A multi-algorithm optimization framework for cooperative electric vehicle fleet routing and charging on realistic road networks. Compares five approaches — Simulated Annealing (SA), Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Teaching-Learning Based Optimization (TLBO), and Deep Q-Network (DQN) — under nonlinear charging physics, shared station capacity with FIFO queuing, and discrete speed-level trade-offs.
Paper — Comparative Metaheuristic and Reinforcement Learning Optimization for Cooperative Electric Vehicle Fleet Routing with Nonlinear Charging and Shared Station Capacity (PDF)
The system routes multiple heterogeneous EVs through fork-shaped road networks with shared charging stations, jointly minimizing a weighted objective:
subject to SOC bounds (10–100 %), energy balance along each route, station capacity with FIFO queuing, nonlinear SOC-dependent charging (
- Nonlinear charging physics — SOC-dependent power curves with 95 % grid-to-battery efficiency
- Queue modeling — FIFO discipline at shared stations with limited plug counts
- 5 discrete speed levels per road edge (see table below)
- Heterogeneous fleet — 3–7 EVs with 40–80 kWh batteries
- Interactive GUI — Streamlit app with real-time animated simulation
- HTML dashboards — per-solution interactive reports with timelines, SOC traces, and queue stats
| Level | Description | Time Factor | Energy Factor |
|---|---|---|---|
| 1 | Very Slow | +40 % | −25 % |
| 2 | Slow | +20 % | −10 % |
| 3 | Normal | Baseline | Baseline |
| 4 | Fast | −15 % | +20 % |
| 5 | Very Fast | −30 % | +50 % |
| Algorithm | Key Mechanism | Parameters | Best Scenario |
|---|---|---|---|
| SA | Trajectory search with geometric cooling ( |
|
Simple networks |
| GA | Tournament selection, two-point crossover, elitism (top 10–15 %) | pop = 50, 100 gens, |
Congested / peak demand |
| PSO | Adaptive inertia ( |
swarm = 50, 100 iters | Uncongested networks |
| TLBO | Teacher + Learner phases, hybrid blending for mixed variables | pop = 60, 150 iters, zero hyperparams | Limited tuning resources |
| DQN | Dueling DQN + Prioritized Replay + Double DQN, 25-dim state | 500 episodes, |
Time-critical applications |
Seven scenarios of increasing complexity were evaluated on single-fork and double-fork topologies:
| Scenario | EVs | Stations | Notes |
|---|---|---|---|
| Single-Small | 3 | S1/S2/S3 | Standard pricing |
| Single-Large | 7 | S1/S2/S3 | Standard pricing |
| Double-Sparse | 4 | 6 stations | Standard pricing |
| Double-Balanced | 5 | 6 stations | Standard pricing |
| Double-Premium | 4 | 6 stations | Standard pricing |
| Double-Congested | 6 | −1 plug/stn | Standard pricing |
| Double-High Demand | 7 | S2/S5 upgraded | +15 % pricing |
| Scenario | Algorithm | Objective ( |
Makespan (min) | Cost (EGP) |
|---|---|---|---|---|
| Single-Small | SA | 757.2 | 304.2 | 453.1 |
| GA | 291.6 | 271.6 | 20.0 | |
| PSO | 274.3 | 274.3 | 0.0 | |
| TLBO | 363.6 | 285.4 | 78.2 | |
| Dbl-Balanced | SA | 2925.2 | 495.8 | 2429.4 |
| GA | 2089.1 | 493.3 | 1595.8 | |
| PSO | 1856.2 | 516.0 | 1340.2 | |
| TLBO | 2742.4 | 492.9 | 2249.4 | |
| Dbl-Premium | SA | 2803.8 | 502.8 | 2301.0 |
| GA | 1919.9 | 469.9 | 1450.0 | |
| PSO | 1254.5 | 513.3 | 741.2 | |
| TLBO | 2453.3 | 472.8 | 1980.6 |
| Scenario | Algorithm | Objective ( |
Makespan (min) | Cost (EGP) |
|---|---|---|---|---|
| Baseline (3 EVs) | SA | 2885.0 | 497.1 | 2388.0 |
| GA | 1805.8 | 475.5 | 1330.2 | |
| PSO | 1483.6 | 520.7 | 962.9 | |
| TLBO | 2386.7 | 486.1 | 1900.6 | |
| RL | 4697.9 | 473.1 | 4224.7 | |
| Congested (6 EVs) | SA | 3495.1 | 551.5 | 2943.6 |
| GA | 2321.9 | 487.2 | 1834.7 | |
| PSO | 3447.9 | 562.5 | 2885.4 | |
| TLBO | 3150.9 | 500.8 | 2650.1 | |
| RL | 5868.1 | 468.2 | 5399.9 | |
| High Demand (7 EVs) | SA | 4657.3 | 505.5 | 4151.8 |
| GA | 2851.3 | 486.7 | 2364.6 | |
| PSO | 4073.0 | 502.9 | 3570.0 | |
| TLBO | 3644.5 | 495.7 | 3148.7 | |
| RL | 7563.9 | 482.1 | 7081.8 |
- PSO dominates uncongested scenarios — 18–40 % lower weighted objective than the next-best algorithm, driven by minimal charging costs.
- GA is superior under congestion — maintains consistent performance as fleet size increases (3 → 7 EVs), outperforming PSO which drops from best to third-best.
- TLBO is competitive with zero tuning — outperforms SA by 10–25 % despite having no algorithm-specific hyperparameters.
- RL achieves fastest makespans (1–5 % faster) but incurs 183–199 % higher charging costs due to reward function weighting.
- Adaptive inertia for PSO improves convergence by 16 % over static schedules (1 483.6 vs 1 760.8 after 150 iterations).
- Extended RL training (1 000 episodes) paradoxically worsens performance, suggesting overfitting.
Figure 1 — Five-algorithm comparison across double-fork scenarios. GA maintains lowest cost under congestion; RL has the highest cost but lowest queue times.
Figure 2 — Benchmark comparison of PSO, GA, and SA on the double-fork baseline. PSO achieves the lowest weighted objective and cost.
Figure 3 — SA vs GA vs PSO across single baseline, congested grid, and high-demand scenarios showing weighted objective, makespan, cost, and queue time.
Figure 4 — Adaptive inertia achieves 16 % lower final objective than a static schedule on the double-fork scenario.
Figure 5 — PSO convergence on the double-fork scenario; objective improves steadily before flattening near iteration 150.
Figure 6 — PSO hyperparameter grid search heatmap ($w \in \{0.4, 0.6, 0.8\}$, $c_1 = c_2 \in \{1.5, 1.7\}$, swarm $\in \{30, 45, 60\}$). Best: $w = 0.6$, $c_1 = c_2 = 1.5$, swarm 45.
├── algorithms/
│ ├── ga/genetic_algorithm.py # Genetic Algorithm
│ ├── pso/particle_swarm.py # PSO with adaptive inertia
│ ├── pso/compare_weights.py # Static vs adaptive inertia study
│ ├── pso/pso_experiments.py # PSO across case studies
│ ├── pso/pso_parameter_sweep.py # Hyperparameter grid search
│ ├── sa/simulated_annealing.py # Simulated Annealing
│ ├── tlbo/teaching_learning_optimization.py # TLBO (parameter-free)
│ ├── rl/ # DQN agent + environment
│ └── examples/example_algorithm.py # Template for new algorithms
│
├── common/
│ ├── params.py # Network topology, fleet, station config
│ ├── objectives.py # Objective functions + queue processing
│ ├── feasibility_repair.py # Constraint repair utilities
│ ├── case_studies.py # Benchmark scenario builders
│ └── visualization.py # HTML dashboard generator
│
├── scripts/
│ ├── compare_all_algorithms.py # Full 5-algorithm comparison
│ ├── run_case_studies.py # Batch runner with JSON/Markdown export
│ ├── run_metaheuristic_studies_parallel.py # Parallel metaheuristic runs
│ ├── run_parameter_sensitivity.py # Parameter sensitivity analysis
│ ├── visualize_comparison.py # Rich HTML comparison reports
│ └── test_*.py # Validation scripts
│
├── docs/ # Algorithm documentation and guides
├── outputs/ # Results (JSON, CSV, plots, dashboards)
├── app.py # Streamlit GUI application
├── requirements.txt
├── .gitignore
└── LICENSE
git clone https://github.com/youssefrfarid/EV-Fleet-Routing-Optimization.git
cd EV-Fleet-Routing-Optimization
pip install -r requirements.txtpython algorithms/sa/simulated_annealing.py # SA demo + dashboard
python algorithms/ga/genetic_algorithm.py # GA demo + dashboard
python algorithms/pso/particle_swarm.py # PSO demo + dashboard
python algorithms/tlbo/teaching_learning_optimization.py # TLBO demopython scripts/compare_all_algorithms.py # Full 5-algorithm comparison
python scripts/run_case_studies.py # All scenarios → JSON + tables
python scripts/run_parameter_sensitivity.py # Sensitivity analysisstreamlit run app.pyfrom common.params import make_toy_params
from common.objectives import objective_weighted
params = make_toy_params() # 5 EVs, 3 stations, fork network
# Run any optimizer
from algorithms.ga.genetic_algorithm import run_ga_double_fork_demo
solution = run_ga_double_fork_demo()
# Evaluate
score = objective_weighted(solution, w_time=1.0, w_cost=1.0)| Vehicle | Battery | Initial SOC | Type |
|---|---|---|---|
| EV 1 | 40 kWh | 62% | Compact (Nissan Leaf) |
| EV 2 | 55 kWh | 48% | Sedan (Chevy Bolt) |
| EV 3 | 62 kWh | 45% | Sedan+ (Leaf Plus) |
| EV 4 | 75 kWh | 52% | SUV (Model Y) |
| EV 5 | 80 kWh | 47% | Large SUV (e-tron) |
| Station | Type | Speed | Price (EGP/kWh) | Plugs |
|---|---|---|---|---|
| S1 | Budget | 50 kW | 13.0 | 2 |
| S2 | Premium | 180 kW | 27.0 | 1 |
| S3 | Standard | 100 kW | 20.0 | 1 |
A → J → { Upper: S1 → S2 → M (longer, more stations)
{ Lower: S3 → M (shorter, hillier)
M → B
| Name | Affiliation |
|---|---|
| Andrew Abdelmalak | Mechatronics Engineering, GUC |
| Daniel Ekdawi | Mechatronics Engineering, GUC |
| Daniel Boules | Mechatronics Engineering, GUC |
| David Girgis | Mechatronics Engineering, GUC |
| Youssef Ramy | Media Engineering Technology, GUC |
The full project report is available in docs/EV_Fleet_Routing_Optimization.pdf.
This project was developed as part of the Mechatronics Engineering program at the German University in Cairo (GUC).
- S. Pelletier, O. Jabali, and G. Laporte, "Goods distribution with electric vehicles: Review and research perspectives," Transportation Science, 50(1), 2016.
- M. Schneider, A. Stenger, and D. Goeke, "The electric vehicle-routing problem with time windows and recharging stations," Transportation Science, 48(4), 2014.
- A. Froger et al., "The electric vehicle routing problem with capacitated charging stations," Transportation Science, 56(2), 2022.
- S. R. Kancharla and G. Ramadurai, "Electric vehicle routing problem with non-linear charging and load-dependent discharging," Expert Systems with Applications, 160, 2020.
- M. Keskin et al., "A simulation-based heuristic for the electric vehicle routing problem with stochastic waiting times," Comput. & Oper. Res., 125, 2021.
- S. Karakatic, "Optimizing nonlinear charging times of electric vehicle routing with genetic algorithm," Expert Systems with Applications, 164, 2021.
Apache 2.0 — see LICENSE.