Coming soon

AI that learns the cheapest way to charge your fleet

Instead of following fixed rules or relying on price forecasts, our AI learns the best charging schedule from your real-world data — and keeps improving over time.

Based on peer-reviewed reinforcement learning research from TU Delft and ETH Zurich

Get early access Explore algorithms

Three RL approaches

Each algorithm tackles the charging problem differently — from exploration-driven learning to constraint-aware optimization.

Safe but Slow

CPO — Constrained Policy Optimization

A safety-first approach that guarantees constraints are respected during training. However, it requires generating fresh data for every update, making it slow and data-hungry — impractical for real-world charging plazas. In testing: €0.163 avg cost with 10.49 kWh in constraint violations.

Key strengths

Built-in safety guarantees during training
Proven mathematical convergence properties
Directly handles constraints without penalty tuning
Monotonic policy improvement at each step

Best for

When you have unlimited training data and time — but impractical for real-world deployment.

Precise but Fragile

DDPG / SAC — with manual penalty tuning

Off-policy algorithms that learn efficiently from past data and output precise kW allocations per socket. The catch: safety constraints must be manually encoded as penalty terms, requiring trial-and-error tuning. Too low a penalty violates safety; too high wastes money. In testing: €0.216–0.283 avg cost with up to 15 kWh in violations.

Key strengths

Sample-efficient — learns from historical data
Precise, continuous power control per socket
DDPG: deterministic, reproducible schedules
SAC: robust to uncertainty via entropy regularization

Best for

When you have time to manually tune penalty coefficients — but risky without expert oversight.

The Best of Both

AL-SAC — Augmented Lagrangian SAC

Combines the data efficiency of SAC with automatic constraint handling — no manual penalty tuning required. The Augmented Lagrangian method automatically learns the right balance between cost and safety. In testing, it achieved the lowest cost (€0.113 avg) with zero constraint violations, outperforming even MPC with 10% forecast error.

Key strengths

Zero constraint violations — guaranteed safety
No manual penalty tuning — fully automated
Data-efficient — reuses past experience for fast learning
Lowest cost of all RL methods in comparative testing

Best for

The best choice for real-world deployment — safe, efficient, and fully automated.

How the AI learns your optimal schedule

The system interacts with a simulated version of your charging plaza, learning from every decision it makes.

The AI training loop

At each timestep the system reads your plaza state, decides how to distribute power, and measures what worked — repeating thousands of times until it finds the cheapest schedule.

Observe

The system reads the current state: EV battery levels, electricity prices, grid limits, and time-of-day.

Decide

A decision model maps the current state to a power plan — how many kW each charger gets this timestep.

Execute

The simulation charges vehicles with those settings, updating battery levels and tracking costs.

Learn

The system measures what worked (low cost + all constraints met) and adjusts its strategy. Over thousands of rounds, it gets better.

Heuristic vs MPC vs RL

How do reinforcement learning approaches compare to the rule-based and optimization methods already in the simulator?

Heuristic vs MPC vs RL
Dimension	Heuristic	MPC	RL NEW
Approach	Hand-crafted rules	Mathematical optimization	Learned from experience
Needs forecasts?	No	Yes — accuracy critical	No — learns implicitly
Adaptability	None — static rules	Re-solve per horizon	Continuous adaptation
Setup effort	Low	Medium — model required	High — training required
Constraint handling	Manual checks	Built into solver	Learned (AL-SAC)
Computation	Instant	Seconds per horizon	Hours to train, ms to infer
Optimality	Feasible, not optimal	Locally optimal	Near-optimal with enough training
Uncertainty handling	Poor	Stochastic MPC possible	Natural — trained on noise

Early access

Let AI optimize your charging costs

We're building AI-powered charging optimization for real-world deployments. Get in touch to discuss a pilot with your data.