Accurate Battery Power Estimation Breakthrough for EVs
In the rapidly evolving landscape of electric vehicles (EVs), where performance, safety, and longevity of the powertrain are paramount, a critical challenge has long persisted: accurately estimating the real-time power capability of lithium-ion battery systems. This capability, known as State of Power (SOP), dictates how much energy can be drawn for acceleration or absorbed during regenerative braking at any given moment. Underestimating it limits vehicle performance; overestimating it risks damaging the battery or triggering safety shutdowns. Now, a significant advancement in this field has emerged from a collaborative research effort, introducing a novel algorithm that promises unprecedented accuracy and robustness across a battery’s entire operational spectrum.
The research, spearheaded by Li Qiang from Weichai Power Co., Ltd., in partnership with a team from the School of Mechanical Engineering at Beijing Institute of Technology—including Zhang Kaixuan, Yuan Wenwen, Xu Yahan, Yang Ruixin, and Fang Yu—has yielded a powerful solution. Their work, published in the prestigious Transactions of China Electrotechnical Society, details a sophisticated method designed to overcome the inherent complexities of battery behavior. Batteries are not static components; their performance degrades over time, fluctuates with temperature, and responds non-linearly to electrical loads. These dynamic, time-varying characteristics have made precise, real-time SOP estimation a persistent technical bottleneck for Battery Management Systems (BMS), the electronic brain that governs every aspect of a battery pack’s operation.
Traditional approaches to SOP estimation have often fallen short in real-world applications. The characteristic map method relies on pre-programmed look-up tables derived from offline testing. While computationally simple, this method fails to adapt to a battery’s changing condition as it ages or operates under untested conditions, leading to inaccurate predictions. Data-driven models, which treat the battery as a “black box” and learn from vast datasets, are highly dependent on the quality and quantity of training data. Crucially, obtaining a reliable, continuous stream of reference SOP data for training is extremely difficult, making these models less practical for online, real-time use. The third category, multi-constraint estimation, uses a dynamic battery model to calculate the maximum allowable current based on limits like voltage, state of charge (SOC), and temperature, then multiplies this by the terminal voltage to get power. The accuracy of this method is heavily dependent on the fidelity of the underlying battery model and the precision with which its internal parameters—such as internal resistance and capacitance—can be identified in real time. Previous methods often struggled to track these parameters as they slowly change with aging and temperature, leading to accumulating errors.
The new methodology developed by Li Qiang and his colleagues directly addresses this core challenge of parameter drift. Their solution is built upon a foundation of rigorous model selection and system-level understanding. The team began by evaluating twelve common equivalent circuit models (ECMs), which are mathematical representations of a battery’s electrical behavior. After extensive testing, including the standard Hybrid Pulse Power Characteristic (HPPC) test and dynamic driving cycles like UDDS and DST, they selected the first-order RC (Thevenin) model. This model, which includes an open-circuit voltage, an ohmic resistance, and one RC network to represent polarization effects, was chosen for its optimal balance between high voltage prediction accuracy and low computational demand—essential for implementation in a resource-constrained BMS.
A key innovation lies in the transition from a single cell model to a full battery pack system. Instead of modeling every single cell, which would be computationally prohibitive, the researchers developed a “dual-characteristic cell uniformly distributed” pack model. This model focuses on the two most critical cells in the pack: the one with the highest voltage and the one with the lowest voltage at any given moment. By assuming a uniform distribution of cell states across the pack, the complex system can be effectively simplified. The BMS only needs to monitor the pack’s total voltage and the voltages of these two representative cells. This elegant simplification allows the high-precision single-cell algorithm to be scaled up to manage an entire pack with minimal computational overhead, a crucial advantage for commercial applications.
At the heart of the breakthrough is the novel “multi-time scale sliding window Double Extended Kalman Filter” (DEKF) algorithm. The Kalman filter is a powerful mathematical tool used for estimating the state of a dynamic system from noisy measurements. The “Double” Extended Kalman Filter (DEKF) uses two interconnected filters: one to estimate the battery’s instantaneous states (like SOC and polarization voltage) and another to estimate the slowly changing model parameters (like internal resistances and capacitance). The “multi-time scale” aspect is the critical refinement. It recognizes that the battery’s state (e.g., SOC) can change rapidly with load, while its fundamental parameters (e.g., internal resistance) change very slowly over hours, days, or weeks due to aging and temperature shifts. The algorithm operates on two different time scales: a “narrow” scale for rapid state updates and a “wide” scale for infrequent parameter updates. This prevents the fast-changing state noise from corrupting the estimation of the slow-changing parameters, a common problem in standard DEKF implementations that leads to instability and error.
The “sliding window” component further enhances the algorithm’s robustness and accuracy. Instead of relying on a single data point, the parameter estimation filter analyzes a window of recent data—300 consecutive data points in their implementation. This averaging effect smooths out transient noise and measurement errors, leading to a more stable and reliable parameter estimate. The algorithm is also designed to be self-updating. By incorporating the results of periodic peak power tests—standardized procedures that measure the battery’s actual power capability at different states—the model’s parameter library is continuously refined. This creates a feedback loop where the algorithm learns from real-world performance, enabling a “slow time-varying” estimation of parameters that accurately tracks the battery’s aging and changing characteristics over its entire lifespan.
To calculate the SOP itself, the team employed a sophisticated approach centered on the concept of constant power. Rather than making a simple instantaneous calculation, the algorithm predicts whether a proposed constant power level can be sustained for a specific duration (e.g., 10 seconds) without violating voltage limits. This involves a forward simulation of the battery’s voltage response over that time period. Because this relationship is complex and non-linear, the researchers used a “bisection method” (a form of binary search). The algorithm starts with an initial guess for the maximum power, simulates the resulting voltage, compares it to the safe voltage limits, and then iteratively adjusts its power guess—up or down—until it converges on the true maximum sustainable power with high precision. This method ensures that the estimated SOP reflects a practical, usable power level, not just a theoretical instantaneous peak.
The true test of any BMS algorithm is its performance under real-world conditions. To validate their method, the research team went beyond simple simulations and built a comprehensive hardware-in-the-loop (HIL) test bench. This sophisticated setup integrates real physical components—such as a 150Ah battery pack made of three parallel-connected 50Ah ternary lithium-ion cells, high-precision battery cycler hardware from ARBIN, and a programmable thermal chamber—with a real-time simulation environment running on a dedicated MPC5644 main control board. This allows the algorithm to be tested against actual battery behavior under controlled but realistic conditions, providing a far more rigorous validation than pure software simulation.
The experimental campaign was exceptionally thorough. The team conducted extensive testing on both a fresh battery pack and an aged one, subjecting them to a wide range of conditions over a six-month period. This included four different temperatures (-10°C, 0°C, 25°C, and 45°C) and four demanding driving cycles (NEDC, UDDS, US06, and WLTC), simulating everything from city driving to aggressive highway maneuvers. The HIL platform executed these complex test protocols, collecting vast amounts of data on voltage, current, temperature, and the algorithm’s internal estimates.
The results of this rigorous validation were impressive. For State of Charge (SOC) estimation, a critical parameter in its own right, the algorithm demonstrated exceptional accuracy. Across 14 different test scenarios on the fresh battery pack, the maximum SOC estimation error was a remarkably low 2.192%. Even more challenging, on the aged battery pack, where internal changes are more pronounced, the maximum error was just 2.82%. These results are well within the industry’s stringent requirements for BMS accuracy.
The performance on SOP estimation was even more compelling. The primary metric for SOP accuracy is the “equivalent voltage error,” which measures the difference between the predicted voltage at the end of a peak power discharge or charge and the actual measured voltage. A small error indicates the algorithm correctly predicted the power limit. The researchers evaluated this at multiple SOC levels (from 3% to 95%) and across four temperatures. The results, summarized in their data, show that in every single test point where data was available, the equivalent voltage error was less than 40 millivolts (mV). In many cases, it was far lower, often under 20 mV. An error of 40 mV on a typical 3-4 volt cell represents an accuracy of over 98%, which is considered excellent for real-world BMS applications. This level of precision means the BMS can confidently push the battery closer to its true physical limits, maximizing vehicle performance and regenerative braking efficiency without compromising safety.
The implications of this research are significant for the entire EV industry. An accurate SOP estimate allows for more aggressive and efficient energy management strategies. During acceleration, the vehicle can deliver its full power potential. During regenerative braking, it can capture more kinetic energy, extending range. It also enables more precise prediction of vehicle range and performance, improving the driver’s experience. Furthermore, by preventing the battery from being pushed beyond its safe operating limits, the algorithm contributes to longer battery life and enhanced safety, addressing two of the biggest concerns for EV consumers.
The work of Li Qiang and his team represents a major step forward in BMS technology. By combining a carefully selected battery model, a scalable system-level approach, and a highly sophisticated multi-time scale DEKF algorithm with a sliding window, they have created a solution that is not only accurate but also robust and adaptable. It successfully navigates the complex, time-varying nature of real-world battery operation across temperature, aging, and diverse driving conditions. While the researchers note that performance at very low temperatures (below -20°C) remains a challenge due to extreme polarization effects, their current work sets a new benchmark for SOP estimation. This research, published in the Transactions of China Electrotechnical Society (DOI:10.19595/j.cnki.1000-6753.tces.230086), provides a powerful tool for BMS engineers and paves the way for safer, more efficient, and higher-performing electric vehicles. Li Qiang, Zhang Kaixuan, Yuan Wenwen, Xu Yahan, Yang Ruixin, Fang Yu, Transactions of China Electrotechnical Society, DOI:10.19595/j.cnki.1000-6753.tces.230086