How Can Artificial Intelligence Predict Electric Vehicle Charging Times?

Imagine planning a road trip in an electric car without knowing how long the charging will take. A new study merges two approaches to solve this: a physics‑based model describing constant‑current/constant‑voltage (CC‑CV) charging dynamics and a reinforcement learning (RL) algorithm. The physics model starts from the battery’s electrochemical laws, including residual capacity and power loss linked to state‑of‑health (SoH). Even with limited historical data, it provides a reliable estimate.

The second layer is RL, which watches real charging sessions and continuously refines predictions. As the battery ages, the algorithm learns to compensate for degradation, keeping accuracy high. Researchers tested the system on 5,000 simulated sessions, calibrated with manufacturer specs and public datasets. The physics model achieved an R² of 98.5 % and a MAPE of 2.1 %, while the hybrid with RL reached 99.2 % R² and 1.6 % MAPE, boosting accuracy by 23 % and robustness to battery aging by 35 %.

This blend promises more predictable charging stops, enhancing driver experience and charging‑infrastructure efficiency.

The physics model relies on the CC‑CV charging equation, where the constant‑current (CC) phase rapidly charges the battery up to a predefined voltage threshold, after which the constant‑voltage (CV) phase gradually reduces the current to avoid overcharging. The effective capacity C(t) is modeled as C₀·(1‑β·SoH), where β represents the degradation rate per cycle. The peak power Pₘₐₓ is also reduced as a function of SoH, reflecting the loss of power delivery capability.

For the reinforcement learning component, a Deep Q‑Network (DQN) was employed with the state defined by: SoH, battery temperature, current state‑of‑charge (SOC), input power, and the time estimate from the physics model. The action is the correction to the charging‑time prediction, while the reward is inversely proportional to the absolute error between the corrected prediction and the observed real time. The algorithm uses a replay buffer of 10,000 experiences and a target network updated every 500 steps to stabilize learning.

The simulated dataset includes 5,000 sessions with variations in battery model, environmental conditions, and charging profiles (fast‑charging, slow‑charging). The physics model parameters were calibrated with real data from Tesla, Nissan Leaf, and Chevrolet Bolt, ensuring good generalizability. Results show that integrating the DQN reduces MAPE from 2.1 % to 1.6 % and enhances robustness to SoH variations, thanks to the RL's ability to learn non‑linear patterns not captured by the analytical model. This hybrid approach can be extended to grid management systems, optimizing real‑time energy distribution.