提供此类服务，本论文探究如何基于多种机器学习模型对加电服务需求进行高效准确的预测。

为开展预测模型的研究，本论文首先完成了数据提取与特征工程。在位置信息上，本文提出了瓦片地图和多边形地图两种地图划分方式，将北京、上海、深圳等六个城市划分成小区域；而在时间维度，本文采用滑动时间窗口和固定时间窗口两种时间划分方式。此外，本文在领域专家的指导下，借助描述性统计方法，选择与加电服务需求相关度较高的空间数据、历史订单数据、电动汽车宏观数据这三大类数据作为加电服务需求预测的数据依据。通过Spark大数据计算引擎，对多个相关源数据表进行数据预处理、数据连接、对应指标计算、聚合、重要性分析等特征工程操作得到最终送入机器学习模型的特征数据。

本论文在DeepST、ST-ResNet（时空残差网络）和Inception系列图像分类模型的结构的基础上，提出针对时空序列预测问题的ST-Inception-ResNet网络结构模型。为验证ST-Inception-ResNet模型的有效性，本论文同时实现了XGBoost、ST-ResNet、和带注意力机制的LSTM等多种机器学习模型对多个城市进行了电动汽车加电服务需求预测实验。模型调参之后，得到了多个模型的最优实验结果。实验结果发现，相比于ST-ResNet基础模型，使用ST-Inception-ResNet模型将预测准确率从77.35%提高到79.70%。

在上述加电服务需求预测模型的基础上，本文构建了一个加电服务需求预测和监控系统。该系统主要有两方面的作用：一方面，系统以直观、友好的可视化形式呈现需求预测模型的结果，提高了模型的可用性；另一方面，系统对模型预测结果是否正确进行跟踪，并对预测模型计算准确率时空分布等统计量，为进一步分析和改进预测模型提供数据支撑。

In recent years, with the development and improvement of the electric vehicle industry and the enhancement of people's awareness of environmental protection, the number of electric vehicles in my country has been increasing. However, at this stage, electric vehicles still have shortcomings such as low cruising range and long single charging time. These shortcomings have seriously affected the user's experience, purchase intention and the promotion of electric vehicles. The valet charging service is considered to be an effective means to reduce the user's waiting time. In order to provide such services more effectively, this paper explores how to efficiently and accurately predict the demand for valet charging services based on a variety of machine learning models.

In order to carry out research on prediction models, this paper first completed data extraction and feature engineering. In terms of location information, this paper proposes two map division methods: tile map and polygon map, which divides Beijing, Shanghai, Shenzhen and other six cities into small areas; and in the time dimension, this paper uses two kinds of time: sliding window and fixed window division method. In addition, under the guidance of domain experts, with the help of descriptive statistical methods, this paper selects three types of data, spatial data, historical order data, and electric vehicle macro data, which are highly relevant to the demand for valet charging service as the forecast for valet charging service demand. Through the Spark which is a big data calculation engine, data preprocessing, data connection, corresponding index calculation, aggregation, importance analysis and other feature engineering operations are performed on multiple related source data tables to obtain feature data that is finally sent to the machine learning model.

In this paper, based on the structure of DeepST, ST-ResNet (spatio-temporal residual network) and Inception series image classification models, this paper proposes the ST-Inception-ResNet network structural model for the prediction of spatiotemporal sequences. In order to verify the effectiveness of the ST-Inception-ResNet model, this paper also implements a variety of machine learning models such as XGBoost, ST-ResNet, and LSTM with attention mechanism to conduct electric vehicle valet charging service demand forecast experiments for multiple cities. After the model was adjusted, the optimal experimental results of multiple models were obtained. The experimental results found that, compared with the ST-ResNet basic model, the ST-Inception-ResNet model was used to increase the prediction accuracy from 77.35% to 79.70%.

Based on the above power supply service demand forecasting model, this paper builds a power supply service demand forecasting and monitoring system. The system has two main functions: on the one hand, the system presents the results of the demand forecast model in an intuitive and friendly visualization, which improves the usability of the model; on the other hand, the system tracks whether the model forecast results are correct and the forecast model calculate statistics such as the spatial and temporal distribution of accuracy, and provide data support for further analysis and improvement of the prediction model.