New Method Boosts Accuracy in Estimating Battery Capacity for EVs with LFP Batteries
As the global push toward electrification and carbon neutrality accelerates, electric vehicles (EVs) have emerged as a cornerstone of sustainable transportation. At the heart of every EV lies the lithium-ion battery, a critical component whose performance directly influences driving range, safety, and overall vehicle efficiency. Among the various types of lithium-ion batteries, lithium iron phosphate (LFP) batteries have gained increasing popularity due to their superior thermal stability, longer cycle life, and lower cost compared to nickel-manganese-cobalt (NCM) variants. However, despite these advantages, accurately estimating the state of health (SOH) of LFP batteries—particularly their remaining capacity—remains a significant challenge in real-world applications.
A recent study published in Energy Storage Science and Technology presents a novel approach to address this challenge. Led by Chen Xingguang from the University of Shanghai for Science and Technology, in collaboration with researchers from Tsinghua University, the team has developed a highly accurate method for identifying the capacity of LFP batteries under actual vehicle operating conditions. Their work introduces a hybrid framework that combines the ampere-hour integration technique with an equivalent circuit model (ECM), enhanced by particle swarm optimization (PSO) algorithms. More importantly, the researchers propose a specific optimization strategy tailored to the unique voltage characteristics of LFP batteries during slow charging, significantly improving estimation accuracy and robustness.
The importance of precise capacity estimation cannot be overstated. Battery capacity is a primary indicator of SOH, which reflects how much energy a battery can store relative to its original design. In practical terms, when an EV battery’s capacity drops below 80% of its nominal value, it is generally considered to have reached the end of its useful life, resulting in reduced driving range and diminished performance. For fleet operators, service providers, and individual owners alike, having reliable, real-time insights into battery health enables better maintenance planning, enhances safety, and supports residual value assessments.
However, unlike voltage or current, battery capacity cannot be directly measured by onboard sensors. It must be estimated using indirect methods, each with its own limitations. Traditional approaches include direct measurement, model-based techniques, and data-driven machine learning models. Direct measurement, while highly accurate, requires full charge-discharge cycles under controlled laboratory conditions—an impractical solution for vehicles in daily use. Data-driven models, such as deep neural networks, offer promising results but depend heavily on large volumes of labeled training data, which are scarce in real-world EV fleets due to the absence of ground-truth capacity records.
Model-based methods, particularly those employing ECMs, strike a balance between accuracy and computational feasibility, making them more suitable for integration into battery management systems (BMS). These models simulate the dynamic electrical behavior of batteries using combinations of resistors, capacitors, and voltage sources. While effective for tasks like state of charge (SOC) and state of power (SOP) estimation, standard ECMs do not inherently include battery capacity as a parameter, necessitating innovative adaptations to support SOH monitoring.
Chen Xingguang and his team recognized this gap and devised a method that embeds capacity as a variable within the ECM framework. By integrating the ampere-hour integration model—which calculates SOC based on current over time—they established a direct link between the ECM’s output and the battery’s total capacity. This linkage allows the system to treat capacity as an unknown parameter to be optimized, rather than assumed or pre-set.
To solve for this parameter, the researchers employed the particle swarm optimization algorithm, a bio-inspired computational technique that mimics the collective behavior of bird flocks searching for food. PSO iteratively adjusts a population of candidate solutions—referred to as particles—by balancing individual and group performance until an optimal solution is found. In this context, the algorithm seeks to minimize the difference between the modeled terminal voltage (predicted by the ECM) and the actual voltage recorded by the vehicle’s BMS.
While this approach works well for many battery chemistries, LFP batteries present a unique challenge due to their flat voltage plateau across most of the SOC range. During charging, the voltage remains nearly constant between approximately 3.2 V and 3.3 V, making it difficult to distinguish subtle changes in SOC based on voltage alone. This phenomenon, rooted in the thermodynamic stability of the lithium iron phosphate crystal structure, reduces the sensitivity of voltage-based estimation methods.
Moreover, during the final stage of slow charging, LFP cells exhibit a rapid voltage rise as they approach full charge. This sharp increase occurs because the phase transition in the cathode material concludes, leading to a steep climb in open-circuit voltage (OCV). Standard ECMs, which rely on fixed OCV-SOC lookup tables derived from lab tests, often fail to capture this abrupt change accurately, especially when polarization effects are minimal under low-current charging conditions. As a result, the model’s predicted voltage lags behind the actual measured voltage, leading to high root mean square error (RMSE) values and unreliable parameter identification.
To overcome these issues, the research team introduced a two-pronged optimization strategy specifically designed for LFP batteries under slow-charging scenarios. The first component involves segmenting the charging data and excluding the terminal voltage spike from the optimization process. Instead of forcing the model to fit the entire voltage curve, the algorithm focuses on the main charging phase up to the onset of the final voltage surge. This truncation prevents the outlier-like behavior at the end of charge from skewing the overall error metric and allows the PSO algorithm to converge more reliably on the true capacity value.
The second component introduces a dual-dimensional loss function that evaluates model accuracy not only in the voltage domain but also in the charge (or capacity) domain. Traditional optimization relies solely on minimizing voltage RMSE, which may overlook discrepancies in the timing or shape of voltage plateaus. By incorporating a second dimension—measured in ampere-hours—the new objective function penalizes deviations in both voltage level and the amount of charge delivered at each voltage point.
This dual-axis evaluation is particularly effective at capturing the subtle inflection points between the two voltage plateaus characteristic of LFP batteries. Even small shifts in these transitions can indicate changes in battery aging, and the enhanced loss function ensures that the model pays closer attention to these critical regions. The weighting between the two dimensions is controlled by a tunable parameter, allowing engineers to adjust the emphasis based on application requirements.
The proposed method was validated using real-world data collected from two different EV models equipped with LFP batteries. The dataset, sourced from a cloud-based vehicle monitoring platform, included timestamped records of cell voltage, current, temperature, and SOC at 10-second intervals. To ensure data quality, the researchers applied rigorous preprocessing steps, including noise filtering, outlier removal, and segmentation of complete charging cycles.
Given the absence of direct capacity labels in field data, the team adopted a two-tier validation strategy. First, they identified a small number of charging events where the vehicle remained idle for extended periods before and after charging—conditions that allow the battery to fully depolarize. Under such static conditions, the terminal voltage closely approximates the OCV, enabling the calculation of accurate SOC values at the start and end of the charge cycle. Using the known change in SOC and the integrated charge input, they computed the actual capacity for these segments, creating a high-confidence reference set labeled as Label-1.
However, due to the rarity of such ideal conditions in everyday driving patterns, only three usable segments were found for one of the two models. To expand the validation dataset, the researchers introduced a second labeling strategy: assuming that vehicles with less than 5,000 kilometers of mileage retain close to their nominal capacity. This assumption, grounded in empirical observations of early-life battery degradation, allowed them to use the manufacturer-specified capacity as a proxy label for low-mileage vehicles, designated as Label-2.
Using these dual validation approaches, the team evaluated the performance of their PSO-ECM method across hundreds of charging cycles. The results demonstrated exceptional accuracy. For the first vehicle model, the mean absolute percentage error (MAPE) was just 2.33%, with individual estimates ranging from a minimal 0.2% error to a maximum of 6.9%. For the second model, the MAPE was slightly higher at 3.38%, with peak errors below 6.1%. These figures represent a significant improvement over conventional methods, particularly considering the inherent noise and variability in real-world driving data.
Beyond numerical accuracy, the study also revealed important insights into the dynamics of battery aging. The estimated capacity values showed a clear downward trend with increasing vehicle mileage, consistent with expected degradation patterns. For example, the first model exhibited a capacity loss of about 15 Ah over 60,000 kilometers, while the second lost approximately 8 Ah over 20,000 kilometers. These trends align with industry benchmarks and validate the method’s ability to track long-term performance decline.
Interestingly, the capacity estimates were not perfectly smooth; some fluctuations and outliers were observed. The researchers attributed these variations to several factors, including sensor inaccuracies, data transmission errors, and the stochastic nature of the PSO algorithm itself. Since PSO relies on random initialization and iterative search, repeated runs on the same data segment can yield slightly different results. While this variability introduces uncertainty, it behaves similarly to Gaussian noise and can be mitigated through statistical smoothing techniques such as moving averages or Kalman filtering in production systems.
One of the most compelling aspects of this research is its practical applicability. Unlike many academic studies that rely on idealized laboratory data, this work was conducted entirely on real-world vehicle telemetry. The fact that the method performs well under messy, uncontrolled conditions speaks to its robustness and readiness for deployment in commercial BMS platforms. Furthermore, the computational demands of the PSO-ECM framework are moderate, making it feasible for implementation in embedded systems with limited processing power.
The authors also noted that while the optimization strategies were specifically designed for slow-charging LFP batteries, the core methodology is broadly applicable. For instance, NCM batteries, which have a more pronounced and nonlinear voltage-SOC relationship, would likely benefit even more from the basic PSO-ECM approach without requiring the specialized enhancements. Similarly, fast-charging scenarios, though more complex due to frequent current interruptions and thermal effects, could still be addressed by adjusting the optimization window or relaxing the voltage fitting constraints.
Nonetheless, the study acknowledges certain limitations. Most notably, the impact of temperature on capacity estimation was not thoroughly investigated. Battery performance is highly temperature-dependent, and variations in ambient or operating temperature can influence both OCV and internal resistance. Future work will need to incorporate thermal effects into the model to ensure consistent accuracy across diverse climates and driving conditions.
Additionally, while the dual-dimensional loss function improves fitting in the mid-SOC range, it may slightly degrade performance at the extremes—particularly near full charge—where the truncated voltage segment resides. This trade-off highlights the importance of application-specific tuning and suggests that hybrid strategies combining multiple optimization objectives may offer further gains.
In conclusion, the research led by Chen Xingguang and his colleagues represents a significant step forward in the field of battery diagnostics for electric vehicles. By addressing the unique challenges posed by LFP chemistry and real-world data constraints, they have developed a method that is not only accurate but also practical and scalable. As automakers continue to adopt LFP batteries for entry-level and fleet vehicles, tools like this will become increasingly vital for ensuring reliability, safety, and customer satisfaction.
The implications extend beyond individual vehicles. Accurate capacity estimation enables smarter grid integration, more efficient second-life battery repurposing, and improved battery swap and leasing models. In an era where data-driven decision-making is transforming industries, this work exemplifies how advanced modeling and optimization can unlock new levels of insight from existing sensor data—without requiring costly hardware upgrades.
With further refinement and integration into next-generation BMS architectures, this approach could become a standard feature in future EVs, providing drivers and operators with trustworthy, real-time health assessments of their most valuable component: the battery.
Chen Xingguang, Shen Yifan, Shao Yuxin, Zheng Yuejiu, Sun Tao, Lai Xin, Shen Kai, Han Xuebing. University of Shanghai for Science and Technology and Tsinghua University. Energy Storage Science and Technology. doi: 10.19799/j.cnki.2095-4239.2024.0144