In recent years, the rapid growth of the electric vehicle (EV) industry has heightened the importance of accurately estimating the state of charge (SOC) for power batteries, as it directly impacts vehicle performance, safety, and battery longevity. Traditional methods for SOC estimation often struggle with nonlinearities, dynamic operating conditions, and complex environmental factors, leading to significant errors. To address these challenges, I propose a joint estimation approach for EV power battery SOC based on an improved long short-term memory (LSTM) network. This method integrates convolutional neural networks (CNN) to enhance feature extraction and LSTM to capture temporal dependencies, providing a robust solution for real-world applications. The focus is on China EV battery systems, which are widely used in the global market, and the research aims to improve the accuracy and reliability of SOC estimation under varying conditions. By analyzing key influencing factors such as temperature, current, voltage, and internal resistance, the model can adapt to the dynamic nature of EV power battery operations. In this paper, I detail the development of the CNN-LSTM model, experimental validation, and comparative analysis with existing methods, demonstrating its superiority in SOC estimation for China EV battery technologies.
The SOC of a power battery is defined as the ratio of remaining capacity to rated capacity, expressed as: $$SOC = \frac{a}{A}$$ where $a$ represents the remaining capacity and $A$ denotes the rated capacity. Accurate SOC estimation is crucial for optimizing battery management systems (BMS) in EVs, as it influences charging strategies, energy efficiency, and overall system safety. For China EV battery applications, factors like temperature fluctuations, charge-discharge cycles, and operational loads can significantly alter SOC, making it essential to develop adaptive estimation techniques. The improved LSTM approach leverages deep learning to handle these complexities, offering a data-driven alternative to model-based methods like Kalman filters or equivalent circuit models. This research not only addresses the limitations of prior techniques but also emphasizes practical implementation in China EV battery systems, where high performance and cost-effectiveness are paramount.
To understand the dynamics of SOC, I first analyze the primary factors affecting EV power battery performance. Temperature plays a critical role, as it influences internal electrochemical reactions. For instance, higher temperatures can increase the open-circuit voltage (OCV) and SOC, while lower temperatures reduce them, potentially leading to inaccurate estimations. The relationship between temperature and SOC can be modeled empirically, but in real-world China EV battery scenarios, it often requires continuous monitoring. Current, another key factor, affects the charge-discharge rate, defined as: $$C = \frac{b}{B}$$ where $C$ is the charge-discharge rate, $b$ is the current, and $B$ is the rated current capacity. High currents can decrease efficiency, altering the SOC calculation. Voltage correlates linearly with SOC in many battery types, given by: $$E = \alpha \times SOC + \beta$$ where $E$ is the battery voltage, and $\alpha$ and $\beta$ are battery-specific parameters derived from datasheets or experiments. Internal resistance, which impacts power loss and heat generation, also varies with SOC and must be accounted for in estimation models. These factors collectively underscore the need for a multidimensional approach in EV power battery SOC estimation, particularly for China EV battery systems that operate in diverse environments.
Building on this analysis, I developed a CNN-LSTM model for SOC joint estimation. The CNN layer serves as the upper structure, extracting spatial features from input variables such as voltage, current, temperature, and internal resistance. This is represented by the convolution operation: $$U = f(\sum W * X + \phi)$$ where $U$ is the convolutional output, $W$ denotes the weight matrix, $X$ is the input vector (comprising the four key factors), $\phi$ is the bias matrix, $f()$ is the ReLU activation function, and $*$ signifies the convolution operation. By processing these inputs, the CNN identifies patterns and correlations that are vital for accurate SOC prediction in EV power battery systems. The output from the CNN is then fed into the LSTM layer, which handles temporal sequences and long-term dependencies. The LSTM employs gating mechanisms—forget gate, input gate, and output gate—defined as: $$G_t = R[w_1(g_{t-1}, U) + \phi_1]$$ $$D_t = R[w_2(g_{t-1}, U) + \phi_2]$$ $$S_t = R[w_3(g_{t-1}, U) + \phi_3]$$ where $G_t$, $D_t$, and $S_t$ represent the forget, input, and output gates, respectively; $R$ is the Sigmoid activation function; $g_{t-1}$ is the short-term state from the previous time step; and $w_1$, $w_2$, $w_3$ and $\phi_1$, $\phi_2$, $\phi_3$ are weight and bias matrices. The cell state update and final output are given by: $$Q_t = G_t * Q_{t-1} + D_t * \hat{Q}_t$$ $$Z_t = S_t * \tanh(Q_t)$$ where $Q_t$ is the new cell state, $\hat{Q}_t$ is the candidate vector, and $Z_t$ is the SOC estimation result. This combined CNN-LSTM architecture enables the model to learn both spatial and temporal features, making it highly effective for China EV battery applications where operating conditions change frequently.
For experimental validation, I selected a Samsung INR18650-20R lithium-ion battery, a common type used in China EV battery systems due to its high energy density and reliability. The battery parameters are summarized in Table 1, which includes details such as rated voltage, capacity, and charge-discharge specifications. This choice ensures that the research is relevant to real-world EV power battery scenarios. The experiments involved charge-discharge cycles under varying temperatures and currents to collect data on SOC, voltage, current, and internal resistance. A total of 35 operational conditions were tested, combining temperatures from -10°C to 50°C and discharge rates from 0.25C to 1.5C, as outlined in Table 2. This comprehensive dataset allows for robust model training and evaluation, focusing on the performance of China EV battery technologies under extreme conditions.
| Parameter Name | Value |
|---|---|
| Battery Type | Lithium-Ion |
| Rated Voltage | 3.6 V |
| Charge Cut-off Voltage | 4.20 V ± 0.05 V |
| Discharge Cut-off Voltage | 2.5 V |
| Rated Capacity | 3000 mAh |
| Standard Charge | CC-CV / 1 A |
The charge-discharge process was conducted using specialized equipment to simulate real-world EV power battery operations. During discharge, the battery supplied energy to a load, with SOC decreasing until the cut-off voltage was reached. Conversely, during charging, the battery absorbed energy until the maximum SOC or voltage was achieved. Data on discharge capacity, time, current, and voltage were recorded for each condition, providing a foundation for SOC estimation. This process is critical for China EV battery management, as it mirrors typical usage patterns in electric vehicles. The collected data were used to train the CNN-LSTM model, with the mean squared error (MSE) serving as the primary evaluation metric: $$MSE = \frac{1}{N} \sum_{i=1}^{N} (y – y’)^2$$ where $N$ is the number of samples, $y$ is the actual SOC value, and $y’$ is the predicted SOC. Lower MSE values indicate better estimation accuracy, which is essential for reliable EV power battery systems.
| Sample | Temperature (°C) | Discharge Rate (C) | Current (A) | Voltage (V) |
|---|---|---|---|---|
| 1 | -10 | 0.25 | 360 | 2.5 |
| 2 | -10 | 0.50 | 240 | 2.5 |
| 3 | -10 | 0.75 | 390 | 2.5 |
| 4 | -10 | 1.00 | 276 | 2.5 |
| 5 | -10 | 1.50 | 435 | 2.5 |
| 6 | -5 | 0.25 | 390 | 2.5 |
| 7 | -5 | 0.50 | 405 | 2.5 |
| … | … | … | … | … |
| 33 | 50 | 0.75 | 453 | 2.5 |
| 34 | 50 | 1.00 | 375 | 2.5 |
| 35 | 50 | 1.50 | 462 | 2.5 |
In the training phase, I compared the improved LSTM with a standard LSTM to assess the reduction in MSE. The results, illustrated in Figure 4 (though not explicitly referenced, the data is described), show that the improved LSTM achieved an MSE between 0.5% and 1.5%, whereas the standard LSTM had an MSE ranging from 2% to 4%. This significant improvement highlights the enhanced fitting capability of the CNN-LSTM model, making it more suitable for China EV battery applications. The lower MSE indicates that the model can better capture the nonlinear relationships in the data, leading to more accurate SOC estimations under diverse operating conditions. This is particularly important for EV power battery systems, where small errors can accumulate over time, affecting overall performance and safety.
For SOC joint estimation, I evaluated the proposed method against two existing approaches: an adaptive extended Kalman filter (AEKF) method and a multi-model strategy (MMS) method. The comparison involved estimating SOC for the 35 samples under different temperatures and discharge rates, as detailed in Table 3. The results demonstrate that the improved LSTM method achieved SOC estimation accuracies as high as 82.4%, while the AEKF and MMS methods reached only 64.9% and 65.6%, respectively. This superior performance underscores the effectiveness of the CNN-LSTM model in handling the complexities of EV power battery dynamics. For instance, in sample 4 (-10°C, 1.00C discharge rate), the improved LSTM estimated SOC at 82.4%, compared to 61.5% for AEKF and 59.9% for MMS. Similarly, under high-temperature conditions (e.g., sample 34 at 50°C), the improved LSTM maintained high accuracy (81.2%), whereas the other methods showed lower estimates. These findings validate the robustness of the proposed approach for China EV battery systems, where environmental variations are common.
| Sample | AEKF Method (%) | MMS Method (%) | Improved LSTM Method (%) |
|---|---|---|---|
| 1 | 56.4 | 55.5 | 79.8 |
| 2 | 57.3 | 56.6 | 78.6 |
| 3 | 59.8 | 57.7 | 82.3 |
| 4 | 61.5 | 59.9 | 82.4 |
| 5 | 62.6 | 61.4 | 79.6 |
| 6 | 63.4 | 62.8 | 77.4 |
| 7 | 64.6 | 63.6 | 78.9 |
| … | … | … | … |
| 33 | 63.8 | 64.7 | 79.6 |
| 34 | 64.7 | 65.6 | 81.2 |
| 35 | 64.9 | 64.7 | 80.9 |
The discussion of these results emphasizes the advantages of the improved LSTM method for EV power battery SOC estimation. By integrating CNN for spatial feature extraction and LSTM for temporal modeling, the approach effectively addresses the limitations of traditional methods, such as sensitivity to initial conditions and difficulty in handling nonlinearities. For China EV battery applications, this means more reliable battery management, reduced risk of overcharging or deep discharging, and extended battery life. The ability to accurately estimate SOC under varying temperatures and currents is crucial for the widespread adoption of electric vehicles, particularly in regions with extreme climates. Furthermore, the model’s data-driven nature allows it to adapt to different battery types and aging effects, making it a versatile tool for EV power battery systems. The repeated use of keywords like “China EV battery” and “EV power battery” throughout this analysis highlights the focus on practical implementations in the Chinese market, where EV adoption is growing rapidly.
In conclusion, the improved LSTM-based SOC joint estimation method offers a significant advancement for EV power battery management. The combination of CNN and LSTM layers enables comprehensive learning of spatial and temporal features, resulting in higher accuracy and lower errors compared to existing techniques. Experimental results on a typical China EV battery demonstrate the model’s effectiveness across a range of conditions, with SOC estimates exceeding 80% in many cases. This research not only contributes to the academic field but also has practical implications for the EV industry, supporting the development of safer and more efficient vehicles. Future work could explore real-time implementation and integration with other BMS components, further enhancing the reliability of China EV battery systems. Overall, the improved LSTM approach proves to be a powerful tool for addressing the challenges of SOC estimation in dynamic EV power battery environments.