Source-linked AI summary
Electric Vehicle Charging Station Location Selection using Geospatial Artificial Intelligence (GeoAI)
Eun Hak Lee, Euntak Lee
TL;DR
Charging-station siting studies often prioritize costs or travel distance, while the spatial context of existing stations is also relevant to their operations. This study uses a GeoAI VAE-GCN framework to assess candidate locations, outperforming benchmark models and producing different outcomes under similarity- and distance-focused policies.
Problem
Many charging-station location studies focus on cost or travel distance, while the geospatial characteristics surrounding existing stations remain insufficiently examined.
Source: §1 · p. 2, §2.3 · p. 7Method
The GeoAI VAE-GCN framework compresses high-dimensional EV-related inputs into a latent representation and uses it to assess candidate locations’ geospatial similarity to existing stations.
Source: §1 · p. 2Results
The VAE-GCN model outperforms GCN, MLP, RF, XGBosost, and CatBoost benchmarks, while the two policy scenarios produce different outcomes suited to distinct strategic objectives.
Source: §5 · p. 24Takeaways & Limitations
The findings indicate that candidate siting informed by geospatial similarity can reinforce demand centers, support route continuity, and fill geographic gaps at the regional scale.
Source: §5 · p. 24The study identifies future work on feature uncertainty and notes its reliance on four aggregated datasets.
Source: §5 · p. 24
Abstract
from arXiv · showhide
As electric vehicle (EV) adoption increases, ensuring efficient and well-distributed charging infrastructure has become a critical challenge. While many EV charging station location problem (CSLP) studies focus on minimizing costs or travel distance, it is crucial to consider the surrounding geospatial characteristics of existing stations that influence operational performance. This study proposes a geospatial artificial intelligence (GeoAI)-based framework that integrates high-dimensional EV-related geospatial data, including EV usage, land-use, population, and traffic attributes. We incorporate a variational autoencoder (VAE) and a graph convolutional network (GCN) into the model to capture similarities among existing charging stations, and to identify suitable locations for future stations. The VAE compresses high-dimensional EV input data into a low-dimensional latent space, and the GCN uses this latent representation to predict locations suitable for charging stations. Using real-world data from Bryan-College Station, Texas, US, the proposed model outperforms state-of-the-art baselines, achieving an F1-score of 0.87 in distinguishing existing station locations from non-station locations. The model also identifies 27 additional candidate locations that show geospatial characteristics similar to those of existing stations, based on a similarity score. We further evaluate two policy implementation scenarios, maximizing geospatial similarity and minimizing total travel distance, each yielding different outcomes aligned with distinct strategic objectives. The findings highlight the importance of incorporating spatial context into CSLP and provide valuable insights for future EV infrastructure planning, promoting both efficiency and accessibility in the rapidly growing electric mobility sector.
1. Introduction
The study targets a gap in charging-station siting: existing approaches often emphasize costs or travel distance without examining stations’ surrounding geospatial relationships. It proposes a GeoAI VAE-GCN framework and evaluates it with real-world data from Bryan-College Station, Texas.
1. Introduction
- Charging supply-demand imbalances include longer queues and travel to find stations, and providers cannot always add stations when shortages arise.The paper motivates managing these imbalances within resource constraints.Source: §1 · p. 2
- Prior CSLP studies largely emphasized operating costs or travel distance, leaving the effects of geospatial attributes and station interdependence unexamined.The paper frames surrounding spatial characteristics as relevant to station operations.Source: §1 · p. 2
- The VAE-GCN framework assesses candidate locations’ geospatial attributes and measures their similarity to existing stations.The study describes reducing high-dimensional inputs to a low-dimensional representation.Source: §1 · p. 2
- The proposed model outperforms existing state-of-the-art models using data from a 185 km2 urban network in Bryan-College Station, Texas.Candidate locations are assessed and ranked by similarity to existing stations.Source: §1 · p. 3
- The study’s contributions include integrating spatial characteristics, designing a VAE-GCN model, and comparing policies that maximize similarity or minimize travel distance.The policy assessment considers priorities including spatial characteristics, user convenience, and accessibility.Source: §1 · p. 3
2. Literature review
The literature review identifies limited attention to charging stations’ spatial context and to high-dimensional data challenges in CSLP. It positions the study’s VAE-GCN approach and two policy scenarios as responses to these gaps.
2.1. EV charging station location problem
- Prior CSLP work has addressed budget and demand constraints, but the spatial relationships among stations remain underexamined.The review describes existing work as primarily focused on measurable operational resources and cost or distance objectives.Source: §2.1 · pp. 3, 5
2.2. Artificial intelligence-based algorithms
- AI-based CSLP studies incorporate factors such as demand, traffic, land use, infrastructure, and station distances, but do not capture the spatial characteristics tied to station performance.The review also notes concerns about computational burden from high-dimensional inputs and limited integration of VAE and GCN in GeoAI.Source: §2.2 · pp. 5–7
2.3. Key research gaps
- The study identifies gaps in spatial-context modeling, handling high-dimensional inputs, and comparing siting policies across objectives.It proposes a compact latent representation and compares maximizing geospatial similarity with minimizing travel distance.Source: §2.3 · pp. 7–8
3. Modeling framework
The framework pairs a VAE that learns compact representations of EV station data with a GCN that uses spatial relationships to predict suitable candidate locations. Its training and ranking workflow produces suitability scores for candidate sites.
3. Modeling framework
- The two-phase framework learns existing stations’ spatial distributions with a VAE, then uses that information to propose optimal locations for additional stations.The low-dimensional representation from the VAE is passed to the GCN.Source: §3 · pp. 8–9
3.1. VAE phase for deriving latent space
- The VAE encodes high-dimensional EV data into a latent space and reconstructs the input, training a compact representation with reconstruction and KL-divergence terms.The reconstruction term measures how well the decoder reproduces the input; the KL term encourages the approximate posterior to remain close to the prior.Source: §3.1 · pp. 9–10
3.2. GCN phase for finding optimal station location
- The proposed GCN processes graph-structured inputs, combining neighboring-node information with the VAE-generated latent space z to learn node representations and capture spatial dependencies between EV charging stations.Its first layer uses ReLU for non-negative feature propagation, and its output sigmoid yields a value between 0 and 1.Source: §3.2 · pp. 11–12
3.3. Optimization process
- The workflow trains the VAE and GCN, scores locations without existing stations, ranks candidates by suitability, and selects the top K sites.The GCN uses a normalized adjacency matrix to propagate latent features and binary cross-entropy over nodes with known station labels.Source: §3.3 · p. 13
4. Application
The Bryan-College Station application combines EV, urban, and traffic data in a road-network buffer analysis to assess station patterns and candidate sites. It compares model performance, ranks new locations, and evaluates two station-siting policies.
4.1. Dataset and model settings
- Four datasets feed a road-network buffer analysis; 1,971 buffers across 185 km² use station presence as the target, with 500 m buffers spaced every 100 m.The analysis excludes unsuitable terrain and proportionally assigns building area and road length when features intersect buffer boundaries.Source: §4.1 · pp. 14–15
4.2. Geospatial similarity analysis between existing nodes
- Similarity patterns identify N10-N15 as strong non-station candidates, while N5-N6 have weak station similarity; station nodes N16-N22 are mutually similar, unlike N23.The analysis also reports clustering among non-station nodes with similar station similarity.Source: §4.2 · p. 17, Figure 5 · p. 17
4.3. Model performance
- 0.87 F1 score: VAE-GCN leads overall, with precision 0.80 and recall 0.97.The authors also note that precision 0.80 reflects some over-prediction of station-free buffers with similar spatial characteristics.Source: §4.3 · p. 19, Table 3 · p. 19
4.4. Locations for new EV charging stations
- 27 candidate locations score above 0.5; the highest-priority candidate L1 scores 0.71 and has 15 similar existing stations.The study interprets scores above 0.5 as indicating high installation suitability.Source: §4.4 · pp. 19–20, Table 4 · p. 20
- Candidates cluster around central urban areas and mobility corridors; commercial areas host many proposed sites, while fewer candidates occur in green or low-activity areas and at the periphery.The paper describes peripheral sites as strategically placed to extend coverage into suburban or less-developed areas.Source: §4.4 · pp. 21–22, Figure 7 · p. 22
4.5. Expected impact of new EV charging stations
- 11.6% travel reduction under distance minimization exceeds the 7.0% reduction under similarity maximization.The similarity-maximizing scenario yields 70.0% average similarity; the authors associate the policies with different strategic priorities.Source: §4.5 · p. 23, Table 5 · p. 23
5. Conclusions
The GeoAI-based VAE-GCN outperforms the compared benchmarks, while two policy scenarios yield different outcomes suited to distinct strategic objectives. The conclusions connect geospatially informed siting to regional planning and identify data and modeling limitations for future work.
5. Conclusions
- The VAE-GCN outperforms GCN, MLP, RF, XGBosost, and CatBoost, and its two policy scenarios suit different strategic objectives.The scenarios maximize geospatial similarity or minimize total travel distance.Source: §5 · p. 24
- Jointly considering spatial characteristics and existing features identifies similarity patterns for siting candidates that reinforce demand centers, connect major routes, and fill geographic gaps.The conclusions describe these locations as supporting a more balanced and efficient regional charging infrastructure.Source: §5 · p. 24
- Compressing high-dimensional inputs into a low-dimensional latent space improves computational efficiency and prediction accuracy, while feature relationships need careful examination when predictions overstate candidates or miss existing stations.The conclusions also flag risks of longer queues or travel distances from supply-demand imbalances.Source: §5 · p. 24
- Charging plans should align similarity maximization or travel-distance minimization with regional EV adoption, population, land use, and road-network conditions.The conclusions state that the two policies produce different outcomes.Source: §5 · p. 24
- The study relies on four aggregated datasets under specific assumptions; future work should examine feature uncertainty, model efficiency and interpretability, fine-grained spatial representation, and dynamic data.Examples of dynamic data include projected population growth and planned urban developments.Source: §5 · p. 24