01
Project Overview
AI-driven site selection PoC to accelerate bp's EV charging network expansion in China
Problem Statement
China's EV market hit 13.1M vehicles in 2022 (+67% YoY), creating urgent demand for high-value charging locations. Existing site selection was manual, inefficient, costly, and difficult to assess at scale. No structured AI modelling tool existed in the industry to systematically identify optimal sites.
Manual & inefficient
High assessment cost
No AI tooling
Fast-moving market
Solution & PoC Objectives
Build a modelling platform combining GIS spatial AI, ML classification, and forecast algorithms to score, rank, and predict performance of candidate EV charging sites — enabling business teams to make faster, data-driven expansion decisions. Pilot city: Guangzhou. Validation against business development on-site investigation results.
AI site scoring
Performance forecast
GIS visualisation
Validated vs. BD data
02
System Architecture
5-layer AI platform: raw data → derived features → ML engine → AI modules → user interface
USER LAYER
🗺️ Heat Map
Density visualisation
Density visualisation
📍 Site Evaluation
Score + grade
Score + grade
🏆 Site Recommendation
Ranked list
Ranked list
⭐ Good Sites
Factor analysis
Factor analysis
📈 Perf. Forecast
Revenue · break-even
Revenue · break-even
↕ SaaS Platform API (Alading)
AI MODULES
🎯 Location Evaluation
Classification algorithms
Classification algorithms
💡 Location Recommendation
Clustering + CF
Clustering + CF
📊 Performance Forecast
Regression algorithms
Regression algorithms
↕ Spark MLlib · Model Training Pipeline · Feature Selection
ML ENGINE
🌐 Kernel Density
Spatial distribution
Spatial distribution
🚗 OD/PA Model
Population flow
Population flow
📐 Spatial Interpolation
Feature coverage
Feature coverage
🕸️ Force Layout
Feature correlation
Feature correlation
🏙️ Business District
Commercial analysis
Commercial analysis
💎 Value Factors
High-value site ID
High-value site ID
↕ Data Processing · Deduplication · Normalisation · Derived Feature Engineering
DATA LAYER
👥 Population Portrait
Demographics · finance
Demographics · finance
📍 AOI / POI
Location attributes
Location attributes
🚘 NEV Data
Static vehicle counts
Static vehicle counts
🛣️ Traffic Flow
Dynamic + real-time
Dynamic + real-time
⚡ Existing EVC Sites
Competitor landscape
Competitor landscape
↕ Multi-source Data Ingestion · Crawler · Telecom APIs · Map APIs
SOURCE DATA
📡 Telecom Providers
🗺️ Map / GIS Companies
🚕 Ride-hailing Platforms
🏦 Insurance Data
🕸️ Web Crawlers
Stack:
Spark MLlib
Python
GIS / Spatial AI
SaaS Platform
REST APIs
03
Data Dashboard & Market Context
Market scale, data coverage, and PoC resource breakdown
EV Site Planning PoC — Project Dashboard
2023 · Guangzhou Pilot
13.1M
EVs in China (2022)
↑ +67% year-on-year growth
6
Core ML Models Deployed
↑ Kernel density · OD/PA · Spatial · …
5+
Data Source Categories
↑ Population · NEV · POI · Traffic · EVC
$40K
PoC Investment (DS&E)
↑ SaaS + validation report
ML Algorithm Coverage by Module
Data Source Diversity
04
Core ML Models
6 spatial AI and statistical models forming the site-scoring engine
Model 01
🌐 Kernel Density Model
Collects feature data within a defined analysis radius, calculates the density of relevant features (population, POIs, NEVs) in the surrounding area. Produces spatial heatmaps to surface demand hotspots.
Spatial statisticsBandwidth selection
Model 02
🚗 OD / PA Flow Model
Processes tens of millions of smart mobile device records to model city-level population travel patterns. Derives Origin-Destination and Production-Attraction matrices to estimate actual demand flows to candidate sites.
Big data processingFlow estimation
Model 03
📐 Spatial Interpolation Model
Analyses the spatial extent of feature distributions and finds the nearest input sample subset to predict feature values at unsampled locations across the entire study area.
IDW / KrigingGIS interpolation
Model 04
🕸️ Force-Oriented Layout Model
Draws a topological relationship network by calculating correlation degrees between all features — visualises the aggregation and compactness of site-relevant factors to reveal key driver clusters.
Graph layoutCorrelation analysis
Model 05
🏙️ Business District Analysis Model
Analyses customer situation, site conditions, and surrounding operational factors within a defined commercial area to identify the optimal location address within a business district.
Commercial circleBoundary analysis
Model 06
💎 Value Factors Analysis Model
Identifies high-value sites by analysing key success factors from existing good-performing sites, then forms a location selection model to systematically discover new high-potential sites.
Feature importanceGood-site learning
05
Data Sources & Feature Engineering
5 data categories across demand signals, supply data, and geospatial intelligence
Population Portrait
Population size, demographics (gender, age), lifestyle preferences (travel, shopping), and financial profiles — used to model demand density and user segment affinity.
Source: Telecom companies · Insurance companies
AOI / POI (Area & Point of Interest)
Name, address, coordinates, telephone and other attributes for commercial zones, residential areas, transit hubs, and destinations — core inputs for spatial feature extraction.
Source: Map companies · Crawler & data processing
New Energy Vehicle Data (Static)
NEV count by brand and model segmented by geographic area — identifies actual EV ownership concentration and reveals high-demand zones.
Source: Insurance companies · Crawler
Traffic Flow Data (Dynamic)
Road grade, time-of-day segmented traffic volumes, real-time NEV counts on road, e-hailing NEV booking counts, and GPS tracking within each POI — provides demand dynamics.
Source: Ride-hailing platforms · Telecom companies
Existing EV Charging Site Data
Competitor and existing site attributes: name, pile count, pricing, real-time availability (free/busy), and power rating — informs supply gap and competitive landscape analysis.
Source: Public charging networks · Web crawlers
Derived Features (computed)
Commercial circle boundaries · Opportunity zones · Peer radiation areas · Competitive zones · Equidistance circles · Surrounding business density — all derived via the 6 ML models from raw source data.
06
Feature Selection & Model Training Pipeline
Rigorous ML pipeline from raw data preprocessing to model optimisation
Filter Methods
Statistical Relevance
Variance selection (low-variance removal)
Pearson & Spearman correlation
Chi-square test
ANOVA (F-statistic)
Kendall's tau
Mutual information method
Wrapper Methods
Model-guided Selection
Recursive Feature Elimination (RFE)
Cross-validated RFE with estimator
Embedded Methods
Penalty-based
L1 / Lasso regularisation
Tree-based feature importance
Dimensionality Reduction
Compression & Decorrelation
PCA (Principal Component Analysis)
LDA (Linear Discriminant Analysis)
Model Evaluation Metrics
KS statistic & PSI
Accuracy, Recall, F1 Score
AUC / ROC curve
Training pipeline:
Data preprocessing→
Feature extraction & selection→
Model training/building→
Model testing→
Model evaluation→
Model optimisation
07
AI Platform — 5 User Modules
End-to-end visual interface for business teams to evaluate and plan EV charging sites
Heat Map
Input: filter conditions (area, EV density, POI type)
↓
Output: colour-coded demand density heat map across city
Site Evaluation
Input: candidate site address
↓
Output: overall composite score + recommendation grade (A/B/C)
Site Recommendation
Input: filter conditions (district, budget, site type)
↓
Output: ranked list of recommended sites sorted by score
Good Sites Analysis
Input: known high-performing site address
↓
Output: key success factors and their relative weights
Performance Forecast
Input: site address + investment cost
↓
Output: customer unit price · annual revenue split · break-even point
08
Project Management & Delivery
End-to-end PoC ownership from vendor selection to model validation
Phase 1 — Market Research
Benchmarking & Vendor Selection
Evaluated multiple China technology providers on data quality, coverage, prior practice experience, and cost. Selected Alading as prioritised partner following structured benchmarking.
Phase 2 — PoC Design
Solution Architecture & Scope Definition
Defined 3-module AI system (Evaluation · Recommendation · Forecast), identified 5 data source categories, scoped Guangzhou as pilot city. Secured $40K DS&E global funding.
Phase 3 — Contract & Build
Contract Signed → First Deployment
Signed contract with Alading. Oversaw SaaS platform configuration with heat map, site evaluation, site recommendation, good-site models, and performance forecast modules.
Phase 4 — Validation
Trial Launch & Model Accuracy Validation
Compared AI modelling results against Business Development team's on-site investigation findings. Generated validation report covering functions, performance, and model accuracy.
PoC Deliverables
SaaS platform modules5 functional
ML models deployed6 models
Validation reportFunction + accuracy
Total budget$40K USD
Pilot cityGuangzhou
Risk · High
Data Availability
Real-time traffic and NEV tracking data subject to provider access restrictions and coverage gaps.
→ Multi-source fallback · data SLA in contract
Risk · Medium
Model Accuracy
Prediction vs. BD on-site investigation divergence — new domain with limited labelled training data.
→ Iterative retraining · accuracy KPI gates
09
Technology Stack
Spatial AI, distributed ML, and GIS-integrated analytics platform
Spark MLlib
Distributed ML
Python
ML + Data Pipeline
GIS / Spatial AI
Geospatial analysis
PCA / LDA
Dimensionality reduction
Kernel Density
Spatial statistics
Graph / Network
Feature correlation
SaaS Platform
Alading / partner
REST APIs
Data ingestion
Collab. Filtering
Recommendation
Regression Models
Performance forecast
Big Data Pipeline
ETL · preprocessing
RFE / Feature Sel.
Model optimisation