A power battery temperature prediction method based on an iPSO-LSTM model

Abstract

The application discloses a power battery temperature prediction method based on an iPSO-LSTM model and belongs to the technical field of power battery manufacturing prediction. The method defines inert particles, and the position of the inert particles is reset to the position of the global optimal particle in the iteration process, but the search speed of the particle is not changed. Through the strategy of resetting the position but not changing the search speed, the particle after replacing the position has the search significance, and time is not wasted. The definition and processing strategy of the inert particle added in the application make the iPSO-LSTM model increase the convergence speed without losing the accuracy, reduce the possibility of invalid particle search, increase the search range near the optimal value in each round, and finally quickly and accurately optimize the optimal hyperparameter. The optimal hyperparameter is brought into the LSTM model to realize the fast and accurate prediction of the power battery temperature.

Description

Technical Field

[0001] This invention relates to a method for predicting the temperature of power batteries based on the iPSO-LSTM model, belonging to the field of power battery manufacturing prediction technology. Background Technology

[0002] In recent years, lithium-ion batteries, as an upgrade to nickel-metal hydride and lead-acid batteries, have become a key focus of technological research and industrialization due to their high energy density, high rate capability, and high safety. Meanwhile, battery manufacturing technology has evolved from workshop-style production to automation and today's intelligent manufacturing, with the industry continuously expanding. China has become the world's largest producer and consumer of lithium-ion batteries.

[0003] Power battery charging mainly relies on current charging. However, current charging leads to a rapid increase in battery temperature, increased internal polarization reactions, prolonged charging time, and decreased charging capacity. Severe polarization reactions can also cause battery overheating and lithium plating, and even thermal runaway. Therefore, temperature monitoring during battery charging is necessary. The most basic temperature monitoring method uses temperature sensors to acquire battery temperature in real time. When the temperature reaches a threshold, corresponding measures are taken to avoid safety issues. However, this method cannot predict the temperature in advance; it can only obtain the current temperature. Therefore, temperature prediction technology has emerged to predict the temperature during battery charging. Most existing temperature prediction technologies establish an equivalent variable parameter thermal model for ternary lithium-ion batteries, use state equation analysis to establish the correlation of battery temperature, and finally use prediction algorithms for real-time prediction (see "Wang Yuchao. Research on Real-time Prediction Algorithm for Internal Temperature and State of Charge of Power Batteries [C]. Jiangxi University of Science and Technology, 2022."). While the above method of mathematically modeling the battery improves the accuracy of prediction, it has high computational complexity and can only be used in professional high-performance servers. However, for battery temperature prediction that does not require ultra-high accuracy, it is not very practical and wastes computing power.

[0004] In recent years, neural network prediction models have matured, leading to the development of LSTM (Long Short-Term Memory) recurrent neural networks for predicting power battery temperatures. These models possess the ability to extract temporal features from multi-temporal data, addressing the nonlinearity of temperature changes and offering significant advantages in battery temperature prediction. However, the accuracy of LSTM prediction models depends heavily on the selection of hyperparameters. Appropriate hyperparameter selection leads to better and more accurate predictions, while biased selection results in inaccurate predictions. In practical applications, hyperparameter tuning is typically done manually, which requires extensive experimentation, demands high technical skills, and is time-consuming and labor-intensive.

[0005] To address the aforementioned issues, some scholars have proposed using the Particle Swarm Optimization (PSO) algorithm to optimize LSTM hyperparameters. This method treats hyperparameters as the position coordinates of the particle swarm and the root mean square error (RMSE) predicted by the LSTM as the fitness of the particle swarm, iterating through the particle swarm to find the optimal hyperparameters. However, this method suffers from low accuracy and is prone to getting trapped in local optima. Some particles engage in meaningless searches near their own positions, and when PSO is combined with LSTM, these particles further increase the model's time consumption. This is extremely detrimental to power battery temperature prediction methods that require high accuracy and real-time performance. Summary of the Invention

[0006] To achieve efficient and high-precision prediction of the maximum temperature during battery charging, thus assisting enterprises in hazard prediction and warning, this invention proposes a deep learning-based method for predicting the temperature of power batteries. This method first defines a novel combined model, using the root mean square error (RMSE) in the time prediction model as the objective function of the particle swarm optimization (PSO) algorithm, and incorporating inert particle decision-making for PSO iteration. Specifically, addressing the problem of easily getting trapped in local optima when iterating the PSO to find the optimal hyperparameters, this application defines the concept of inert particles and proposes a decision-making scheme for inert particles, establishing an iPSO-LSTM model. Adding the definition and handling strategy of inert particles increases the convergence speed of the iPSO-LSTM model without sacrificing accuracy, reduces the possibility of invalid particle searches, and increases the search range near the optimal value in each round. Ultimately, it quickly and accurately finds the optimal hyperparameters, which are then fed into the LSTM model to predict the temperature of the power battery.

[0007] A deep learning-based method for predicting the temperature of a power battery is disclosed. The method uses the root mean square error (RMSE) of an LSTM model as the fitness value of a dynamic decision particle swarm optimization (PSO) algorithm. The number of hidden layer nodes and the learning rate of the LSTM model are set as optimization parameters for the PSO. Based on changes in fitness and iteration count, the optimal position of a single particle is updated, along with its coordinates and velocity, until the RMSE stabilizes. The optimal values ​​of the optimization parameters are then determined. An LSTM model, i.e., an iPSO-LSTM model, is constructed using these optimal values. The iPSO-LSTM model is trained, and the trained iPSO-LSTM model is used for battery temperature prediction.

[0008] Optionally, the method includes:

[0009] Step 1: Obtain battery temperature data and construct training and testing sets; the temperature data is the highest temperature of the battery during a single charge.

[0010] Step 2: Initialize the particle population X by setting the particle population size, number of iterations, initial velocity, and position range.

[0011] Let x(α,β) be any particle in the particle swarm X, and let its position coordinates (α,β) be the optimization parameters of the particle swarm;

[0012] Step 3: Construct an LSTM prediction model using the particle coordinates in particle population X, where α is the number of hidden layer units in the LSTM prediction model and β is the learning rate of the LSTM prediction model.

[0013] Step 4: Use the root mean square error (RMSE) of the LSTM prediction model as the fitness value for the particle swarm iteration update to perform particle swarm optimization. During the optimization process, define inert particles and use inert particle decisions to iteratively update the position coordinates and velocity of the particle swarm until the fitness value (RMSE) meets the evaluation criteria for the end of the iteration or the number of iterations reaches the maximum, and determine the optimal value of the optimization parameters (α,β).

[0014] Step 5: Using the optimal value of the optimization parameters (α,β) as the number of hidden layer units and β as the learning rate, construct the iPSO-LSTM model and train it to obtain the trained iPSO-LSTM model.

[0015] Step 6: Obtain the temperature data of the battery to be predicted and input it into the trained iPSO-LSTM model for temperature prediction.

[0016] Optionally, the inert particle decision refers to, after each iteration update of the particles, arranging the positions of the particle swarm in descending order of fitness value, defining the bottom 20% of the particles as inert particles, resetting the position of the inert particles to the position of the globally optimal particle, but without changing the search speed of the particle, and continuing to seek optimization.

[0017] Optionally, during the particle swarm initialization process in step 2, the particle swarm coordinates and velocities are initialized according to the following formula:

[0018] pso i 0 (γ)=lb(γ)+r*(ub(γ)-lb(γ))γ=1, 2

[0019] v i 0 (γ)=-vmax(γ)+2vmax(γ)*rγ=1, 2

[0020] Among them, pso i 0 (γ) represents the value of the γ-th optimization parameter in the initial position coordinates of the i-th particle, v i 0(γ) represents the search speed of the γth optimization parameter of the initial i-th particle, vmax(γ) represents the upper limit of the search speed of the γth optimization parameter of the particle swarm, ub(γ) and lb(γ) are the upper and lower limits of the γth optimization parameter of the particle swarm, respectively, and r is a random number uniformly distributed between [0,1].

[0021] Optionally, in step 4, the root mean square error (RMSE) of the LSTM prediction model is expressed as the standard deviation of the deviation between the true and predicted values:

[0022]

[0023] In the formula, n is the number of the highest temperature sample points in the battery temperature data obtained in step 1; x j This represents the actual value of the j-th sample point; This is the predicted value for the j-th highest temperature sample point.

[0024] Optionally, in step 4, during the particle swarm optimization process, inert particles are defined, and the position coordinates and velocities of the particle swarm are iteratively updated using inert particle decisions, including:

[0025] During particle swarm optimization, the inertia weight w and learning factors c1 and c2 are updated based on fitness and iteration number.

[0026]

[0027]

[0028]

[0029] Where, the subscript max represents the maximum value of the corresponding parameter, and the subscript min represents the minimum value of the corresponding parameter; t is the cumulative number of iterations, and T is the maximum number of iterations;

[0030] The particle velocity and coordinates are iteratively updated based on the updated variables w, c1, and c2, as shown in the following expression:

[0031] v i k+1 (γ)=wv i k (γ)+c1r1(p i k -x i k )+c2r2(g k -x i k )γ=1、2

[0032] pso i k+1 (γ)=psoi k (γ)+v i k+1 (γ)γ=1、2

[0033] Where v i k (γ) is the search velocity of the γ-th optimization parameter of the i-th particle in the k-th iteration, pso i k (γ) is the value of the γ-th optimization parameter of the i-th particle in the k-th iteration; g k Let x be the global optimum value in the k-th iteration. i k Let r1 and r2 be the fitness value of the i-th particle in the k-th iteration, and r1 and r2 be random numbers that are uniformly distributed in the interval [0,1].

[0034] After solving for the fitness value in each round of particle swarm optimization, the positions of the particles are sorted in descending order of fitness, and the bottom 20% of the particles are defined as inert particles.

[0035] Change the position of the inert particle to the position of the globally optimal particle, but do not change the particle's search speed. The expression is as follows:

[0036] pso y k =pso m k

[0037] Where y represents that the y-th particle is an inert particle, k represents the iteration number, and pso y k represents the position coordinates of the y-th particle in the k-th iteration; m represents the fitness of the m-th particle, which is the global optimum in this iteration, pso. m k This represents the position coordinates of the m-th particle in the k-th iteration.

[0038] Optionally, in step 1, when acquiring battery temperature data and constructing training and test sets, the first 90% of the acquired battery temperature data is divided into training data to form the training set, and the remainder is test data to form the test set.

[0039] Optionally, the LSTM prediction model includes an input layer, a hidden layer, an output layer, and a Dropout layer. When constructing the LSTM prediction model using particle coordinates in the particle population X, the method further includes setting the learning rate decay rate of the LSTM prediction model to 0.2, setting the number of learning rate decay rounds of the LSTM prediction model to 125, setting the number of layers of the input layer, output layer, and Dropout layer to 1, and setting the maximum number of iterations to 250.

[0040] Optionally, the fitness value RMSE tends to stabilize and meets the criteria for ending the iteration, indicating that:

[0041] g k -g k-2 <0.0001

[0042] Among them, g k Let g represent the global optimum in the k-th particle swarm iteration. k-2 This represents the global optimum value in the (k-2)th particle swarm iteration.

[0043] This application also provides a method for monitoring the temperature during the charging process of a power battery. The method uses the above-mentioned temperature prediction method to predict the temperature of the battery during charging, thereby realizing temperature monitoring during the battery charging process.

[0044] The beneficial effects of this invention are:

[0045] After solving for the fitness value in each round of particle swarm optimization, the fitness of the particle swarm is sorted in descending order, and the bottom 20% of particles are defined as inert particles. Searches performed by inert particles at these positions are considered meaningless searches. Therefore, the positions of inert particles are changed, resetting them to the positions of the globally optimal particles, without altering their search speed. This strategy of resetting positions without changing search speed gives the repositioned particles meaningful search opportunities without wasting time. This operation preserves the randomness of particle search directions and reduces the possibility of particles performing invalid searches on previously searched paths. Adding the definition and handling strategy for inert particles allows the iPSO-LSTM model to increase convergence speed without sacrificing accuracy, reduce the possibility of invalid particle searches, and increase the search range near the optimal value in each round. Ultimately, it quickly and accurately finds the optimal hyperparameters, which are then fed into the LSTM model to achieve rapid and accurate prediction of power battery temperature. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart of a power battery temperature prediction method based on an iPSO-LSTM model disclosed in one embodiment of the present invention.

[0048] Figure 2 This is a structural diagram of a power battery temperature prediction method based on an iPSO-LSTM model disclosed in one embodiment of the present invention.

[0049] Figure 3 This is a diagram showing the prediction results of the maximum temperature of a power battery using the method of this application and based on an existing LSTM model, as disclosed in one embodiment of the present invention.

[0050] Figure 4 This is a graph showing the error prediction result of the maximum temperature of a power battery using the method of this application and based on an existing LSTM model, as disclosed in one embodiment of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0052] Example 1:

[0053] This embodiment provides a method for predicting the temperature of a power battery based on an iPSO-LSTM model. The method is based on an adaptive weighted iPSO-LSTM model. The root mean square error of the LSTM model is used as the fitness value of the dynamic decision particle swarm optimization algorithm. The number of hidden layer nodes and the learning rate of the LSTM are set as optimization parameters for the particle swarm. Based on changes in fitness and iteration count, the optimal position of each individual particle is updated, along with its coordinates and velocity.

[0054] The particle swarm optimization is iteratively updated until the comfort value RMSE meets the criteria for ending the iteration or the number of iterations reaches the maximum. The optimization parameters are then determined, and an LSTM model is constructed using the optimal hyperparameters. After training, the model is used for power battery temperature prediction.

[0055] See Figure 1 The method includes:

[0056] Step 1: Obtain battery temperature data and construct training and testing sets; the temperature data is the highest temperature of the battery during a single charge.

[0057] Step 2: Initialize the particle population X by setting the particle population size, number of iterations, initial velocity, and position range.

[0058] Let x(α,β) be any particle in the particle swarm X, and let its position coordinates (α,β) be the optimization parameters of the particle swarm;

[0059] Step 3: Construct an LSTM prediction model using the particle coordinates in particle population X, where α is the number of hidden layer units in the LSTM prediction model and β is the learning rate of the LSTM prediction model.

[0060] Step 4: Use the root mean square error (RMSE) of the LSTM prediction model as the fitness value for the particle swarm iteration update to perform particle swarm optimization. During the optimization process, define inert particles and use inert particle decisions to iteratively update the position coordinates and velocity of the particle swarm until the fitness value (RMSE) meets the evaluation criteria for ending the iteration or the number of iterations reaches the maximum. Stop the iteration and determine the optimal value of the optimization parameters (α,β).

[0061] Step 5: Using the optimal value of the optimization parameters (α,β) as the number of hidden layer units and β as the learning rate, construct the iPSO-LSTM model and train it to obtain the trained iPSO-LSTM model.

[0062] Step 6: Obtain the temperature data of the battery to be predicted and input it into the trained iPSO-LSTM model for temperature prediction.

[0063] Example 2

[0064] This embodiment provides a deep learning-based method for predicting the temperature of a power battery, the method comprising:

[0065] Step 1: Import the dataset and extract the highest temperature data of the power battery during a single charge.

[0066] This embodiment uses an existing publicly available dataset, which battery manufacturers can also collect in practical applications.

[0067] The publicly available dataset used in this embodiment is NASA lithium-ion battery data tested at NASA's PCoE Research Center at the Idaho National Laboratory. This dataset uses 18650 commercially available lithium-ion batteries with a rated capacity of 2 A·h, divided into multiple groups for experiments, and tested at room temperature using three different operating curves (charge, discharge, and impedance). Table 1 shows the experimental parameters, namely ambient temperature (AT), charging current (CC), discharging current (DC), end-of-discharge voltage (EOC), and end-of-life criterion (EOLC).

[0068] Table 1: Experimental Parameters

[0069]

[0070] This experimental example selects data from battery #5 in the dataset. Battery #5 data has a "cycle" field, which includes four fields: type, ambient_temperature, time, and data. The "type" field indicates the type of measurement data, including charging, discharging, and impedance types. "ambient_temperature" is the temperature at the time of measurement, "time" is the measurement time, and "data" is the measurement data. Charging types include fields such as "Voltage_measured," "Current_measured," "Temperature_measured," "Current_charge," "Voltage_charge," and "Time." Discharging types include fields such as "Voltage_measured," "Current_measured," "Temperature_measured," "Current_load," "Voltage_load," "Time," and "Capacity." Impedance types include fields such as "Sense_current," "Battery_current," "Current_ratio," "Battery_impedance," "Rectified_Impedance," "Re," and "Rct."

[0071] Step 2: Initialize the particle population X. Set the particle population size, number of iterations, initial velocity, and position range.

[0072] The upper and lower limits of the parameters set in this embodiment and the initialization can be seen in Tables 2 and 3:

[0073] Table 2: Particle population size, number of iterations, initial velocity, and upper and lower limits of the position interval

[0074]

[0075] Where v(1) and v(2) represent the search speeds of the optimization parameters α and β, respectively.

[0076] Table 3: Particle population size, number of iterations, initial velocity and position interval initialization

[0077]

[0078] In Table 3, SearchAgents_no represents the maximum number of iterations, dim represents the dimension, and Max_iter represents the population size. Considering the experimental time, the maximum number of iterations and the population size are set to 10 in this embodiment. In actual applications, these values ​​can be set according to actual needs.

[0079] Step 3: Initialize particle velocity and position.

[0080] A particle x(α,β) is randomly generated in X, where (α,β) are the optimization parameters of the particle swarm, α is the number of hidden units in the LSTM layer, and β is the learning rate. The particle swarm coordinates and velocity are initialized using the following formula.

[0081] pso i 0 (γ)=lb(γ)+r*(ub(γ)-lb(γ))γ=1, 2

[0082] v i 0 (γ)=-vmax(γ)+2vmax(γ)*rγ=1, 2

[0083] Among them, pso i 0 (γ) represents the value of the γ-th optimization parameter in the initial position coordinates of the i-th particle, v i 0 (γ) represents the search speed of the γth optimization parameter of the initial i-th particle, vmax(γ) represents the upper limit of the search speed of the γth optimization parameter of the particle swarm, ub(γ) and lb(γ) are the upper and lower limits of the γth optimization parameter of the particle swarm, respectively, and r is a random number uniformly distributed between [0,1].

[0084] Step 4: Set the root mean square error (RMSE) of the iPSO-LSTM model as the fitness value. RMSE is usually expressed as the standard deviation of the deviation between the true value and the predicted value.

[0085]

[0086] In the formula, n is the number of sample points with the highest temperature; x j This represents the actual value of the j-th sample point; This is the predicted value for the j-th highest temperature sample point.

[0087] Step 5: Using the particle coordinates from Step 3, construct an LSTM prediction model. After training and prediction, use RMSE as the particle swarm fitness, and update the inertia weight w and learning factors c1 and c2 based on the fitness and the number of iterations.

[0088]

[0089]

[0090]

[0091] Where, the subscript max represents the maximum value of the corresponding parameter, and the subscript min represents the minimum value of the corresponding parameter; t is the cumulative number of iterations, and T is the maximum number of iterations.

[0092] The LSTM cell construction method is as follows:

[0093]

[0094] Among them, h t C represents short-term historical information at time t. t This represents long-term historical information at time t. For candidate long-term historical information at time t, x t Let W represent the input sample at time t, σ be the Sigmoid activation function, tanh be the hyperbolic tangent activation function, and W be the input sample at time t. f and b f These are the weight matrix and bias vector of the forget gate, respectively, f t W is the output of the forget gate of the Length Short-Term Memory network at time t. i and b i These are the weight matrix and bias vector of the input gate, i. t W is the output of the input gate of the Length Short-Time Memory network at time t. o and b o These are the weight matrix and bias vector of the output gate, respectively. t The output of the short-term memory network at time t is the predicted temperature of the power battery.

[0095] Step Six: Iteratively update the particle velocity and coordinates using the updated variables w, c1, and c2, as shown in the following expression:

[0096] v i k+1 (γ)=wv i k (γ)+c1r1(p i k -x i k )+c2r2(g k -x i k )γ=1、2

[0097] pso i k+1 (γ)=pso i k (γ)+v i k+1 (γ)γ=1、2

[0098] Where v i k (γ) is the search velocity of the γ-th optimization parameter of the i-th particle in the k-th iteration, pso i k(γ) is the value of the γ-th optimization parameter of the i-th particle in the k-th iteration; g k Let x be the global optimum value in the k-th iteration. i k Let r1 and r2 be the fitness value of the i-th particle in the k-th iteration, and r1 and r2 be random numbers uniformly distributed between [0,1]. During the particle swarm iteration, the particle's velocity v i k =[v i k (1), v i k (2)],v i k (1) is the search speed of the first optimization parameter α, v i k (2) is the search speed of the second optimization parameter β.

[0099] After each round of particle swarm optimization to determine the fitness value, the particle positions are arranged in descending order according to fitness. The bottom 20% of particles are defined as inert particles. Searches performed by inert particles at these positions are considered meaningless searches, so their positions are changed to the positions of the globally optimal particles, without altering their search speed. The expression is as follows:

[0100] pso y k =pso m k

[0101] Where y represents that the y-th particle is an inert particle, k represents the iteration number, and pso y k represents the position coordinates of the y-th particle in the k-th iteration; m represents the fitness of the m-th particle, which is the global optimum in this iteration, pso. m k This represents the position coordinates of the m-th particle in the k-th iteration.

[0102] The strategy of resetting the position without changing the search speed gives the repositioned particle a search purpose and avoids wasting time. This operation preserves the randomness of the particle's search direction and reduces the possibility of the particle performing an invalid search on an already searched path.

[0103] The iteration continues until the fitness RMSE stabilizes or the predetermined maximum number of iterations is reached, at which point the iteration stops and the values ​​of the optimization parameters (α,β) are determined.

[0104] Step 7: Input the optimal parameters into the LSTM network for training and prediction.

[0105] The first 90% of the data is divided into training data and the remainder into test data. The optimal hyperparameter set (α, β) is found using the particle swarm optimization algorithm and then fed into the LSTM architecture.

[0106]

[0107] The final prediction of the highest temperature during the power battery charging cycle was obtained.

[0108] To evaluate the predictive performance of the method in this application (hereinafter referred to as the iPSO-LSTM method), this embodiment compares it with a base model to determine the superiority or inferiority of the method. The base model adopts the Long Short-Term Memory Network method (hereinafter referred to as the LSTM method).

[0109] For an introduction to the Long Short-Term Memory (LSTM) network method, please refer to: "Wang Qiuwen, Chen Yanru, Liu Yuanchun. Short-term passenger flow prediction of urban rail transit based on convolutional long short-term memory neural network [J]. Control and Decision, 2021, 36(11):2760-2770."

[0110] Table 4: Comparison of Prediction Errors for Battery Manufacturing Capacity under Different Prediction Methods

[0111] Mean Absolute Error (MAE) 0.1398 0.1715 Root Mean Square Error (RMSE) 0.1687 0.2123

[0112] To verify the accuracy and effectiveness of the deep learning-based power battery temperature prediction method proposed in this application, the following simulation experiments were conducted using the method of this application and the existing LSTM method. For actual production process conditions, Figure 3 , Figure 4 The prediction results and errors of each method are presented separately. The dataset used in this embodiment is NASA lithium-ion battery data tested at NASA's PCoE Research Center at the Idaho National Laboratory. This dataset uses commercially available 18650 lithium-ion batteries with a rated capacity of 2 Ah, divided into multiple groups for experiments. The highest temperature during all charging cycles was collected as data samples, with the first 90% selected as the training set and the last 10% as the test set.

[0113] Depend on Figure 3 As can be seen, the circled lines represent the actual value of the highest temperature during a single charge, the star shape represents the prediction result of the LSTM model, and the square represents the prediction result of the iPSO-LSTM model method based on adaptive weights. All these methods can roughly predict the overall trend of manufacturing capability changes. Furthermore, it can be seen that at time periods of {0, 12} and {15, 35}, the prediction result of the iPSO-LSTM model method (i.e., the method of this application) represented by the square is closer to the actual value, while the prediction result of the LSTM model represented by the star can only predict the general trend of the actual value.

[0114] Depend on Figure 4 It can be seen that when the time is {25, 35}, the prediction error of the iPSO-LSTM model method provided in this application is smaller than that of the LSTM model. The LSTM model fails to achieve the optimal prediction effect due to improper selection of hyperparameters, resulting in a larger residual.

[0115] After incorporating particle swarm optimization, the LSTM model shows significant improvement, with predicted data more closely matching the original data and fluctuation trends still predictable. It has wider applicability than the simple LSTM model, saving most of the time spent adjusting LSTM hyperparameters. Compared to ordinary PSO-LSTM, the iPSO-LSTM model algorithm based on adaptive weights used in this application can achieve rapid convergence to find the global optimum, avoiding meaningless searches of particles trapped in local optima near their own positions. It saves most of the time spent traversing the LSTM model, saving costs while improving efficiency.

[0116] After combining the PSO particle swarm algorithm with adaptive weights, it can be seen that the prediction accuracy of LSTM is significantly improved, and the predicted trend is more consistent with the trend of the original data.

[0117] In the comparison of the two error metrics, the iPSO-LSTM model combination algorithm improved the mean absolute error (MAE) by 18.5% and the root mean square error (RMSE) by 20.5% compared to the LSTM algorithm. This demonstrates that in predicting the maximum battery temperature, the iPSO-LSTM model algorithm based on adaptive weights proposed in this application can further improve the prediction accuracy of the LSTM method.

[0118] Some steps in the embodiments of the present invention can be implemented using software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk.

[0119] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the temperature of a power battery based on an iPSO-LSTM model, characterized in that, The method uses the root mean square error (RMSE) of the LSTM model as the fitness value of the dynamic decision particle swarm optimization algorithm. It sets the number of hidden layer nodes and the learning rate of the LSTM model as optimization parameters for the particle swarm. Based on the changes in fitness and iteration count, it updates the optimal position of a single particle, as well as the particle's coordinates and velocity, until the fitness value (RMSE) stabilizes. Then, it determines the optimal values ​​of the optimization parameters, constructs an LSTM model (iPSO-LSTM model) using these optimal values, trains the iPSO-LSTM model, and uses the trained iPSO-LSTM model for battery temperature prediction. The method includes: Step 1: Obtain battery temperature data and construct training and testing sets; the temperature data is the highest temperature of the battery during a single charge. Step 2: Initialize the particle population X by setting the particle population size, number of iterations, initial velocity, and position range. Let any particle in particle population X Its position coordinates These are the optimization parameters for the particle swarm optimization. Step 3: Construct an LSTM prediction model using the particle coordinates in particle population X, where... The number of hidden layer units in the LSTM prediction model is given by... The learning rate for the LSTM prediction model; Step 4: Using the root mean square error (RMSE) of the LSTM prediction model as the fitness value for particle swarm optimization, particle swarm optimization is performed. During the optimization process, inert particles are defined, and their decisions are used to iteratively update the position coordinates and velocities of the particle swarm until the fitness value (RMSE) meets the criteria for ending the iteration or the number of iterations reaches its maximum. The optimization parameters are then determined. Optimal value; Step 5, to find the optimal parameters In the optimal value The number of hidden layer units, in The iPSO-LSTM model is constructed and trained to obtain the trained iPSO-LSTM model. Step 6: Obtain the temperature data of the battery to be predicted and input it into the trained iPSO-LSTM model for temperature prediction. The inert particle decision refers to sorting the positions of the particle swarm in descending order of fitness value after each iteration update, defining the bottom 20% of particles as inert particles, resetting the position of the inert particles to the position of the globally optimal particle, but without changing the search speed of the particle, and continuing to seek optimization.

2. The method according to claim 1, characterized in that, In step 2, during the particle swarm initialization process, the particle swarm coordinates and velocities are initialized according to the following formula: in, Represents the initial position coordinates of the i-th particle. The value of the optimization parameter, Represents the initial i-th particle's... The search speed for each optimization parameter The first particle swarm is represented by the second particle swarm. The upper limit of the search speed for each optimization parameter. ( ) are respectively the particle swarm number The upper and lower limits of the optimization parameters, where r is a random number uniformly distributed between [0,1].

3. The method according to claim 2, characterized in that, In step 4, the root mean square error (RMSE) of the LSTM prediction model is represented by the standard deviation of the deviation between the true value and the predicted value: In the formula, n is the number of the highest temperature sample points in the battery temperature data obtained in step 1; This represents the actual value of the j-th sample point; This is the predicted value for the j-th highest temperature sample point.

4. The method according to claim 3, characterized in that, In step 4, during the particle swarm optimization process, inert particles are defined, and the position coordinates and velocities of the particle swarm are iteratively updated using inert particle decisions, including: During particle swarm optimization, the inertia weight w and learning factors c1 and c2 are updated based on fitness and iteration number. Where, the subscript max represents the maximum value of the corresponding parameter, and the subscript min represents the minimum value of the corresponding parameter; t is the cumulative number of iterations, and T is the maximum number of iterations; The particle velocity and coordinates are iteratively updated based on the updated variables w, c1, and c2, as shown in the following expression: in It is the i-th particle in the k-th iteration. The search speed for each optimization parameter It is the i-th particle in the k-th iteration. The values ​​of the optimization parameters; This is the global optimum value in the k-th iteration. Let r1 and r2 be the fitness value of the i-th particle in the k-th iteration, and r1 and r2 be random numbers that are uniformly distributed in the interval [0,1]. After solving for the fitness value in each round of particle swarm optimization, the positions of the particles are sorted in descending order of fitness, and the bottom 20% of the particles are defined as inert particles. Change the position of the inert particle to the position of the globally optimal particle, but do not change the particle's search speed. The expression is as follows: Where y represents that the y-th particle is an inert particle, and k represents the number of iterations. represents the position coordinates of the y-th particle in the k-th iteration; m represents the fitness of the m-th particle being the global optimum in this iteration. This represents the position coordinates of the m-th particle in the k-th iteration.

5. The method according to claim 4, characterized in that, In step 1, when acquiring battery temperature data and constructing training and test sets, the first 90% of the acquired battery temperature data is divided into training data to form the training set, and the remainder is test data to form the test set.

6. The method according to claim 5, characterized in that, The LSTM prediction model includes an input layer, a hidden layer, an output layer, and a Dropout layer. When constructing the LSTM prediction model using particle coordinates in the particle population X, the method further includes setting the learning rate decay rate of the LSTM prediction model to 0.2, setting the number of learning rate decay rounds of the LSTM prediction model to 125, setting the number of layers of the input layer, output layer, and Dropout layer to 1, and setting the maximum number of iterations to 250.

7. The method according to claim 6, characterized in that, The fitness value RMSE tends to stabilize and meets the criteria for ending the iteration, which means: in, This represents the global optimum value in the k-th particle swarm iteration. This represents the global optimum value in the (k-2)th particle swarm iteration.

8. A method for monitoring temperature during the charging process of a power battery, characterized in that, The method uses the method described in any one of claims 1-7 to predict the temperature of the battery during charging, thereby achieving temperature monitoring during the battery charging process.

Images