Adaptive battery state-of-health estimation for electric vehicles under diverse driving styles using reinforcement learning.

Singh, S; Wani, NA; Chaudhary, RK · 2026 · PubMed Central

DOI: 10.1038/s41598-026-49169-9

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of accurately estimating the State of Health (SoH) of electric vehicle (EV) batteries under heterogeneous driving conditions. Existing SoH estimation methods, including equivalent circuit models and static machine learning regressors, typically assume stationary operating conditions. Consequently, they struggle to generalize across diverse real-world driving styles—such as urban stop-and-go traffic, mixed logistics routes, and steady highway cruising—which induce distinct battery degradation mechanisms. The authors propose an adaptive framework that utilizes reinforcement learning (RL) to dynamically select the most appropriate prediction model based on real-time driving behavior, thereby improving robustness and accuracy in varying operational scenarios. The methodology employs a multi-expert system combined with a Deep Q-Network (DQN) agent. First, the authors simulate battery dynamics for three specific driving profiles: taxis, delivery vans, and family EVs. These simulations generate time-series data including current, voltage, state of charge, temperature, and internal resistance, incorporating degradation effects. Statistical features are extracted from these signals to train three style-specific gradient boosting regressors, each acting as an "expert" for a particular driving regime. The DQN agent operates within a custom environment to select the optimal expert at each time step. The RL state consists of the current sensor measurements and the previous prediction error, while the reward is defined as the negative squared prediction error, incentivizing the agent to minimize estimation inaccuracies. The agent learns via experience replay and target network updates, allowing it to adapt to transitions between driving styles without explicit labels. Experimental results demonstrate that the proposed adaptive framework achieves a prediction accuracy of over 97%. The model effectively handles abrupt transitions between driving patterns, maintaining stability and robustness where single-model baselines typically degrade. The RL agent successfully learns to associate latent patterns in sensor data with the most suitable expert, leveraging prediction-error feedback to refine its selection policy continuously. This dynamic selection mechanism ensures high fidelity in SoH estimation across the diverse load profiles simulated. The significance of this work lies in its ability to address the non-stationarity of real-world EV usage, which is often overlooked in static estimation models. By integrating RL with specialized machine learning experts, the framework provides a scalable solution for fleet management, enabling more reliable battery health monitoring for ride-sharing and logistics applications. This approach supports safer operation, optimized maintenance schedules, and cost-effective management of diverse EV fleets by ensuring that SoH estimates remain accurate despite varying driving behaviors and environmental factors.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• drowsiness detection algorithms