EVNextTrade: Learning-to-Rank-Based Recommendation of Next Charging Nodes for EV-EV Energy Trading
Authors: Md Mahfujur Rahmana, Alistair Barros, Raja Jurdak, Darshika Koggalahewa
Organizations: School of Information Systems, Queensland University of Technology, Brisbane, QLD, 4000, Australia · School of Computer Science, Queensland University of Technology, Brisbane, QLD, 4000, Australia
Peer-to-peer energy trading among electric vehicles (EVs) has been increasingly studied as a promising solution for improving supply-side resilience under growing charging demand and constrained charging infrastructure. While prior studies on EV-EV energy trading and related EV research have largely focused on transaction management or isolated mobility prediction tasks, the problem of identifying which charging nodes are more suitable for EV-EV trading in journey contexts remains open. We address this gap by formulating next charging nodes recommendation as a learning-to-rank problem, where each EV decision event is associated with a set of candidate charging locations. We propose a supervised ranking framework applied to a large-scale urban EV mobility dataset comprising millions of journey records and multidimensional EV trading-related features, including EV energy level, trading role, distance to charging locations, charging speed, and temporal station popularity. To account for uncertainty arising from the mobility of both energy providers and consumers, as well as the presence of multiple viable charging nodes at a decision point, we employ probabilistic relevance refinement to generate graded labels for ranking. We evaluate gradient-boosted learning-to-rank models, including LightGBM, XGBoost, and CatBoost, on EV journey records enriched with candidate charging nodes. Experimental results show that LightGBM consistently achieves the strongest ranking performance across standard metrics, including NDCG@k, Recall@k, and MRR, with particularly strong early-ranking quality, reflected in the highest NDCG@1 (0.9795) and MRR (0.9990). These results highlight the effectiveness of uncertainty-aware learning-to-rank for charging node recommendation and support improved coordination and matching in decentralized EV-EV energy trading systems.
Figures & tables
Field/Features
Category
Static/dynamic
Description
Ref
Used in EV–EV trading
VId
Identifier
S
Unique identifier for the conventional vehicle
[ 25 , 20 , 36 , 39 , 18 , 47 , 14 ]
Yes
Vehicle model
Vehicle
S
Name of the vehicle model
[ 30 ]
Yes
Timestamps
Temporal
D
Current date and time
[ 25 , 30 , 20 , 36 , 39 , 18 , 47 , 14 ]
Yes
Locations (latitude, longitude)
Spatial
D
Locations of vehicle at different time
[ 25 , 30 , 20 , 36 , 39 , 18 , 47 , 14 ]
Yes
Stay locations
Spatial
D
Vehicles stay location at different stoppage
[ 25 ]
No
Visit frequency
Behavioral
D
Indicate frequent visit at specific locations
[ 25 , 20 , 42 ]
No
Table 1: Common data attributes or features used in existing studies for supporting mobility prediction: mobility data description
Figure 1: Overall process for the EV–EV trading datasets generation.
Field
Type
Static/dynamic
Description
EV journey ID
String
D
Unique identifier for each journey
EV ID
String
S
Unique identifier for each EV
EV Model name
String
S
An integer from 1 to 9 representing the EV model type allows easy vehicle category identification
Battery capacity
int64
S
Represents the maximum energy storage capacity of the EV’s battery, measured in watt-hours (Wh), indicating the total energy the battery can store.
Battery range
int64
D
Represent the maximum distance the EV can travel on a full charge, measured in kilometers. This value is reduced after every fare (trip) based on the km covered in the journey.
Community area ID
int64
S
An integer between 1 and 77, where each ID represents a designated community area in Chicago. Each community area is a geographic unit.
Table 2: Summary of Dataset Attributes for EV–EV energy trading
Figure 2: Distribution of trips per EV model shown as a bar plot, ordered from most to least frequent model.
Figure 3: Distribution of unique EVs per model (distinct EV IDs), ordered by model frequency.
Figure 4: Spatial distribution of EV locations by energy level (20–40%, 40–60%, 60–80%, 80–100%) for 30-minute windows, high peak (08:30, 17:30) and off-peak (06:00, 22:00) periods; legends report per energy level counts.
Figure 5: Label generation pipeline. Unlabeled EV decision events from the NextTrade-EV dataset are first evaluated using fuzzy TOPSIS to compute multi-criteria station suitability scores. These scores are subsequently refined through EM-based probabilistic modeling to capture latent contextual uncertainty. The resulting soft relevance scores are discretized into graded relevance labels, yielding a weakly labeled dataset suitable for learning-to-rank without requiring observed station selection data.
Figure 6: Spatial distribution of EV charging infrastructure and mobility patterns in Chicago.
Category
Feature
Description
Candidate-level
Distance to station
Distance from the EV’s current location to the candidate charging station, capturing travel effort and feasibility.
Charging speed
Maximum charging or discharging rate supported by the station, reflecting energy transfer capability.
Station popularity
Historical usage intensity of the station, reflecting recurring demand and supply pressure and local availability conditions that influence EV–EV matching feasibility.
EV-level
Battery capacity
Maximum battery capacity of the EV.
Battery energy level
Available battery energy at the decision point.
State-of-charge ( SoCe )
Normalized battery energy level, capturing charging or discharging urgency.
Table 3: Feature set used for learning-to-rank in EV–EV charging node recommendation.The following EV-level, temporal, spatial, context-level, and vehicle-level attributes are treated as contextual features, as they characterize the decision environment under which candidate charging nodes are ranked.
Figure 7: Pearson Correlation Analysis among features: This analysis measures the linear relationships among features, identifying potential multicollinearity and redundant information.
Figure 8: Next charging nodes recommendation framework based on learning-to-rank.
Figure 9: Query-level training, validation, and testing protocol for learning-to-rank models.
Model
NDCG@1
Recall@1
NDCG@3
Recall@3
NDCG@5
Recall@5
NDCG@10
Recall@10
MRR
LightGBM (LambdaRank)
0.9795
0.1931
0.9725
0.5112
0.9762
0.6848
0.9811
0.8889
0.9990
XGBoost (Pairwise)
0.9655
0.1926
0.9623
0.5119
0.9689
0.6854
0.9749
0.8899
0.9984
CatBoost (PairLogit)
0.9678
0.1924
0.9646
0.5111
0.9709
0.6854
0.9772
0.8900
0.9982
Table 4: Comparison of learning-to-rank models for next charging nodes recommendation
Figure 10: LightGBM LambdaRank convergence and early stopping. Training (blue) and validation (purple) NDCG@10 versus boosting iteration for the LightGBM ranker. The best iteration selected by early stopping occurs at Iter = 499 with a peak validation NDCG@10 = 0.9716 (blue marker; dashed vertical line).
Threshold Configuration ( r )
Providers (%)
Consumers (%)
NDCG@1
Recall@1
NDCG@5
Recall@5
NDCG@10
Recall@10
(r≤0.30,r>0.85)
55.99
44.01
0.9793
0.1930
0.9761
0.6849
0.9811
0.8889
Baseline (r≤0.30,r>0.90)
56.48
43.52
0.9795
0.1931
0.9762
0.6848
0.9811
0.8889
(r≤0.30,r>0.95)
56.78
43.22
0.9795
0.1931
0.9763
0.6849
0.9812
0.8889
Table 5: Sensitivity Analysis of SoC Thresholds
Radius (km)
#Events
Mean
Median
P90
P95
0 WCPs (%)
≤3 (%)
≤5 (%)
≤10 (%)
1
2190331
7.21
6
16
18
16.84
28.72
48.31
74.80
2
2190331
8.64
8
16
19
2.67
17.75
38.16
68.32
3
2190331
9.07
8
16
19
0.23
13.85
35.33
66.34
5
2190331
9.14
8
16
19
0.02
13.18
35.27
66.33
10
2190331
9.15
8
16
19
0.00
12.54
35.15
66.32
Table 6: Distribution of candidate WCPs per EV decision event under varying geographic radius limits.
Figure 11: Percentage of decision events with zero candidate WCPs under varying geographic filtering radius (1–10 km). The proportion of infeasible events decreases sharply as the radius expands, dropping from 16.84% at 1 km to nearly zero beyond 5 km. This demonstrates that small radius can lead to substantial candidate sparsity, whereas radius ≥ 3 km effectively ensure candidate availability for almost all decision events.
Figure 12: The empirical cumulative distribution of candidate charging pads per decision event under varying radius limits. For a given value x , the cumulative distribution represents the proportion of decision events with at most x candidate WCPs. The curves demonstrate that candidate availability stabilizes beyond 3 km, as the distributions for 5 km and 10 km are nearly indistinguishable.
Radius (km)
NDCG@1
Recall@1
NDCG@3
Recall@3
NDCG@5
Recall@5
NDCG@10
Recall@10
1
0.9764
0.2212
0.9693
0.5160
0.9726
0.6866
0.9782
0.8862
2
0.9750
0.2442
0.9668
0.5164
0.9704
0.6840
0.9765
0.8828
3
0.9768
0.2040
0.9679
0.5115
0.9716
0.6817
0.9782
0.8848
5
0.9796
0.1959
0.9705
0.5143
0.9746
0.6854
0.9813
0.8886
10
0.9796
0.1930
0.9706
0.5123
0.9747
0.6842
0.9813
0.8881
Table 7: Sensitivity of ranking performance to candidate-generation radius.
Variant
EM
Context
NDCG@k
Recall@k
MRR
@1
@3
@5
@10
@1
@3
@5
@10
TOPSIS labels + full features
×
✓
0.9423
0.9616
0.9724
0.9759
0.2516
0.7401
0.9596
0.9986
0.9981
TOPSIS labels + candidate-only features
×
×
0.7508
0.8478
0.8845
0.8989
0.2426
0.6936
0.9120
0.9925
0.9764
EM labels + candidate-only features
✓
×
0.8802
0.8933
0.9130
0.9317
0.1902
0.5061
0.6815
0.8835
0.9960
EM labels + full features (Full Model)
✓
✓
0.9795
0.9725
0.9762
0.9811
0.1931
0.5112
0.6848
0.8889
0.9990
Table 8: Ablation study evaluating the impact of probabilistic smoothing (EM) and feature components on ranking performance using LightGBM (LambdaRank).
Vehicle-to-vehicle (V2V) energy trading enables decentralized peer-to-peer energy exchange among electric vehicles (EVs), reducing grid dependency while monetizing surplus capacity. However, coordinating self-interested EV agents with diverse charging needs and uncertain arrival-departure schedules remains challenging. Existing approaches either require centralized optimization with computational limitations or lack fairness guarantees. This paper integrates Nash Bargaining Solution into Multi-Agent Deep Deterministic Policy Gradient, namely Nash-MADDPG, for incentive-aligned V2V energy trading. Nash bargaining determines efficient bilateral pricing, while Nash-guided price proximity rewards align agent learning toward bargaining-optimal strategies. Evaluation over 30-day continuous operation demonstrates an improvement of 61.6% in social welfare and 62.9% improvement in trading volume over Double Auction, while achieving superior fairness, such as 40.1% improvement in Jain's index. Testing across 6-100 agents over a 30-day horizon with continuous vehicle turnover confirms scalability across population size and empirically stable pricing near the Nash Bargaining benchmark.
Yujin Lin, Yue Yang, Hao Wang
Department of Data Science and AI, Faculty of IT, Monash University, Australia · Monash Energy Institute, Monash University, Australia
The electrification of transportation through electric vehicles introduces new challenges for power grid management, such as increased peak demand, voltage fluctuations, line overloads, and the integration of variable renewable energy sources. To enable efficient integration of EVs while minimizing costs for users and avoiding network overloads, implicit coordination between EVs is required. This work compares two independent multi-agent reinforcement learning approaches for optimizing such decentralized EV charging: contextual combinatorial bandits and policy gradient algorithms. Using a realistic simulation environment with autonomous agents making decisions based on local environmental information (including price signals, state-of-charge, and temporal constraints), we evaluate their performance across varying congestion levels, and mixed-strategy configurations with heterogeneous agent groups under dynamic electricity pricing derived from real photovoltaic production data.
Accurate forecasting of electric vehicle (EV) charging demand is critical for grid stability, infrastructure planning, and real-time charging optimization. In this work, we study the problem of early prediction of charging demand, where the total energy of a session is estimated using only information available at plug-in time and during the first minutes of charging. This enables actionable decisions while the session is still in progress, which is of direct importance for EV network operators. We construct a session-level dataset from the Adaptive Charging Network (ACN), combining session metadata with early-window charging measurements, and derive tabular features capturing user intent, temporal patterns, and initial charging behavior. We focus on a single operational depot, Caltech, and model intra-depot heterogeneity through station-level client partitions while evaluating multiple model families in a federated learning (FL) setting. Our results show that federated models can approach centralized predictive performance while keeping data in-depot, enabling privacy-enhanced training across distributed charging infrastructures. Overall, we demonstrate that reliable demand estimates can be obtained early in the session with minimal data, and that FL provides a practical pathway toward scalable and privacy-aware analytics for EV charging networks. Code is available at https://github.com/Indigma-Innovations/federated-learning-ev-charging-demand.