Wireless Charging Alignment Breakthrough for Electric Vehicles
As electric vehicles (EVs) continue to gain momentum across global markets, the focus on improving charging infrastructure has intensified. Among the most promising technologies is wireless power transfer, which offers a seamless and user-friendly alternative to traditional plug-in systems. However, one persistent challenge has limited its widespread adoption: precise coil alignment between the vehicle’s receiver and the ground-based transmitter. Misalignment can significantly reduce charging efficiency, making accurate positioning essential. A recent study published in Electrical Measurement & Instrumentation introduces a novel solution using artificial intelligence to solve this problem, paving the way for more reliable and efficient EV wireless charging.
The research, led by Pan Zhixin of Jiangsu Fangtian Electric Power Technology Co., Ltd., in collaboration with experts from State Grid Jiangsu Electric Power Co., Ltd. and Southeast University, presents a breakthrough approach based on Backpropagation (BP) neural networks. This method enables real-time detection of the receiving coil’s position relative to the transmitting coil, offering critical feedback for either driver guidance or autonomous parking systems. The findings were detailed in the February 2024 issue of the journal, under the title “Research on coil positioning technology based on BP neural network,” and have since drawn attention from both academic and industry circles.
At the heart of modern wireless EV charging systems lies magnetic coupling resonance, a principle that allows energy to be transferred efficiently over short air gaps through electromagnetic fields. For optimal performance, the primary (transmitter) and secondary (receiver) coils must be closely aligned—typically within a 10 cm tolerance—to maintain high power transfer efficiency, often above 80%. While some mechanical aids exist, such as visual guides or sensor-assisted parking, these methods are either imprecise or costly. Moreover, manual alignment remains impractical for everyday users, especially in public charging stations where speed and convenience are paramount.
Existing solutions have explored various sensing techniques. Some rely on magnetoresistive sensors placed around the charging pad to detect spatial magnetic field variations. Others use auxiliary systems like RFID tags embedded in the vehicle chassis. However, these approaches often require additional hardware, increase system complexity, and raise manufacturing costs. Furthermore, many fail to provide full two-dimensional positional data necessary for fine-tuning alignment in both x- and y-axes.
The team’s innovation circumvents these limitations by leveraging machine learning to interpret signals from simple detection coils integrated into the charging infrastructure. These small inductive coils capture voltage fluctuations induced by the electromagnetic field generated by the main transmitter. Because the strength and distribution of this field vary depending on the relative position of the receiver coil, the induced voltages serve as indirect indicators of spatial location. The relationship between voltage readings and physical coordinates, however, is highly nonlinear and complex—too intricate for conventional mathematical modeling.
This is where the BP neural network comes into play. Unlike rule-based algorithms, neural networks excel at identifying patterns in large datasets and approximating complex functions without requiring explicit equations. By training the network on a dataset of known positions and corresponding voltage inputs, it learns to predict unknown positions with remarkable accuracy. In essence, the system treats coil localization as a supervised learning problem: given four input voltages from strategically placed detection coils, the network outputs the estimated x and y coordinates of the receiver.
To validate their concept, the researchers conducted both simulation and experimental studies. Using ANSYS Maxwell, a finite element analysis tool widely used in electromagnetic design, they modeled a typical LCC-S compensated wireless charging system—a topology favored for its ability to deliver constant current on the primary side and stable output voltage on the secondary. The transmitter and receiver coils were square-shaped, measuring 550 mm and 450 mm per side respectively, with a standard vertical separation of 20 cm, simulating realistic vehicle ground clearance.
A grid of 1,600 points was defined over a 20 cm × 20 cm area centered on the transmitter coil, with measurements taken every 0.5 cm. From this comprehensive map, 25 representative data points were selected for training the neural network. Of these, 17 were used for actual training, while 4 served for validation and another 4 for final testing. The MATLAB Deep Learning Toolbox was employed, utilizing the Levenberg-Marquardt algorithm—an advanced optimization technique known for fast convergence in small-to-medium-sized networks.
One key decision in neural network design is determining the number of neurons in the hidden layer, which directly impacts model capacity and computational load. Too few neurons may result in underfitting, while too many can lead to overfitting and longer processing times. Guided by empirical rules suggesting a range between √(n + r) + a (where n is input count, r is output count, and a is a small integer), the team tested configurations with 5, 10, and 15 hidden neurons. Surprisingly, the simplest configuration—with just five neurons—yielded the best performance, achieving an average error of only 0.907 cm in simulations. Increasing the neuron count did not improve accuracy, indicating that the problem’s complexity could be captured efficiently without excessive computation.
Encouraged by these results, the team moved to physical experimentation. An actual testbed was constructed featuring the same LCC-S circuit driven by a 390V AC source operating at resonant frequency. Four detection coils were mounted beneath the transmitter, each connected to a signal conditioning circuit comprising diodes for half-wave rectification, RC filters for peak holding, and resistive dividers to scale down voltages suitable for microcontroller analog-to-digital conversion. Data acquisition was handled via a single-chip processor, which fed raw voltage values into the pre-trained BP network running on a host computer.
The experimental protocol mirrored the simulation setup: the same 25 calibration points were measured manually, then used to train the network. Five new test positions were evaluated, including off-center locations such as (-2.5 cm, -7.5 cm) and (7.0 cm, 3.5 cm). The system successfully predicted all positions within centimeter-level precision, though slight deviations were observed compared to simulation results. For instance, at the point (-2.5, -7.5), the network estimated (-4.14, -8.4), resulting in an error of approximately 1.8 cm. Similar discrepancies occurred elsewhere, primarily attributed to measurement noise, minor misplacement during manual positioning, and non-idealities in analog circuitry.
Despite these variances, the overall consistency between simulated and real-world outcomes confirmed the robustness of the method. As shown in comparative error distribution charts, experimental errors remained within acceptable bounds, reinforcing the feasibility of deploying such a system in commercial applications. Notably, the entire process—from voltage sampling to coordinate estimation—occurred rapidly, enabling near real-time feedback ideal for integration with automated parking systems.
What sets this approach apart is its balance of simplicity, cost-effectiveness, and performance. Unlike sensor-heavy alternatives requiring arrays of Hall effect devices or external RFID infrastructure, this solution uses only passive inductive elements already compatible with the existing charging hardware. There is no need for additional transmitters, cameras, or ultrasonic modules. The intelligence resides entirely in software, making upgrades and recalibrations straightforward through firmware updates.
Moreover, the reliance on BP neural networks brings inherent adaptability. Should environmental conditions change—such as temperature shifts affecting coil resistance or nearby metallic objects distorting the magnetic field—the network can be retrained with updated data to maintain accuracy. This contrasts sharply with fixed-parameter models that degrade over time without manual recalibration.
From a practical standpoint, integrating this technology into current EV architectures would require minimal modifications. Vehicle manufacturers could embed the trained neural network within the onboard charging control unit, while charging station operators could implement the detection circuitry during pad installation. Feedback could then be relayed visually through the dashboard display, audibly via alerts, or automatically through vehicle-to-infrastructure communication protocols enabling self-parking maneuvers.
The implications extend beyond consumer convenience. Fleet operators managing electric buses or delivery vans stand to benefit significantly, as consistent and rapid alignment reduces downtime and maximizes operational efficiency. Public charging hubs could see higher throughput, reducing congestion and improving user satisfaction. Additionally, precise alignment ensures maximum energy transfer, minimizing waste heat and prolonging component lifespan—key considerations for long-term sustainability.
Safety is another dimension enhanced by accurate positioning. When coils are misaligned, stray electromagnetic fields may increase, potentially interfering with pacemakers or other sensitive electronics. Regulatory standards, such as those set by the International Electrotechnical Commission (IEC), impose strict limits on electromagnetic emissions. By ensuring optimal coupling, this AI-driven method helps keep emissions within safe thresholds, contributing to broader acceptance of wireless charging in urban environments.
While the current implementation focuses on planar displacement, future work could expand to include vertical distance (z-axis) and angular orientation (pitch and yaw), providing full six-degree-of-freedom tracking. This would be particularly useful for uneven surfaces or sloped driveways. Incorporating dynamic learning capabilities—where the network continuously refines itself based on live data—could further enhance resilience against wear and environmental drift.
Another avenue involves multi-coil transmitter arrays, where multiple overlapping coils allow greater spatial flexibility. In such setups, knowing the exact receiver position enables selective activation of the nearest sub-coil, boosting efficiency and reducing standby losses. The BP neural network framework described here provides a solid foundation for such intelligent switching strategies.
Industry response to the study has been positive. Experts note that while wireless charging has long been viewed as a premium feature, advancements like this bring it closer to mainstream viability. Standardization bodies, including SAE International and the Wireless Power Consortium, are actively developing interoperability guidelines, and incorporating AI-based positioning could become part of future specifications.
In conclusion, the research conducted by Pan Zhixin, Yang Xiaomei, Wang Chengliang, Fei Yijun, Xu Qingqiang, and Li Chengyun represents a significant leap forward in EV wireless charging technology. By combining fundamental electromagnetic principles with modern machine learning, they have created a scalable, accurate, and economically viable solution to one of the field’s most enduring challenges. Their work demonstrates how interdisciplinary collaboration—spanning power engineering, signal processing, and artificial intelligence—can yield practical innovations with wide-reaching impact.
As cities worldwide accelerate their transition to electrified transportation, technologies that simplify and optimize the charging experience will be crucial. This BP neural network-based positioning system not only enhances technical performance but also improves user trust and engagement, removing a major psychological barrier to adoption. With continued refinement and deployment, it may soon become a standard feature in next-generation wireless charging systems.
Research on coil positioning technology based on BP neural network – Pan Zhixin, Yang Xiaomei, Wang Chengliang, Fei Yijun, Xu Qingqiang, Li Chengyun; Electrical Measurement & Instrumentation; DOI: 10.19753/j.issn1001-1390.2024.02.029