New AI-Driven Method Detects Faulty EV Chargers in Real Time
As the global shift toward electric mobility accelerates, the reliability of charging infrastructure has become a critical concern—not just for drivers, but for grid operators, regulators, and automakers alike. A newly published study in Electric Power Construction introduces a breakthrough approach that leverages advanced signal processing and deep learning to detect metering inaccuracies in electric vehicle (EV) charging stations in real time, without requiring on-site inspections or hardware modifications.
The method, developed by a team led by Dongxiang Jiao from the Metrology Center of State Grid Jibei Electric Power Co., Ltd., in collaboration with researchers from Tangshan Power Supply Company and Tianjin University’s School of Electrical and Information Engineering, addresses a growing pain point in the EV ecosystem: the silent drift of charger accuracy over time due to component aging, environmental stress, or tampering. Traditional verification methods rely on manual, periodic calibration—costly, labor-intensive, and increasingly impractical as the number of public and semi-public chargers surges into the millions worldwide.
What sets this new technique apart is its fusion of physics-based modeling with data-driven intelligence. At its core, the system applies the principle of energy conservation across the entire charging station topology. For alternating current (AC) chargers, which deliver power to the vehicle’s onboard converter, the total energy drawn from the grid must equal the sum of energy recorded by individual charging points plus predictable losses from cabling and fixed station loads. For direct current (DC) fast chargers—where AC-to-DC conversion happens inside the charger—the equation must also account for the efficiency of that power conversion process, which varies by model, load, temperature, and other operational factors.
Here’s where the innovation deepens. While past attempts to model these systems often assumed a fixed or average conversion efficiency, the research team recognized that real-world efficiency is dynamic. To capture this variability, they designed a novel deep learning architecture—dubbed Transformer-CNN—that ingests time-series data from voltage, current, power factor, ambient temperature, and manufacturer-rated efficiency specs. This hybrid model combines the global contextual awareness of Transformer networks with the multi-scale temporal feature extraction capabilities of convolutional layers, enabling highly accurate, real-time estimation of the AC/DC conversion efficiency for each charging session.
With this efficiency estimate in hand, the system then feeds into a refined recursive damped least squares (RDLS) algorithm to solve for individual charger metering errors. Unlike standard recursive least squares—which can become unstable or diverge when faced with noisy or ill-conditioned data—the RDLS variant introduces a damping term that penalizes abrupt parameter shifts between iterations. This not only stabilizes convergence but also enhances robustness against measurement noise and transient anomalies.
In extensive testing using both synthetic datasets and real-world operational data from 30 charging stations over a 14-month period, the method demonstrated exceptional performance. On simulated data with known faulty meters (intentionally biased by up to ±4.2%), the system achieved a 98% detection accuracy and an F1-score of 84%—significantly outperforming baseline RLS, dual-parameter RLS, and kernel ridge regression alternatives. Even more impressively, on actual field data where ground-truth faults were later confirmed via physical inspection, the algorithm correctly identified all malfunctioning units with 100% precision and recall.
This level of reliability has profound implications. For utility operators managing thousands of chargers, it means moving from reactive, calendar-based maintenance to predictive, condition-based oversight. For EV drivers, it ensures fair billing and consistent charging performance. And for regulators, it offers a scalable, non-intrusive tool to enforce metrological standards across a fragmented and rapidly evolving infrastructure landscape.
The approach is particularly timely given the massive public and private investments flowing into EV charging networks. In the United States alone, the National Electric Vehicle Infrastructure (NEVI) program has allocated $5 billion to build a coast-to-coast fast-charging corridor. Similar initiatives are underway in the European Union, China, and India. Yet without reliable mechanisms to verify the accuracy and integrity of these chargers, consumer trust—and the broader adoption of electric vehicles—could be undermined.
Critically, the proposed method requires no hardware retrofits. It operates entirely on existing telemetry data already collected by most modern charging stations and transmitted to central management systems. This “software-only” upgrade path dramatically lowers the barrier to deployment, especially for legacy installations.
The research also underscores a broader trend in smart infrastructure: the convergence of classical engineering principles with artificial intelligence. Rather than treating AI as a black box, the team anchored their solution in first-principles physics—energy conservation—then used machine learning to handle the complex, nonlinear variables that defy analytical modeling. This hybrid paradigm aligns with emerging best practices in trustworthy AI for critical infrastructure, where interpretability, safety, and compliance cannot be sacrificed for marginal gains in prediction accuracy.
From a technical standpoint, the Transformer-CNN architecture represents a thoughtful adaptation of cutting-edge AI to a domain-specific challenge. Transformers, originally developed for natural language processing, excel at capturing long-range dependencies in sequences—ideal for understanding how a charger’s efficiency evolves over a multi-hour session. Meanwhile, the multi-scale depthwise convolutions efficiently extract local patterns (e.g., spikes in current during battery preconditioning) without the computational overhead of dense layers. The result is a lightweight model suitable for edge deployment or cloud-based analytics, depending on the operator’s architecture.
The recursive damped least squares component further exemplifies engineering pragmatism. By incorporating a forgetting factor, the algorithm can adapt to gradual changes in station behavior—such as increasing cable resistance due to corrosion—while the damping term prevents overreaction to short-term noise. This balance between adaptability and stability is essential for real-world systems where data quality is imperfect and operational conditions are constantly shifting.
Looking ahead, the methodology could be extended beyond metering accuracy. The same energy-balance framework might detect energy theft, identify failing power electronics, or even estimate battery state-of-health by analyzing charging inefficiencies at the vehicle end. Moreover, as vehicle-to-grid (V2G) services emerge, accurate bidirectional energy accounting will become even more crucial—making robust, online verification not just desirable but mandatory.
For now, the immediate impact lies in operational efficiency. Manual calibration of a single DC fast charger can cost hundreds of dollars and require hours of technician time. Scaling that across a national network is economically and logistically untenable. By enabling remote, continuous self-diagnosis, this new method could reduce maintenance costs by orders of magnitude while simultaneously improving service quality.
The work also sets a high bar for academic-industry collaboration. Led by seasoned engineers from State Grid Jibei—a major regional utility—and supported by academic experts in AI and power systems from Tianjin University, the project bridges theoretical innovation and practical deployment. The inclusion of real-world data spanning diverse charger models, environmental conditions, and usage patterns adds significant credibility to the results.
As EV adoption continues its exponential growth—global sales surpassed 10 million units in 2022 and are projected to exceed 40 million annually by 2030—the pressure on charging infrastructure will only intensify. Solutions like this one, which enhance reliability without adding complexity or cost, will be essential to maintaining the momentum of the electric transition.
In an era where charging speed and network coverage dominate headlines, this research reminds us that accuracy and trust are equally vital. A charger may be fast and widely available, but if it bills you for 50 kWh when you only received 45, its value evaporates. By ensuring that every kilowatt-hour is accounted for—fairly, transparently, and automatically—this new method helps build the foundation for a truly sustainable and user-centric electric mobility ecosystem.
Authors: Dongxiang Jiao¹, Yachao Wang¹, Di Han¹, Xuechao Li², Shengli Yuan², Zhaoshuai Dang³, Ting Yang³
Affiliations:
¹ State Grid Jibei Electric Power Co., Ltd. Metrology Center, Beijing 100032, China
² Tangshan Power Supply Company, State Grid Jibei Electric Power Co., Ltd., Tangshan 063099, Hebei Province, China
³ School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China
Published in: Electric Power Construction, Vol. 45, No. 8, August 2024
DOI: 10.12204/j.issn.1000-7229.2024.08.013