1 of 57

数据科学2023 年夏季学期优化电动汽车充电站布局的空间经济分析

巴赫拉姆-汗-巴洛赫 (7047281)

赛拉-索海尔-安瓦里 (7047706)

乌默尔-布特 (7024124)

Cicy Kuriakose Anges (7047703)

阿坎什-毛里亚 (7047939)

1

2 of 57

内容

问题陈述

相关工作

我们的建议

数据收集

数据分析

机器学习建模

仪表板和代码

行业和初创企业

Data Science – Summer Semester 2023 - Prof. Wolfgang Maaß

2

26.05.23

3 of 57

问题

说明

3

4 of 57

私人交通工具数量增加

碳排放量增加

全球变暖

燃料枯竭

4

问题的出现

The rise in the number of private transport vehicles has contributed to the growing need for EV charging stations. Here are some key factors that highlight the urgency of addressing this problem:

Increased Number of Private Transport Vehicles: The global population has witnessed a significant increase in the number of private vehicles, particularly in urban areas. This surge in vehicle ownership has led to a higher demand for charging infrastructure to support the growing number of electric vehicles.

Rise of Carbon Emissions: Conventional vehicles powered by fossil fuels emit greenhouse gases, contributing to climate change. Electric vehicles offer a greener alternative by producing zero tailpipe emissions. However, to fully realize the environmental benefits of EVs, an efficient and widespread charging network is essential.

Global Warming: The excessive emissions of carbon dioxide and other greenhouse gases have led to global warming, causing detrimental effects on the planet. By promoting the adoption of electric vehicles through an optimal charging infrastructure, we can mitigate the environmental impact and work towards a sustainable future.

Exhaustion of Fuels: The reliance on fossil fuels for transportation is not only environmentally harmful but also unsustainable in the long run. The depletion of finite fuel resources necessitates a shift towards renewable energy sources and electric mobility. A well-planned network of EV charging stations plays a vital role in facilitating this transition.

Addressing the rise of these challenges requires careful analysis and planning to ensure the effective placement of EV charging stations. Our spatial-economic analysis approach aims to tackle these issues and provide solutions for a cleaner and more sustainable transportation system.

5 of 57

一种碳排放量较少的交通工具

电动汽车

其动力来自风能、太阳能或水能。

问题出在哪里？

5

替代方案？

6 of 57

6

电动汽车充电站

基础设施

电动汽车站设在哪里？

7 of 57

当前的解决方案？

花时间了解当地情况

与当地人谈论常见的地方

寻找加油/加气站

花钱请专家

7

8 of 57

人工智能可以帮助我们吗？

利用人工智能，我们可以

了解电动汽车站的使用趋势和模式

预测电动汽车站的位置

预测电动汽车站的利用率

8

欧洲哪些地方可为电动汽车充电

来源：OpenStreetMap

9 of 57

来源：https://www.statista.com/statistics/646075/total-number-electric-cars-germany/

但为什么是现在？

9

电动汽车注册数量

年份

10 of 57

提出电动汽车充电站问题

4 类问题

我们的范围

电动汽车充电桩的布置取决于哪些因素

充电需求利用率用户行为

道路连通性兴趣点地理特征

10

11 of 57

Related Work

11

12 of 57

13 of 57

这些方法的局限性

复杂且具有挑战性

为强化学习方法定义奖励函数可能存在问题

严重依赖数据的可用性

没有可用数据，这些方法不起作用
这些方法不适用于新城市

13

Formulating a Multiple-Objective Optimization Problem (MOOP) for optimal EV station placement can be a complex and challenging task. It often requires defining a loss function that effectively represents the various constraints involved, such as range anxiety, cost, and utilization. Finding the right balance and accurately capturing the trade-offs between these constraints can pose difficulties in the formulation process.

Similarly, when using reinforcement learning-based approaches, defining an appropriate reward function becomes crucial. Designing a reward function that incentivizes the reinforcement learning agent to make optimal decisions for charging station placement can be a non-trivial task. Ensuring that the reward function captures the desired objectives while taking into account the constraints and avoiding unintended consequences is a challenge that researchers must overcome.

Furthermore, machine learning methods that rely on traffic patterns, historic charging data, and EV locations are heavily dependent on the availability of quality data. If data is not readily accessible or comprehensive for a specific location, these methods may not perform optimally or even fail. Additionally, these methods may not work effectively in newly developed cities where there is a lack of historical data. Thus, data availability and quality play a significant role in the success and applicability of these machine learning-based techniques for optimal EV charging station placement.

14 of 57

我们的建议

数据收集 → 分析 → 建模

14

15 of 57

我们的建议

15

城市地图

道路网络

土地利用

POIs

人口密度

机器学习

电动汽车充电站位置

16 of 57

我们工作的利益相关者

16

Stakeholder	Examples	Reason
电动汽车制造商	大众、宝马、梅赛德斯-奔驰	有助于制造商了解德国哪些地区可能购买电动汽车作为主要或次要交通工具
EV charge point installers or service providers	Allego Germany, EnBW	Useful for any enterprise or business to be proactive about any bids or investment for the installation of charging stations
Policy Makers, Think tanks & Local Councils	German Federal Government, Berlin City Council	Useful for any marketing campaign or incentive scheme that involves the public to ensure it will be cost-effective and focused

our analysis caters to several key stakeholders.

EV manufacturers such as Volkswagen, BMW, and Mercedes-Benz can benefit from our findings, as they gain insights into areas in Germany where EV adoption is likely to be high.

For EV charge point installers or service providers like Allego Germany and EnBW, our work enables them to be proactive in planning and investment decisions for charging station installations.

Policy makers, think tanks, and local councils, including the German Federal Government and Berlin City Council, can utilize our analysis to design effective marketing campaigns and incentive schemes that encourage cost-effective EV adoption. By considering the interests of these stakeholders, our project contributes to informed decision-making and promotes the sustainable transition to electric mobility in Germany.

17 of 57

数据收集

17

18 of 57

空间数据：形状文件

城市形状文件：由 Esri 开发的一种流行的地理空间矢量数据格式。

我们使用 OpenStreetMaps（OSM）和 OSMNX API 提取形状文件。
出于研究目的，我们将研究范围限定在德国的八个城市。为了使我们的数据集具有多样性，我们对这些城市进行了精心挑选。
提取城市形状文件后，我们对整个城市进行栅格化处理，栅格尺寸为 500 米。

数据收集

18

划分网格

19 of 57

空间数据：

道路连通性：

道路连接对于确定城市中电动汽车 (EV) 的可达区域至关重要。
没有道路连通性的电动汽车站是不切实际的，也是无效的。
为了测量服务区的道路连通性，需要计算节点（路口）和边（线）的数量。我们只提取车行道。
节点代表路口，边代表连接这些路口的路段。

数据收集

19

柏林公路网

萨尔布吕肯公路网

网格中的边和节点

20 of 57

空间数据：

兴趣点 (POI) 数据：

电动汽车充电站的位置受区域内兴趣点 (POI) 的影响。
与加油相比，电动汽车充电所需的时间更长，因此加油站的战略位置应允许用户在充电过程中从事其他活动。
放置电动汽车充电站的常见区域包括停车位、大学、娱乐中心、购物中心、餐厅和办公室。

数据收集

20

POIs(Berlin)

21 of 57

空间数据：

土地使用数据：

电动汽车充电站的位置受土地使用因素的影响。
在住宅区，用户通常喜欢在家中通宵为电动汽车充电，从而减少了对公共充电站的需求。
商场和办公室等商业场所则需要电动汽车充电站，以满足无法在家充电的用户的充电需求。
土地使用模式有助于确定特定区域对电动汽车充电站的需求和必要性。

数据收集

21

22 of 57

数据收集

2. 电动汽车充电站数据：

城市形状文件: 政府通过联邦网络局网站提供德国现有充电站数据，网址为https://www.bundesnetzagentur.de 。

数据以逗号分隔值 (CSV) 格式提供，最新数据集可使用至 2023 年 3 月。
数据集包括插头数量、插头类型、所有者、充电模式和经纬度坐标等信息。
这些数据将作为 mac.hine-learning 模型的目标变量，用于分析和预测德国电动汽车充电站的相关信息。
我们还使用 OpenCharge.org 数据来确定电动汽车充电站的位置。

22

23 of 57

数据收集

3 人口数据：

GeoTiFF ：人口密度分布数据为了解城市电动汽车充电的流动性、普及程度和潜在需求提供了重要依据。 .

人口密度较高的地区往往拥有更多的电动汽车和更高的充电需求。
人口密度数据有助于识别森林、农场或丘陵等可能不需要电动汽车站的地区。
分析中使用了 Meta（之前为 Facebook）提供的高分辨率人口密度数据。
原始数据的分辨率为 30 米，但比较稀疏，因此将分辨率降低到 500 米，以改进分析。

23

Population density of Regionalverband Saarbrücken

City Center �EUROPA-Galerie�Saarbrücken Hbf

St. Johanner Markt

Rathaus

Uds

24 of 57

数据处理

数据集结构：

每个数据点代表一个城市内的网格多边形。
每个数据点都可计算出该多边形内有哪些兴趣点（POI）以及有多少兴趣点（POI）。

人口数据

本任务的目标变量：

网格多边形中的电动汽车站

24

25 of 57

数据收集概述

25

数据集

26 of 57

Data Analysis

26

27 of 57

德国每年新安装的充电站数量

27

28 of 57

每个城市的电动汽车站数量

28

29 of 57

哪些特征对放置

电动汽车站的位置？

29

30 of 57

萨尔布吕肯简介

30

31 of 57

Machine Learning Modeling

31

32 of 57

预测电动汽车充电站位置的两种方法，基于以下因素

无法获得利用率数据 -> 电动汽车充电站布局

有利用率数据 -> 计算电动汽车充电站利用率

在这两种情况下，我们使用所有其他 POI 进行预测

[公园、餐馆、停车场、大学......]

32

33 of 57

EV Charging Station Placement

Machine Learning Modeling

33

34 of 57

城市划分

我们根据人口将数据分为大城市和小城市。

大城市柏林 ,斯图加特,法兰克福

小城市：卡尔斯鲁厄,特里尔,萨尔布吕肯,美因茨

这有助于了解大城市和小城市在趋势、模式或模型方面的差异

34

35 of 57

Experiments - Model Performance

Experiment 1: Random Split:

Randomly split grid polygons of all cities into train and test sets.

Experiment 2: Big Cities Split

Big cities separated into train (2 cities) and test (1 city) sets.
All possible combinations tried and average performance reported.

Experiment 3: Small Cities Split

Small cities separated into train (3 cities) and test (1 city) sets.
All possible combinations tried and average performance reported.

Experiment 4: Cross-Domain Testing

Training on big cities and testing on small cities, and vice versa.

Performance Metrics: F1, AUC, and Balanced Accuracy.

35

36 of 57

Best Performing Models

After trying around 40 models, the following models demonstrated the best performance:�

Decision Tree Classifier

Bernoulli Naive Bayes

Nearest Centroid

Complement Naive Bayes

Multinomial Naive Bayes

�

36

37 of 57

模型结果

37

我们的模型在大数据和小数据上的表现一致，这证明了模型的通用性

38 of 57

EV Station Utilization Computation

Machine Learning Modeling

38

39 of 57

数据

英国 40 多个电动汽车站的利用率数据

来自 2017-2018 年

按周汇总数据，帮助模型更好地学习

创建日期时间特征：月、年、周数、纬度、经度和利用率

39

https://data.europa.eu/data/datasets/electric-vehicles-charging-points

40 of 57

Model

Objective: Determine the future usage of a given EV station location

Task: Predictive Modeling / Regression

Model : Random Forest

Model inputs: datetime features and EV station latitude and longitude

40

41 of 57

实际应用和优势

准确的布局决策

优化基础设施投资

改进规划和资源分配

41

42 of 57

使用范围：0-1000 千瓦时

成果

42

Models	MAE	RMSE
Linear Regression	350	383
Decision Tree Regressor	92	130
Random Forest Regressor	105	143

43 of 57

仪表板和代码

43

44 of 57

44

仪表板显示萨尔布吕肯地图，其中包含所有兴趣点

和预测的电动汽车站

在 Streamlit 上开发，在 Huggingface 上托管

https://huggingface.co/spaces/Saira20/Optimal-EV-charging-stations

45 of 57

Github Repository

45

https ://github.com/akansh12/data-science-Optimal-EV-station-placement

46 of 57

行业和初创企业

为进一步了解问题陈述并获取数据，我们联系了一些初创公司：

Smart EMOTION: https://www.smart-emotion.de

Nikol EV: https://nikolev.in/

Airlo: https://airlo.in/

采埃孚集团： https://www.zf.com

46

47 of 57

Thank You

47

48 of 57

Appendix

48

49 of 57

文献综述（摘要）

电动汽车充电基础设施规划问题大致可分为以下四类。

EVCS 的布置取决于多种因素（不完全）：

充电需求
使用情况
用户行为
兴趣点
道路连接性
旅行时间（焦虑范围）
地理特征
....
....
.... 等

传统方法是通过提出多目标优化问题（MOOP）来解决问题。
然后采用遗传算法（GA）、粒子群优化（PSO）来解决问题。
问题：MOOPs 比较复杂，需要对问题进行仔细的表述。
大多数情况下都没有考虑到现实世界的数据。

大多数方法都使用机器学习建模来预测充电需求和利用率；在使用机器学习来布置电动汽车站方面的研究还很有限。
用于训练 ML 模型的历史数据很难收集。

49

50 of 57

导言和问题理解

为什么改用电动汽车很困难？

与汽油车和柴油车相比，公众可选择购买的电动汽车种类有限。
对于普通居民来说，购买一辆新电动车的成本要高于购买或保留一辆现有的汽油/柴油车。
电动汽车充电基础设施分布稀少且不足，导致采用速度缓慢。

解决方案

建立模型，预测可能安装基础设施的电动汽车充电地点。

我们解决方案的利益相关者

50

利益相关者holder	Examples	Reason
Policy Makers, Think tanks & Local Councils	German Federal Government, Berlin City Council	Useful for any marketing campaign or incentive scheme that involves the public to ensure it will be cost-effective and focused
EV charge point installers or service providers	Allego Germany, EnBW	Useful for any enterprise or business to be proactive about any bids or investment for the installation of charging stations
EV Manufacturers	Volkswagen, BMW, Mercedes-Benz	Useful for manufacturers to understand areas in Germany that are likely to purchase an EV for their primary or secondary mode of transportation

51 of 57

52 of 57

Limitations Of These Approaches

The problem with MOOP is that they are complicated, and formulating these problem with constraints is challenging.
Similar drawback arises when defining reward function for approaches like Reinforcement learning.
Most of the research had access to data that includes:

Traffic pattern
Historic charging patterns
EV Demand
EV utilization
Existing EV charging station location

This limits these studies from predicting EV charging station when no information about users, their utilization is known.
These method are also not applicable to new locations where no EV information is available.

Can we do propose optimal EV station location when no historic charging data is available?

52

53 of 57

Real world Disclaimer: Industries, and Startups

53

54 of 57

Data processing

Dataset Structure:

Each datapoint represents a grid polygon within a city.
The dataset is represented as a dataframe.
Each row in the Data Frame corresponds to a grid polygon within the city.
Columns:

Geometry: Polygon coordinates defining the shape of the grid polygon.
Other columns: Numerical counts of different types of POIs within each grid polygon.

POI Data

POIs can have multiple polygons.
The dataset captures the presence of POIs within the city by checking if the center of each POI polygon falls within the city's grid polygons.
If a POI's center falls within a grid polygon, the count for that POI is added to the respective city grid.

Population Data

Original data was of 30 meter resolution but very sparse. We decreased the resolution to 500 meters by merging

Target Variable: EV

54

55 of 57

Introduction:

To model the dataset, several columns were selected as features for training.
The columns used for modeling are: ['EV_stations', 'parking', 'edges', 'parking_space', 'civic', 'restaurant', 'park', 'school', 'node', 'Community_centre', 'place_of_worship', 'university', 'cinema', 'library', 'commercial', 'retail', 'townhall', 'government', 'residential', 'population'].
The columns 'geometry' and 'city' were used for creating splits and were not part of the training process.�

Experiment Overview:

Various experiments were conducted to evaluate the performance of the models.
The dataset was divided into big cities and small cities, and the following experiments were performed:�

Two Methods to Predict Placement Of EV stations Based on:

Availability of utilization data
Unavailability of utilization data

With Availability of Utilization data:

55

56 of 57

Methodology

Introduction:

The dataset consists of a cities divided into polygons using a grid system.
It includes information about various Points of Interest (POIs) within the cities.�

Dataset Structure:

The dataset is represented as a dataframe.
Each row in the Data Frame corresponds to a grid polygon within the city.
Columns:

Geometry: Polygon coordinates defining the shape of the grid polygon.
Other columns: Numerical counts of different types of POIs within each grid polygon.�

POI Data

POIs can have multiple polygons.
The dataset captures the presence of POIs within the city by checking if the center of each POI polygon falls within the city's grid polygons.
If a POI's center falls within a grid polygon, the count for that POI is added to the respective city grid.

56

57 of 57

Utilization Based Approach

Utilized EV station utilization data from the UK
Data included charging session information for 40+ EV stations
Aggregated data on a weekly level
Created datetime features: month, year, week number, latitude, longitude, and utilization
Developed a predictive model
Model inputs: datetime features and EV station latitude and longitude
Objective: Determine the future usage of a given EV station location
Application: Input new location coordinates and future date/time to predict usage of a new EV station
Benefits:

Accurate placement decisions, optimized infrastructure investments, improved planning and resource allocation
More power for user to try out multiple locations based on other constraint i.e property price
Based on actual usage so discards sub optimal EV stations

�

57