ISSA-KELM Model Enhances EV Battery SOC Accuracy
In the rapidly evolving landscape of electric mobility, one of the most critical challenges remains the precise estimation of a battery’s state of charge (SOC). As electric vehicles (EVs) continue to gain market share and become central to global transportation strategies, the demand for smarter, more reliable battery management systems has never been higher. A recent breakthrough in this domain comes from a research team at State Grid Jingmen Power Supply Company, who have developed a novel method that significantly improves the accuracy of SOC estimation using an advanced machine learning framework.
The study, published in the August 2024 issue of Northeast Electric Power Technology, introduces a new model known as ISSA-KELM—short for Improved Sparrow Search Algorithm-optimized Kernel Extreme Learning Machine. This approach marks a pivotal advancement in the field of battery diagnostics, offering enhanced precision, faster convergence, and superior stability compared to existing methods. The work was led by Chen Baihan, along with co-authors Yang Wei, Wan Wenxin, and Liu Chuang, whose interdisciplinary expertise in power systems and intelligent algorithms has culminated in a solution with far-reaching implications for EV manufacturers, grid operators, and energy researchers alike.
At the heart of modern electric vehicles lies the lithium iron phosphate (LFP) battery—a technology praised for its safety, longevity, and thermal stability. However, accurately determining how much energy remains in such batteries under real-world conditions is a complex task. Traditional methods like ampere-hour integration and open-circuit voltage measurement often fall short due to their sensitivity to noise, temperature fluctuations, and aging effects. These limitations can lead to inaccurate range predictions, inefficient charging cycles, and even safety risks if the battery is over-discharged or overcharged.
To overcome these shortcomings, researchers have increasingly turned to artificial intelligence (AI)-driven models capable of learning from vast datasets and adapting to dynamic operating conditions. Among these, extreme learning machines (ELMs) have emerged as promising tools due to their fast training speed and strong generalization capabilities. However, standard ELMs are not without drawbacks, particularly when it comes to parameter selection and kernel function optimization. This is where the innovation of the ISSA-KELM model becomes evident.
The core idea behind the ISSA-KELM framework is to enhance the performance of the Kernel Extreme Learning Machine (KELM) by optimizing its key parameters—namely the kernel coefficient and the penalty parameter—using a modified version of the Sparrow Search Algorithm (SSA). The original SSA is a bio-inspired optimization technique that mimics the foraging and anti-predation behaviors of sparrows. While effective in many applications, the conventional SSA can sometimes get trapped in local optima, especially when dealing with high-dimensional, non-linear problems like SOC estimation.
To address this limitation, the research team implemented three strategic improvements to the base algorithm, transforming it into what they call the Improved Sparrow Search Algorithm (ISSA). First, they introduced Logistic chaos initialization, which ensures a more uniform distribution of initial search agents across the solution space. Unlike random initialization, which may leave certain regions underexplored, chaotic mapping promotes diversity and enhances global search capability from the outset.
Second, the team incorporated a non-linearly decreasing inertia weight into the position update mechanism of the searchers. This adaptive weighting allows the algorithm to balance exploration and exploitation more effectively over successive iterations. In the early stages, larger weights enable broader exploration of the search space, while smaller weights in later stages facilitate fine-tuning around promising solutions. This dynamic adjustment prevents premature convergence and supports a more robust path toward the global optimum.
Third, and perhaps most crucially, the researchers integrated Levy flight dynamics into the movement of the sentinels—those individuals responsible for detecting threats and guiding the flock to safer areas. Levy flight, characterized by a series of short steps interspersed with occasional long jumps, is a natural search pattern observed in various animals and has been shown to optimize search efficiency in unknown environments. By embedding this behavior into the ISSA, the algorithm gains a greater ability to escape local minima and explore distant regions of the solution landscape, thereby increasing the likelihood of discovering the best possible parameter configuration for the KELM.
With the ISSA fully refined, the next step was to apply it to optimize the KELM model for SOC estimation. The input features selected for the model were voltage, current, and temperature—three readily available signals that reflect the instantaneous operating condition of the battery. The output, naturally, was the SOC value, expressed as a percentage of the battery’s total capacity. Before feeding the data into the model, all inputs were normalized to ensure consistent scaling and prevent any single variable from dominating the learning process.
The experimental dataset consisted of 400 samples collected from a single LFP cell using a Neware battery cycler. The cell had a nominal capacity of 1.75 Ah and a rated voltage of 4.2 V, representative of typical small-format cells used in modular battery packs. Of the total samples, 380 were used for training the model, while the remaining 20 formed the test set, allowing for an unbiased evaluation of predictive performance.
During the training phase, the ISSA iteratively adjusted the KELM’s parameters to minimize the root mean square error (RMSE) between predicted and actual SOC values. The convergence curve revealed that the ISSA-KELM model reached a stable solution after just 57 iterations, achieving a training RMSE of 0.0113. In contrast, the same model optimized with the original SSA required 109 iterations to converge, with a final RMSE of 0.0155. This stark difference highlights the efficiency gains offered by the improved algorithm, not only in terms of speed but also in the quality of the solution found.
Once trained, the ISSA-KELM model was tested against two benchmark approaches: the SSA-KELM and the PSO-LSSVM (Particle Swarm Optimization-optimized Least Squares Support Vector Machine), the latter being a widely used method in battery diagnostics. The comparison focused on three key metrics: maximum relative error (MAX), mean absolute percentage error (MAPE), and RMSE. These indicators provide a comprehensive view of both the worst-case deviation and the overall accuracy and consistency of the predictions.
The results were compelling. The ISSA-KELM model achieved a maximum relative error of 6.232%, significantly lower than the 8.643% recorded by SSA-KELM and the 13.150% seen in PSO-LSSVM. This means that even in the most challenging scenarios—such as sudden load changes or extreme temperatures—the ISSA-KELM predictions remained within a narrow margin of the true SOC, reducing the risk of misjudgment.
On average, the ISSA-KELM model demonstrated a MAPE of 3.964%, outperforming both comparison models. This level of precision translates into more reliable range estimation for drivers, enabling better trip planning and reducing range anxiety—a major psychological barrier to EV adoption. Moreover, the model’s RMSE of 0.0197 indicated excellent overall stability, suggesting that its predictions are not only accurate but also consistently so across diverse operating conditions.
Beyond numerical metrics, the visual analysis of prediction curves further confirmed the superiority of the ISSA-KELM approach. When plotted alongside actual SOC values, the ISSA-KELM outputs closely followed the reference line, with minimal oscillation or lag. In contrast, the other models exhibited noticeable deviations, particularly during transitions between charging and discharging phases, where rapid changes in current and voltage can confuse less sophisticated algorithms.
One of the most significant advantages of the ISSA-KELM model is its practical applicability. Unlike deep neural networks that require extensive computational resources and long training times, the KELM architecture is inherently lightweight and efficient. Combined with the fast convergence of ISSA, this makes the model well-suited for real-time implementation in onboard battery management systems (BMS). Given the limited processing power and memory available in automotive embedded systems, such efficiency is not just desirable—it is essential.
Moreover, the model’s reliance on only three input variables—voltage, current, and temperature—means it can be deployed without requiring additional sensors or complex signal processing. These parameters are already monitored by virtually every EV’s BMS, making integration straightforward and cost-effective. There is no need for open-circuit voltage measurements, which require the battery to rest for extended periods, nor does it depend on detailed electrochemical models that are difficult to calibrate and maintain.
The implications of this research extend beyond individual vehicle performance. Accurate SOC estimation plays a vital role in fleet management, smart charging, and vehicle-to-grid (V2G) operations. For example, in a V2G scenario, utilities rely on precise knowledge of each vehicle’s available energy to schedule discharging events and stabilize the grid. An error of just a few percentage points in SOC estimation could lead to over-dispatching, potentially damaging batteries or failing to meet power demands. The ISSA-KELM model, with its high accuracy and reliability, provides a solid foundation for such advanced applications.
From a sustainability perspective, improved SOC estimation contributes to longer battery life by preventing deep discharges and enabling optimal charging strategies. It also supports better utilization of renewable energy sources by allowing EVs to charge when solar or wind generation is abundant, thus reducing reliance on fossil-fueled power plants. In this way, the technological advancement described in this study aligns with broader environmental and energy policy goals.
It is worth noting that while the current study focused on LFP batteries, the underlying methodology is not limited to this chemistry. With appropriate adjustments to the training data and parameter ranges, the ISSA-KELM framework could be adapted to other types of lithium-ion batteries, such as nickel manganese cobalt (NMC) or lithium nickel cobalt aluminum oxide (NCA), which are commonly used in high-performance EVs. Future work may explore such extensions, as well as the integration of additional factors like battery age and internal resistance to further refine the model.
In conclusion, the development of the ISSA-KELM model represents a significant leap forward in the science of battery state estimation. By combining the strengths of bio-inspired optimization and kernel-based machine learning, the research team has created a tool that is not only more accurate and stable than existing methods but also practical for real-world deployment. As the world moves toward a fully electrified transportation future, innovations like this will play a crucial role in ensuring that EVs are not only cleaner and more efficient but also smarter and more trustworthy.
Chen Baihan, Yang Wei, Wan Wenxin, Liu Chuang, State Grid Jingmen Power Supply Company, Northeast Electric Power Technology, DOI: 10.12061/j.issn.1004-7913.2024.08.001