As a researcher in the field of electric vehicle technology, I have devoted significant effort to the study of battery remaining capacity prediction. The EV battery pack serves as the core energy source for electric vehicles, and its performance directly determines the vehicle’s driving range, acceleration capability, and climbing performance. Therefore, accurate estimation of the remaining capacity of an EV battery pack is of great theoretical and practical importance. In this paper, I present my work on developing a prediction model based on neural networks and genetic algorithms, supported by extensive experimental data and systematic analysis.
The development of electric vehicles has accelerated worldwide due to rising oil prices and increasing environmental concerns. Electric vehicles, including pure electric vehicles, hybrid electric vehicles, and fuel cell vehicles, are being widely adopted. The EV battery pack is the most critical component, and its technology is often considered the bottleneck restricting the widespread application of electric vehicles. Traditional battery testing methods, such as constant current discharge tests, are accurate but time-consuming and destructive to the battery. Therefore, a fast and reliable method for predicting the remaining capacity of an EV battery pack is urgently needed.
In my research, I focused on lithium iron phosphate (LiFePO4) batteries, which are widely used in electric vehicles due to their high safety, long cycle life, and environmental friendliness. I conducted a series of discharge experiments under various conditions to collect data on voltage, current, and temperature. Based on these data, I established a prediction model using a back-propagation (BP) neural network. To overcome the limitations of the BP algorithm, such as slow convergence and susceptibility to local minima, I combined it with a genetic algorithm (GA), resulting in a GA-BP hybrid algorithm. This approach improved both the accuracy and efficiency of the remaining capacity prediction for the EV battery pack.
Introduction
The EV battery pack is the energy storage system that powers electric vehicles. Unlike conventional vehicles that use internal combustion engines, electric vehicles rely on batteries as their primary power source. The performance of the EV battery pack directly influences the vehicle’s range, acceleration, and hill-climbing capability. However, battery technology still faces challenges, including high cost, heavy weight, and limited energy storage. These issues make battery technology the key factor restricting the development of electric vehicles.
There are various types of batteries used in electric vehicles, including lead-acid batteries, nickel-cadmium batteries, nickel-metal hydride batteries, sodium-sulfur batteries, lithium-ion batteries, and fuel cells. Among these, lithium-ion batteries have become the preferred choice for modern electric vehicles due to their high energy density, long cycle life, and low self-discharge rate. Specifically, lithium iron phosphate (LiFePO4) batteries are widely used because of their excellent thermal stability, safety, and environmental friendliness.
The remaining capacity, also known as the state of charge (SOC), is a critical parameter for evaluating the performance of an EV battery pack. Accurate SOC estimation is essential for battery management systems to ensure safe and efficient operation. Traditional methods for SOC estimation include the open-circuit voltage method, internal resistance method, electrolyte density method, ampere-hour counting method, and recovery effect method. Each of these methods has its limitations, such as poor accuracy, inability to perform online monitoring, or high complexity. Therefore, more advanced techniques, such as fuzzy logic and neural networks, have been explored.
The objective of my research is to develop a reliable prediction model for the remaining capacity of an EV battery pack using neural networks and genetic algorithms. The specific research contents include:
- Conducting discharge experiments on LiFePO4 batteries to collect sample data.
- Analyzing the characteristics of battery voltage, current, and temperature.
- Establishing a BP neural network-based prediction model.
- Improving the model using a GA-BP hybrid algorithm.
- Building a test system to verify the accuracy and reliability of the prediction model.
Performance Analysis of LiFePO4 Batteries
Lithium iron phosphate batteries are a type of lithium-ion battery where the cathode material is lithium iron phosphate instead of traditional materials like lithium cobalt oxide. The chemical formula for a lithium-ion battery can be expressed as:
$$ \text{LiFePO}_4 \rightleftharpoons \text{Li}_{1-x}\text{FePO}_4 + x\text{Li}^+ + x e^- $$
During charging, lithium ions move from the cathode to the anode, and during discharging, they move back. The electrochemical performance of a LiFePO4 battery includes a specific capacity of more than 130 mAh/g, a voltage range of 2.0 V to 4.2 V, and a cycle life of over 2000 cycles with 80% capacity retention.
Key Performance Indicators of Chemical Batteries
Several important parameters are used to describe the performance of an EV battery pack:
| Parameter | Description | Unit |
|---|---|---|
| Electromotive Force (EMF) | Potential difference between positive and negative electrodes | V |
| Open-Circuit Voltage (OCV) | Terminal voltage when the battery is not under load | V |
| Rated Voltage | Voltage specified under standard operating conditions | V |
| Working Voltage | Terminal voltage during discharge, affected by internal resistance | V |
| Termination Voltage | Minimum voltage allowed during discharge | V |
| Theoretical Capacity | Maximum capacity calculated by Faraday’s law | Ah |
| Actual Capacity | Capacity delivered under specific discharge conditions | Ah |
| Rated Capacity | Minimum capacity specified by the manufacturer | Ah |
| Specific Energy | Energy per unit mass | Wh/kg |
| Specific Power | Power per unit mass | W/kg |
| Self-Discharge Rate | Rate of capacity loss when the battery is not in use | %/month |
| Cycle Life | Number of charge/discharge cycles before capacity drops to a specified level | cycles |
The discharge rate is expressed in terms of C-rate, which is defined as the discharge current divided by the rated capacity. For example, a 1C discharge rate for a 3 Ah battery means a discharge current of 3 A.
$$ \text{Discharge Rate} = \frac{\text{Discharge Current (A)}}{\text{Rated Capacity (Ah)}} $$
Discharge Experiment of LiFePO4 Battery
In my experiments, I used a 26650P power-type LiFePO4 battery manufactured by Hangzhou Sciens Energy Technology Co., Ltd., with a nominal voltage of 3.2 V and a rated capacity of 3 Ah. The battery testing equipment used was the ZEEMOO850 battery tester, and the internal resistance was measured with the ZEEMOO3000E internal resistance tester.
The experimental procedure was as follows:
- Charge the battery to 3.65 V and let it rest for 5 hours.
- Discharge the battery using the ZEEMOO850 tester at different constant currents.
- Record the discharge voltage, current, and capacity data via the host computer.
- Measure the internal resistance at open-circuit voltage using the ZEEMOO3000E.
The voltage measurement was performed using the four-wire method to eliminate the influence of contact resistance, especially at high currents. The measurement schematic can be represented as:
$$ U = E_0 + I \cdot R $$
where \(U\) is the working voltage, \(E_0\) is the open-circuit voltage, \(I\) is the discharge current, and \(R\) is the internal resistance.
The discharge capacity was calculated by integrating the current over time:
$$ C = I \cdot t $$
where \(C\) is the capacity in Ah, \(I\) is the discharge current in A, and \(t\) is the discharge time in hours.
Experimental Results and Analysis
The discharge curves of the LiFePO4 battery at different discharge rates are shown in the following table and analysis. It was observed that the higher the discharge rate, the faster the voltage drops, and the lower the discharge capacity at the same time.
| Discharge Current (A) | Discharge Voltage (V) | Discharged Capacity (Ah) |
|---|---|---|
| 3 | 3.2267 | 0.04991 |
| 3 | 3.2142 | 0.29978 |
| 3 | 3.2132 | 0.34967 |
| 3 | 3.2117 | 0.39971 |
| 3 | 3.2015 | 0.69943 |
| 4 | 3.0848 | 1.2242 |
| 4 | 3.0817 | 1.1797 |
| 4 | 3.078 | 1.1352 |
| 4 | 3.0737 | 1.0908 |
| 4 | 3.0693 | 1.0463 |
| 5 | 3.0833 | 1.9852 |
| 5 | 3.0817 | 1.9435 |
| 5 | 3.0802 | 1.9018 |
| 5 | 3.0784 | 1.8602 |
| 5 | 3.0765 | 1.8185 |
| 6 | 3.1104 | 2.6189 |
| 6 | 3.1078 | 2.519 |
| 6 | 3.105 | 2.4191 |
| 6 | 3.1022 | 2.3189 |
| 6 | 3.099 | 2.2191 |
| 7 | 3.0169 | 2.5266 |
| 7 | 3.0157 | 2.4877 |
| 7 | 3.015 | 2.4488 |
| 7 | 3.0137 | 2.4099 |
| 7 | 3.0131 | 2.371 |
| 8 | 2.7566 | 0.47167 |
| 8 | 2.7461 | 0.44945 |
| 8 | 2.7351 | 0.4272 |
| 8 | 2.7237 | 0.40495 |
| 8 | 2.7117 | 0.38274 |
| 9 | 2.9892 | 2.7233 |
| 9 | 2.9859 | 2.5731 |
| 9 | 2.9857 | 2.4231 |
| 9 | 2.9847 | 2.2734 |
| 9 | 2.9823 | 2.1234 |
Additionally, the discharge curves at different temperatures were also obtained. It was found that at higher temperatures, the discharge capacity may decrease due to the reduced efficiency of the electrochemical reactions. The influence of temperature becomes more significant at high discharge rates, especially at low temperatures.
From the experimental data, I derived a theoretical relationship for predicting the remaining capacity. Let \(C(t)\) be the remaining capacity at time \(t\), \(U(t)\) the battery voltage, \(I(t)\) the discharge current, \(C_0\) the rated capacity, \(R_0\) the average internal resistance, and \(\Delta R(t)\) the change in internal resistance. The discharged capacity up to time \(t\) is \(C_1(t)\). Then:
$$ C(t) = C_0 – C_1(t) $$
The implicit relationship among these parameters can be expressed as:
$$ f_1(U(t), I(t), \Delta R(t), t, C(t), C_0, R_0) = 0 $$
Since \(C_0\) and \(R_0\) are constant for a given battery, the equation simplifies to:
$$ f_2(U(t), I(t), \Delta R(t), t, C(t)) = 0 $$
However, the internal resistance change \(\Delta R(t)\) was found to be negligible during the discharge process, as shown in the following table:
| Open-Circuit Voltage (V) | Internal Resistance (mΩ) |
|---|---|
| 3.65 | 22.32 |
| 3.37 | 22.46 |
| 3.082 | 22.36 |
| 2.854 | 22.37 |
| 2.564 | 22.45 |
| 2.372 | 22.43 |
| 2.005 | 22.48 |
Thus, the relationship can be reduced to:
$$ f_3(U(t), I(t), t, C(t)) = 0 $$
In practice, the discharge time \(t\) is often unknown, and the discharged capacity \(C_1(t)\) is also unknown. Therefore, the remaining capacity can be expressed as:
$$ C(t) = g(U(t), I(t)) $$
After discretization with a sampling period \(T\), we obtain:
$$ C(k) = f(U(k), I(k), T) $$
For a constant discharge current, the expression simplifies to:
$$ C(k) = f(U(k), I) $$
Based on this derivation, I determined that the inputs to the prediction model should be the discharge voltage \(U(t)\) and discharge current \(I(t)\), and the output should be the remaining capacity \(C(t)\).
Establishment of the Prediction Model Based on Neural Networks
Neural networks are powerful tools for modeling nonlinear and uncertain systems. They have the ability to approximate any continuous function with arbitrary accuracy, making them suitable for predicting the remaining capacity of an EV battery pack. The back-propagation (BP) algorithm is one of the most widely used neural network training algorithms. It adjusts the connection weights by minimizing the mean squared error between the predicted output and the actual output.
Modeling Mechanism of Neural Networks
System identification is the process of building a mathematical model of a system based on observed input-output data. For nonlinear systems, the NARMAX (Nonlinear AutoRegressive Moving Average with eXogenous inputs) model is often used:
$$ y(k) = f(y(k-1), \ldots, y(k-n_y), u(k-1), \ldots, u(k-n_u)) $$
In neural network-based identification, the network structure and the learning algorithm are selected to approximate the unknown function \(f(\cdot)\). The main steps are:
- Model selection: Determine the number of layers and nodes in each layer.
- Input signal selection: Choose appropriate excitation signals such as white noise or pseudo-random signals.
- Error criterion: Define the performance function, often the mean squared error (MSE).
The mean squared error is defined as:
$$ \text{MSE} = \frac{1}{m p} \sum_{j=1}^{p} \sum_{i=1}^{m} (\hat{y}_{ij} – y_{ij})^2 $$
where \(m\) is the number of output nodes, \(p\) is the number of training samples, \(\hat{y}_{ij}\) is the expected output, and \(y_{ij}\) is the actual output.
Common Neural Network Learning Algorithms
| Algorithm | Description |
|---|---|
| Hebb’s rule | Weights are adjusted based on the correlation between neuron activations. |
| Gradient descent | Weights are updated in the direction of the negative gradient of the error surface. |
| Delta rule | An extension of gradient descent using an error correction signal. |
| Back-propagation (BP) | Error is propagated backwards from the output layer to the hidden layers, adjusting weights. |
In the BP algorithm, the weight update formula is:
$$ w_{ij}(k+1) = w_{ij}(k) + \eta \delta_j y_i $$
where \(\eta\) is the learning rate, \(\delta_j\) is the local gradient, and \(y_i\) is the input from the previous layer.
Modeling Steps for the Battery Remaining Capacity Prediction
The procedure for building a neural network prediction model is as follows:
- Preprocess the sample data by normalizing the inputs to the range [0, 1] or [-1, 1].
- Partition the data into training and testing sets.
- Select the network architecture (number of layers and neurons).
- Train the network using the BP algorithm.
- Validate the model using unseen data.
- Denormalize the output to obtain the actual predicted values.
In my research, I used a three-layer BP network with one hidden layer. The number of hidden nodes was determined through experiments. I found that a network with fewer than 10 hidden nodes yielded poor accuracy, while increasing the number of nodes improved accuracy but increased computational cost. Finally, I selected 11 hidden nodes as a compromise.
The structure of the BP network used for the EV battery pack remaining capacity prediction is shown in the figure above. The network has two input nodes (voltage and current), one hidden layer with 11 nodes, and one output node (remaining capacity). The activation function used in the hidden layer is the sigmoid function:
$$ \sigma(v) = \frac{1}{1 + \exp(-v)} $$
and its derivative is:
$$ \sigma'(v) = \sigma(v) (1 – \sigma(v)) $$
Prediction Results Using BP Neural Network
To validate the model, I used the discharge data at 6 A as the verification data, which was not used during training. The prediction results are shown in the following table and analysis.
| Voltage (V) | Current (A) | Actual Capacity (Ah) | Predicted Capacity (Ah) | Relative Error (%) |
|---|---|---|---|---|
| 3.434 | 8.9973 | 2.985 | 2.8380 | 4.91 |
| 3.284 | 8.9966 | 2.084 | 1.9550 | 4.32 |
| 3.255 | 8.9966 | 1.938 | 1.8300 | 3.61 |
| 3.202 | 8.9966 | 0.563 | 0.6530 | 3.09 |
The training process of the BP network was monitored through the error curve. The training time was approximately 22.13 seconds. From the prediction curves, the BP model was able to follow the actual discharge trend, but the accuracy was limited, with relative errors up to about 5%. This indicated that the standard BP algorithm had room for improvement, especially in terms of convergence speed and global optimization capability.
Improvement of the Prediction Model Using Genetic Algorithm
The BP algorithm has several drawbacks, including slow convergence, susceptibility to local minima, and dependence on initial weights and thresholds. To address these issues, I introduced a genetic algorithm (GA) to optimize the network. GA is a metaheuristic inspired by natural selection and genetics. It maintains a population of candidate solutions and evolves them through selection, crossover, and mutation operations to find the global optimum.
GA-BP Hybrid Algorithm
The GA-BP algorithm first uses GA to search for optimal initial weights and thresholds, which are then used as the starting point for BP training. This approach helps the BP network avoid local minima and speeds up convergence.
The steps of the GA-BP algorithm are as follows:
- Initialize the population of chromosomes, where each chromosome represents a set of weights and thresholds.
- Evaluate the fitness of each chromosome based on the inverse of the sum of squared errors (SSE).
- Apply selection, crossover, and mutation operations to generate a new population.
- Repeat steps 2-3 until the fitness or error criterion is met.
- Use the best chromosome to initialize the BP network.
- Train the BP network using the standard back-propagation algorithm.
The fitness function is defined as:
$$ f(i) = \frac{1}{E(i)} $$
where \(E(i)\) is the sum of squared errors for individual \(i\):
$$ E(i) = \sum_{p} \sum_{k} (T_{kp} – V_{kp})^2 $$
In this equation, \(T_{kp}\) is the target output and \(V_{kp}\) is the network output for sample \(p\) and output node \(k\).
The GA parameters used in my experiments are summarized in the following table:
| Parameter | Value |
|---|---|
| Population size | 30 |
| Crossover probability (Pc) | 0.6 |
| Mutation probability (Pm) | 0.2 |
| Termination generation | 100 |
| Encoding | Real-coded |
After running the GA, I obtained the optimal initial weights and thresholds. These were then used to train the BP network. The prediction results using the GA-BP algorithm showed significant improvement over the standard BP algorithm. The training time was reduced to approximately 2.73 seconds, and the prediction accuracy improved, with relative errors mostly below 2%.
Optimization of Network Training Parameters
In addition to initial weights and thresholds, the training parameters of the BP network, such as the maximum number of training epochs, learning rate, and the training goal (target MSE), significantly affect the convergence speed. I further applied GA to optimize these parameters.
The training goal (goal) is the desired mean squared error. I optimized the goal first, keeping the learning rate and epochs fixed. The GA search for the optimal goal resulted in a value of 0.02, which reduced the training time to 22.95 seconds compared to the original fixed goal. However, the improvement was not dramatic.
Next, I optimized the learning rate (lr) and the goal together. The GA search found the optimal lr = 0.013 and goal = 0.07, reducing the training time to 9.90 seconds.
Finally, I optimized all three parameters: epoch, lr, and goal simultaneously. The optimal values were epoch = 1550, lr = 0.012, and goal = 0.08, resulting in a training time of only 2.75 seconds. The GA parameter settings for each optimization level are listed in the following table:
| Optimization stage | Population size | Search range | Termination generation | Encoding length | Crossover probability | Mutation probability |
|---|---|---|---|---|---|---|
| goal only | 4 | 0.01~0.08 | 10 | 3 | 0.6 | 0.01 |
| lr and goal | 8 | 0.001~0.016 (lr) | 10 | 4 | 0.6 | 0.01 |
| epoch, lr, goal | 32 | 500~2050 (epoch) | 10 | 5 | 0.6 | 0.01 |
Comprehensive Optimization
To achieve both high accuracy and fast training, I combined the optimization of network structure parameters and training parameters. The final optimized settings were:
- Initial weights and thresholds: obtained from GA optimization
- Training goal (goal) = 0.08
- Learning rate (lr) = 0.012
- Maximum number of epochs (epoch) = 1550
With these settings, the BP network training time was 6.375 seconds when using the GA-optimized initial weights and thresholds and the optimized training parameters. The prediction results showed that the predicted remaining capacity closely matched the actual values throughout the entire discharge process. The average error was reduced to approximately 0.3 Ah, which is acceptable for engineering applications.
Comparison of Prediction Results
To demonstrate the effectiveness of the GA-BP algorithm, I compared the prediction results obtained with the standard BP algorithm and the GA-BP algorithm for the same set of data. The comparison is shown in the following table:
| Current (A) | Voltage (V) | Actual Capacity (Ah) | BP Predicted (Ah) | GA-BP Predicted (Ah) | BP Relative Error (%) | GA-BP Relative Error (%) |
|---|---|---|---|---|---|---|
| 3 | 3.2267 | 2.955 | 2.8110 | 2.8800 | 4.81 | 2.53 |
| 3 | 3.2132 | 2.6552 | 2.5202 | 2.7242 | 4.52 | 2.34 |
| 3 | 3.2015 | 2.3055 | 2.3355 | 2.2175 | 1.32 | 2.93 |
| 3 | 3.1917 | 2.0556 | 2.1636 | 1.9906 | 3.61 | 2.16 |
| 3 | 3.1876 | 1.9556 | 1.8146 | 2.0006 | 4.69 | 1.51 |
| 3 | 3.1757 | 1.6558 | 1.5958 | 1.5958 | 2.01 | 2.03 |
| 3 | 3.1696 | 1.5059 | 1.6259 | 1.5899 | 4.05 | 2.83 |
The results clearly show that the GA-BP algorithm reduced the prediction error to about 1-2% on average, compared to 3-5% for the standard BP algorithm. This improvement is significant for estimating the remaining capacity of an EV battery pack.
Furthermore, the convergence speed of the GA-BP algorithm was much faster. The training time was reduced from 22.13 seconds (BP) to 2.75 seconds (GA-BP with optimized parameters), representing an eightfold improvement. This is crucial for real-time battery management systems in electric vehicles.
Establishment of the Test System and Verification
To verify the generalization capability of the improved prediction model, I built a test system based on the GA-BP model. The model was trained using discharge data from 15 different current levels: 0.5 A, 1 A, 2 A, 3 A, 4 A, 5 A, 6 A, 7 A, 8 A, 9 A, 15 A, 20 A, 30 A, 50 A, and 60 A. The network architecture was the same as described earlier: 2 inputs, 11 hidden nodes, and 1 output.
First, I used the model to predict the remaining capacity at non-sampled voltage points for a discharge current of 3 A. I selected voltages from 3.3 V to 3.1 V in steps of 0.02 V. The prediction results are shown in the table below:
| Current (A) | Voltage (V) | Predicted Capacity (Ah) |
|---|---|---|
| 3 | 3.3 | 2.9735 |
| 3 | 3.28 | 2.9682 |
| 3 | 3.26 | 2.8653 |
| 3 | 3.24 | 2.7192 |
| 3 | 3.22 | 2.5634 |
| 3 | 3.2 | 2.3436 |
| 3 | 3.18 | 1.9438 |
| 3 | 3.16 | 1.7466 |
| 3 | 3.14 | 1.5968 |
| 3 | 3.12 | 1.3569 |
| 3 | 3.1 | 1.2442 |
To further test the model’s generalization, I predicted the remaining capacity for a discharge current of 2.5 A, which was not included in the training data set. The voltages selected were in the same range. The predictions are presented below:
| Current (A) | Voltage (V) | Predicted Capacity (Ah) |
|---|---|---|
| 2.5 | 3.4 | 2.9846 |
| 2.5 | 3.38 | 2.9666 |
| 2.5 | 3.36 | 2.8935 |
| 2.5 | 3.34 | 2.7752 |
| 2.5 | 3.32 | 2.6024 |
| 2.5 | 3.3 | 2.4368 |
| 2.5 | 3.28 | 2.0456 |
| 2.5 | 3.26 | 1.8985 |
| 2.5 | 3.24 | 1.7123 |
| 2.5 | 3.22 | 1.4650 |
| 2.5 | 3.2 | 1.3368 |
To validate these predictions, I performed actual discharge experiments at 3 A and 2.5 A with a shorter sampling interval (10 seconds instead of 60 seconds). The actual measured capacities were compared with the predicted values. The results are shown in the following tables:
| Current (A) | Voltage (V) | Predicted Capacity (Ah) | Measured Capacity (Ah) | Relative Error (%) |
|---|---|---|---|---|
| 3 | 3.4 | 2.9735 | 2.9315 | 1.39 |
| 3 | 3.38 | 2.9682 | 2.9552 | 0.43 |
| 3 | 3.36 | 2.8653 | 2.8893 | 0.80 |
| 3 | 3.34 | 2.7192 | 2.6862 | 1.09 |
| 3 | 3.32 | 2.5634 | 2.5114 | 1.73 |
| 3 | 3.3 | 2.3436 | 2.3836 | 1.33 |
| 3 | 3.28 | 1.9438 | 1.9098 | 1.13 |
| 3 | 3.26 | 1.7466 | 1.7676 | 0.69 |
| 3 | 3.24 | 1.5968 | 1.5648 | 1.07 |
| 3 | 3.22 | 1.3569 | 1.3309 | 0.87 |
| 3 | 3.2 | 1.2442 | 1.2142 | 1.00 |
| Current (A) | Voltage (V) | Predicted Capacity (Ah) | Measured Capacity (Ah) | Relative Error (%) |
|---|---|---|---|---|
| 2.5 | 3.4 | 2.9846 | 2.8966 | 2.93 |
| 2.5 | 3.38 | 2.9666 | 2.8906 | 2.53 |
| 2.5 | 3.36 | 2.8935 | 2.9545 | 2.03 |
| 2.5 | 3.34 | 2.7752 | 2.7062 | 2.29 |
| 2.5 | 3.32 | 2.6024 | 2.6544 | 1.73 |
| 2.5 | 3.3 | 2.4368 | 2.3708 | 2.21 |
| 2.5 | 3.28 | 2.0456 | 1.9686 | 2.56 |
| 2.5 | 3.26 | 1.8985 | 1.9765 | 2.62 |
| 2.5 | 3.24 | 1.7123 | 1.6453 | 2.23 |
| 2.5 | 3.22 | 1.4650 | 1.4130 | 1.73 |
| 2.5 | 3.2 | 1.3368 | 1.2888 | 1.60 |
The relative errors for the 3 A discharge predictions were all below 1.8%, and for the 2.5 A predictions, the errors were below 3%. These results confirm that the GA-BP prediction model exhibits good generalization capability and can accurately predict the remaining capacity of an EV battery pack even for conditions not explicitly used in training.
Conclusion
In this research, I have successfully developed a prediction model for the remaining capacity of an EV battery pack using a combination of BP neural networks and genetic algorithms. The main contributions of this work are summarized as follows:
- I conducted systematic discharge experiments on LiFePO4 batteries and analyzed the dynamic characteristics of the EV battery pack under various discharge currents and temperatures.
- I derived a theoretical relationship between the remaining capacity and measurable parameters (voltage and current), providing a basis for the neural network model.
- I established a BP neural network model with two inputs and one output, and demonstrated its capability to predict the remaining capacity, albeit with limited accuracy and slow convergence.
- I proposed a GA-BP hybrid algorithm that optimizes the initial weights and thresholds, as well as the training parameters (goal, learning rate, and maximum epochs). This improved both the prediction accuracy and the training speed.
- I built a test system and verified the model’s generalization ability using data that were not part of the training set. The prediction errors were within acceptable engineering tolerances (less than 3%).
The research confirms that the GA-BP algorithm is an effective method for predicting the remaining capacity of an EV battery pack. It addresses the limitations of traditional methods and standard BP networks by providing a more accurate and faster solution. However, there are still some limitations:
- The model does not explicitly account for environmental temperature, self-discharge, recovery effects, and battery aging, which may affect the prediction accuracy under extreme conditions.
- The model was developed specifically for LiFePO4 batteries; its applicability to other battery chemistries needs further investigation.
Future work could extend the model to incorporate additional influencing factors and validate it across a wider range of battery types and operating conditions. Nevertheless, the developed GA-BP model provides a solid foundation for real-time battery management systems in electric vehicles, contributing to the advancement of EV battery pack technology.