A method for predicting the state of charge of an energy storage battery

Abstract

This invention discloses a method for predicting the state of charge (SOP) of an energy storage battery, belonging to the field of energy storage battery management technology. Addressing the problems of low prediction accuracy, susceptibility to local optima, and lack of physical constraints in existing methods, this invention first acquires battery operating time-series data and extracts intrinsic mode components using adaptive noise complete ensemble empirical mode decomposition. Then, a multi-scale adaptive temporal convolutional neural network model is constructed. Next, an improved detective behavior algorithm is used to globally optimize the number and size of the model's convolutional kernels. A multi-scale loss function incorporating data fitting, physical boundary constraints, and ampere-hour integral consistency is then constructed to train the optimized model. Finally, the complete data is normalized and input into the trained model for SOP prediction. This invention combines data decomposition, physical prior knowledge, and intelligent optimization algorithms, ensuring the physical feasibility of the prediction results while further reducing prediction errors and improving the model's convergence speed and accuracy.

Description

Technical Field

[0001] This invention relates to the field of energy storage charge state prediction technology, and specifically to a method for predicting the charge state of an energy storage battery. Background Technology

[0002] In recent years, the global transition to clean and low-carbon energy has accelerated. Energy storage systems, as a core support for renewable energy consumption and grid peak and frequency regulation, are increasingly widely used on a large scale. Energy storage batteries, as a core component of energy storage systems, directly determine the safe and efficient operation of the entire system based on their operational stability. According to relevant statistics, approximately 60% of energy storage system failures are related to errors in battery state of charge (SOC) estimation. Accurately determining SOC is not only a prerequisite for safe battery operation but also crucial for improving the utilization efficiency of energy storage systems and reducing operation and maintenance costs. Therefore, promoting the research and development of accurate SOC prediction technology has become an urgent need for the high-quality development of the energy storage industry.

[0003] The state of charge (SOC) of a battery refers to the ratio of its current remaining charge to its rated capacity. It directly reflects the battery's power supply capability and is a core parameter for the battery management system (BMS) to achieve charge / discharge control, lifespan management, and safety warnings. As a dynamic and complex electrochemical system, battery SOC is influenced by a combination of factors, including charge / discharge current, ambient temperature, cycle count, aging level, and resting time. These factors interact and influence each other, resulting in SOC exhibiting strong nonlinear and time-varying characteristics, posing a significant challenge to accurate prediction. According to industry standards, when the SOC estimation error exceeds 5%, it significantly increases the risk of overcharging and over-discharging, leading to accelerated capacity decay, shortened lifespan, and even safety accidents. Therefore, achieving high-precision SOC prediction is crucial.

[0004] Currently, battery SOC prediction still faces many technical bottlenecks. Traditional prediction methods are mainly divided into three categories: traditional statistical methods based on the ampere-hour integration method, model methods based on the battery equivalent circuit model, and estimation methods based on empirical formulas. Among them, the ampere-hour integration method will accumulate errors over long-term use; the equivalent circuit model method is difficult to accurately match the complex electrochemical characteristics of batteries; and the empirical formula method relies on fitting a large amount of experimental data, has poor generalization ability, and cannot adapt to the SOC prediction requirements under different operating conditions and different aging levels. These traditional methods cannot fully take into account the complex influencing factors in the battery operation process and cannot meet the requirements of large-scale energy storage systems for SOC prediction accuracy and real-time performance. Therefore, developing efficient, accurate, and robust SOC prediction methods to overcome existing technical bottlenecks has important theoretical significance and engineering application value. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art by employing adaptive noise complete ensemble empirical mode decomposition (CEEMDAN) for feature decomposition. Secondly, a multi-scale adaptive temporal convolutional neural network (TCN) is constructed, introducing an electrochemical physical constraint mechanism to reconstruct the loss function. Thirdly, an improved detective behavior algorithm is used to globally optimize the initial convolutional kernel parameters of the TCN, overcoming the limitation of single models being prone to local optima. Finally, normalized data is input into the trained model for prediction. The method proposed in this invention combines data decomposition, physical prior knowledge, and intelligent optimization algorithms, thereby improving the prediction accuracy of battery charge state.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is: a method for predicting the state of charge of an energy storage battery, comprising the following steps: Step 1: Obtain battery runtime time series data; Step 2: Perform adaptive noise complete ensemble empirical mode decomposition (CEEMDAN) on the battery running time series data to obtain intrinsic mode components (IMF). Step 3: Build a multi-scale adaptive temporal convolutional neural network (TCN) model; Step 4: Optimize the parameters of the multi-scale adaptive temporal convolutional neural network (TCN) model using the improved detective behavior algorithm. These parameters include the number of convolutional kernels and the kernel size. The improved detective behavior algorithm introduces... The population is initialized using chaotic mapping, an adaptive switching mechanism is used to control the exploration and development phases, the search range is expanded during the exploration phase, Lévy flight is introduced during the development phase, and random operators and reverse learning mechanisms are combined for global optimization. Step 5: Construct a multi-scale loss function that includes a physical regularization term, and train a multi-scale adaptive temporal convolutional neural network (TCN) model using the optimized parameters; Step 6: Normalize the complete battery operating time series data, and input the normalized data into the trained multi-scale adaptive temporal convolutional neural network (TCN) model to predict the state of charge.

[0007] Further optimization involves performing adaptive noise-complete set empirical mode decomposition (EMD) on battery operating time series data to obtain intrinsic mode components (EMCs). The process is as follows: Gaussian white noise with a mean of 0 is added multiple times to the battery operating time series data to construct the sequence to be decomposed; EMD is performed on the sequence after each noise addition to obtain the first EMC and the residual signal; this operation is repeated until the residual signal is a monotonic function, yielding multiple EMCs and the final residual. By repeatedly adding Gaussian white noise with a mean of zero and performing EMD, the mode aliasing problem in traditional mode decomposition is effectively solved, decomposing the original non-stationary signal into multiple EMCs at different frequency scales, thus reducing the non-stationarity of the data.

[0008] Further optimization and construction of a multi-scale adaptive temporal convolutional neural network model are as follows: Step 3.1: Set the number of residual blocks, the number of convolutional layers in each residual block, and the basic parameters of each convolutional layer; Step 3.2: Calculate the center frequency of the intrinsic mode components by adaptive noise complete set empirical mode decomposition, and calculate the dilation factor of the corresponding convolutional layer for dynamic configuration, so that the high frequency components correspond to a small dilation rate to capture details, and the low frequency components correspond to a large dilation rate to capture trends. Step 3.3: Construct residual blocks, select the number of stacked convolutional layers, select activation functions, and set random deactivation and batch normalization; Step 3.4: Set the length of the input and output sequences, the learning rate, and the optimizer.

[0009] By setting up structures such as residual blocks and dilated convolutions, the model can effectively extract long-term and short-term dependency features of time series. Specifically, the model dynamically configures the dilation factor of the corresponding convolutional layer based on the center frequency of each intrinsic mode component obtained from CEEMDAN decomposition: high-frequency components correspond to a small dilation rate to capture detailed features, while low-frequency components correspond to a large dilation rate to capture long-term trends. This multi-scale adaptive mechanism allows the model to fully match the characteristics of different frequency components, improving the targeting and efficiency of feature extraction.

[0010] Further optimization involves using the improved detective behavior algorithm to refine the number and size of convolutional kernels in the TCN model. The specific process is as follows: Step 4 utilizes the improved detective behavior algorithm to optimize the parameters of the multi-scale adaptive temporal convolutional neural network model, specifically including the following sub-steps: Step 4.1: Set the population size and maximum number of iterations; Step 4.2: Use The chaotic mapping initializes the population, with each individual representing a set of parameters for the number and size of convolution kernels; this solves the problem of uneven population distribution caused by random initialization and improves the diversity of initial solutions. Step 4.3: Enter the iterative optimization process, where: The exploration and development phases are controlled by an adaptive switching mechanism; the algorithm is dynamically controlled to enter the exploration or development phase based on the current iteration number and random factors, thus balancing the capabilities of global search and local mining. During the exploration phase, an update formula that expands the search range is used to update the location; Levi Flight was introduced during the development phase for location updates; By incorporating random operators to update the positions of some individuals, the diversity of the population can be further enhanced; A reverse learning mechanism is used to generate reverse individuals. If the reverse individual has better fitness, it replaces the original individual, thereby accelerating the convergence speed. Step 4.4: Determine whether the current iteration has reached the maximum number of iterations. If not, continue iterating; otherwise, output the optimal individual as the optimal value for the number of convolution kernels and the kernel size of the multi-scale adaptive temporal convolutional neural network model.

[0011] Traditional loss functions only consider the fitting error between predicted and true values, which can easily lead to predictions that do not conform to electrochemical laws (such as SOC values ​​exceeding the [0,1] range or violating ampere-hour integral consistency). Therefore, the loss function in step 5, which includes a physical regularization term, comprises a data fitting term, a physical boundary constraint term, and an ampere-hour integral consistency term. The physical boundary constraint term constrains the predicted charge state values ​​to be between 0 and 1; the ampere-hour integral consistency term constrains the predicted charge state values ​​at adjacent time points to conform to the ampere-hour integral relationship. By introducing these physical priors, the model output strictly follows electrochemical laws, fundamentally eliminating non-physical explanations and reducing prediction errors.

[0012] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the energy storage battery charge state prediction method.

[0013] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the energy storage battery charge state prediction method.

[0014] This invention uses CEEMDAN to decompose the original data, reducing data stationarity and thus minimizing its impact on prediction. An improved detective behavior algorithm optimizes the number and size of convolutional kernels in the TCN model, overcoming the problem of traditional gradient descent easily getting trapped in local optima. Simultaneously, by setting a dynamic inflation factor within the TCN, it adaptively matches the multi-scale features of the CEEMDAN decomposition. Furthermore, ampere-hour integral consistency constraints and boundary constraints are introduced into the loss function, ensuring that the model corresponding to the optimal parameters searched by the detective behavior algorithm strictly follows electrochemical laws in its output, fundamentally eliminating non-physical interpretations. This multi-faceted approach produces a synergistic effect, significantly improving prediction accuracy compared to using the algorithm alone for parameter tuning. Attached Figure Description

[0015] Figure 1 This is a flowchart of the energy storage battery charge state prediction method described in this invention; Figure 2 This is a flowchart of the improved detective behavior algorithm.

[0016] Figure 3 Comparison chart showing improvements to the detective behavior algorithm.

[0017] Figure 4 This is a comparison chart of SOC prediction results. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0019] Reference Figure 1 A method for predicting the state of charge of an energy storage battery includes the following steps: Step 1: Obtain battery runtime time series data; Specifically, the battery management system (BMS) collects parameters such as terminal voltage, charging and discharging current, and surface temperature during battery operation, forming battery operation time series data, denoted as x(t), at a fixed sampling frequency (e.g., 1Hz). In this embodiment, the acquisition duration is a complete charge and discharge cycle, and the data length includes at least one complete operating cycle.

[0020] Step 2: Perform adaptive noise complete set empirical mode decomposition (CEEMDAN) on the battery operating time series data to obtain intrinsic mode components (IMF). Step 2.1: Add K times Gaussian white noise with a mean of 0 to the battery running time series data x(t) to construct the decomposition sequence x for N experiments. i (t), where i = 1, 2, ..., N: ; In the formula, xi (t) is the noisy sequence obtained in the i-th experiment. B represents the weighting coefficients of Gaussian white noise. i (t) represents the white noise sequence added for the i-th time; Step 2.2: For each x i (t) Perform Empirical Mode Decomposition (EMD) to obtain the first intrinsic mode component. and residual signal The formula is: ; ; In the formula, This represents the i-th eigenmode component obtained in the first-order decomposition. This represents the residual signal obtained in the first-order decomposition; Step 2.3: Determine the residual signal (t) Is it a monotonic function? If not, repeat steps 2.1-2.2 to extract the 2nd, 3rd... IMF components in sequence until the residual signal is a monotonic function. Then the CEEMDAN decomposition process ends, and N intrinsic mode components are obtained. The residual signal obtained from the Nth order decomposition ; ; in For the i-th intrinsic mode component, by adding Gaussian white noise with zero mean multiple times and performing EMD decomposition, the mode mixing problem is effectively avoided, and the original non-stationary signal is decomposed into multiple IMF components with different frequency scales, thus reducing the non-stationarity of the data.

[0021] Step 3: Build a multi-scale adaptive temporal convolutional neural network (TCN) model; Step 3.1: Set the number of residual blocks (3 in this embodiment), the number of convolutional layers in each residual block (2 layers per block), and the basic parameters of each convolutional layer (kernel size range [2,8], stride of 1, padding method is causal padding to ensure temporal causality).

[0022] Step 3.2: Calculate the center frequencies of each intrinsic mode component obtained from the CEEMDAN decomposition. Calculate the dilation factor of the corresponding convolutional layer for dynamic configuration. This allows high-frequency components to correspond to small expansion rates to capture details, while low-frequency components correspond to large expansion rates to capture trends. ; ; In the formula, Let i be the center frequency of the i-th intrinsic mode component. As the reference frequency, The maximum frequency of all intrinsic modal components. The minimum frequency of all intrinsic modal components. This represents the rounding function; this mechanism enables the multi-scale adaptive temporal convolutional neural network (TCN) model to adaptively match the frequency characteristics of different intrinsic mode components, thereby improving the multi-scale feature extraction capability.

[0023] Step 3.3: Construct residual blocks. Each residual block contains two dilated convolutional layers with ReLU activation function. After each convolutional layer, a dropout layer (with a dropout rate of 0.2) and a batch normalization layer are added to accelerate convergence and prevent overfitting.

[0024] Step 3.4: Set the input sequence length (100 time steps in this embodiment), the output sequence length (1 time step, i.e., single-step prediction), initialize the learning rate to 0.001, and use Adam as the optimizer.

[0025] Step 4: Optimize the number and size of convolutional kernels in the multi-scale adaptive temporal convolutional neural network (TCN) model using the improved detective behavior algorithm; for example... Figure 2 As shown, the flowchart of the improved detective behavior algorithm illustrates the collaborative working mechanism of modules such as chaotic initialization, adaptive switching, Levy flight, random operators, and reverse learning.

[0026] Step 4.1: Set the population size to 30 and the maximum number of iterations to 200.

[0027] Step 4.2: Introduce Logistic-Tent chaotic mapping to initialize the population to solve the problem of uneven distribution caused by random initialization.

[0028] ; ; ; In the formula, Represents the initial population. For the i-th individual in the j-th dimension Individuals after chaotic mapping Represents the i-th individual in the j-th dimension Individuals before chaotic mapping, i∈1,2,…,n;j∈1,2,…,d;n is the population size,d is the dimension of the problem;r is a random number between 0 and 4, each individual represents a set of TCN model convolution kernel number and convolution kernel size parameters,ub andlb are the upper and lower bounds of the problem respectively, andRand represents a random number between 0 and 1.

[0029] Step 4.3: Enter the iterative optimization process: 4.3.1 Switching Mechanism Between Exploration and Development To better control the transition between exploration and development modes, the original formula was improved.

[0030] ; In the formula, This represents the current iteration number. This represents the maximum number of iterations. Rand is a randomly generated value between 0 and 1. This is a constant used to control the balance between exploration and utilization; in this embodiment, it is set to 0.5. A threshold value of 0.5 is used to maintain a balance between the two strategies. Specifically, when f(it) is greater than 0.5, candidate solutions tend to be utilized more; while when f(it) is less than or equal to 0.5, exploration is preferred. This mechanism balances the capabilities of global search and local mining.

[0031] 4.3.2 Exploration Phase To address the issue of insufficient exploration capability in the original algorithm's exploration phase, the algorithm was improved in terms of formula, enabling it to have a wider exploration range. ; ; ; In the formula, This represents the globally optimal position obtained from the search history. This represents the j-th dimension position of the i-th individual in the (t+1)-th iteration. This represents the position of the j-th dimension of the i-th individual in the t-th iteration. This represents the upper boundary of the individual at the (t-1)th iteration. This represents the lower boundary of the individual at the (t-1)th iteration. To search for the center point of an individual, This represents the fitness value of an individual before the update. The fitness value is the optimal value for the individual, and r1 and r2 are random numbers between 0 and 1. The product of two random numbers enables controlled shrinkage mapping to high-quality regions (exploration phase) while maintaining global feasibility.

[0032] 4.3.3 Development Phase ; In the formula, It is the stride length of Levy's flight. It is a constant between 0 and 1.

[0033] 4.3.4 Random Operators ; ; In the formula, and These are two individuals randomly selected from the population. This represents the j-th dimension position of the i-th individual in the t-th iteration. It is a d-dimensional binary vector, where each element has a 50% probability of taking the value 0 or 1. It is a set of random vectors of dimension d, where each dimension is generated by uniformly distributing the values ​​between 0 and 1. ° represents the Hadamard product operation. As weight, and Represent and The optimal value, It is a constant value.

[0034] 4.3.5 Reverse Learning Mechanism ; ; In the formula, For the i-th reverse individual, For the i-th individual, Let be the fitness value of the i-th reverse individual. Let be the fitness value of the i-th individual.

[0035] 4.4 Determine if the current iteration has reached the maximum number of iterations. If not, continue iterating; otherwise, stop iterating and output the optimal individual, which is the optimal value of the number of convolution kernels and the kernel size of the TCN model.

[0036] Step 5: Construct a multi-scale loss function that includes a physical regularization term, and train the TCN model using the optimized parameters; 5.1 Constructing a multi-scale loss function ; ; In the formula, For multi-scale loss functions, For data fitting terms, For physical boundary constraints, For ampere-hour integral consistency term, and These are constant coefficients.

[0037] 5.2 Data Fitting Term ; In the formula, For the sample size, Let SOC be the value predicted by the TCN model at time t. This corresponds to the actual SOC value.

[0038] 5.3 Physical Boundary Constraints ; 5.4 Consistency term for ampere-hour integral ; In the formula, For the predicted SOC at the next moment, For the predicted SOC at the current moment, For sampling current, The sampling time interval, For the battery's rated capacity, is the Coulomb efficiency coefficient.

[0039] During training, the TCN model is initialized with the optimal convolutional kernel parameters obtained in step 4. The IMF components obtained in step 2 are normalized and used as inputs, and the true SOC is used as the output. The Adam optimizer is used to minimize the loss function. The number of training epochs is set to 100, and the batch size is set to 32.

[0040] Step 6: Normalize the complete battery operating time series data, and input the normalized data into the trained TCN model for charge state prediction.

[0041] During prediction, the newly acquired battery operating time series data is first normalized using the same methods as the training data (using the mean and standard deviation of the training set). Then, the normalized data is directly input into the trained TCN model, and the model output is the predicted charge state value. It should be noted that this embodiment does not perform CEEMDAN decomposition again during the prediction stage; instead, the original normalized sequence is input as a whole. This is because the TCN model has already learned the feature mappings of the IMF components during training, and the original sequence can be directly input for real-time prediction.

[0042] This embodiment uses publicly available battery datasets (such as the NASA or CALCE lithium-ion battery dataset) for validation. The method of this invention is compared with single TCN, CEEMDAN+TCN, and TCN optimized with the unimproved detective behavior algorithm (i.e., TCN+DBA). Prediction errors are evaluated using root mean square error (RMSE) and mean absolute error (MAE). Table 1 shows the error comparison results for each method: the single TCN model has an RMSE of 3.24% and an MAE of 2.51%; the CEEMDAN+TCN method reduces the error to an RMSE of 2.17% and an MAE of 1.86%; the method combining TCN with the unimproved detective behavior algorithm (DBA) further optimizes the RMSE to 2.03% and the MAE to 1.72%. The method proposed in this invention (integrating CEEMDAN decomposition, multi-scale adaptive TCN, improved detective behavior algorithm optimization, and physical constraint loss function) achieves the best prediction accuracy, with an RMSE as low as 1.08% and a MAE of only 0.89%. Compared with the TCN+DBA method, the MAE is reduced by about 48% and the RMSE is reduced by about 47%. Figure 3 A comparison of the convergence curves before and after the improvement of the detective behavior algorithm is given. It can be seen that the improved algorithm proposed in this invention converges to a better solution after about 80 iterations, while the original algorithm requires more than 150 iterations and has a poor final fitness. Figure 4 The comparison of prediction results shows that the SOC prediction curve of the method of this invention highly coincides with the actual curve, while the error of the comparative method increases significantly at the end of the discharge. The above experimental data fully verify the superiority of this invention in SOC prediction.

[0043] Table 1

[0044] Another embodiment of the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the energy storage battery charge state prediction method.

[0045] Another embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the energy storage battery charge state prediction method.

[0046] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the state of charge of an energy storage battery, characterized in that, Includes the following steps: Step 1: Obtain battery runtime time series data; Step 2: Perform adaptive noise complete set empirical mode decomposition on the battery operating time series data to obtain intrinsic mode components; Step 3: Build a multi-scale adaptive temporal convolutional neural network model; Step 4: Optimize the parameters of the multi-scale adaptive temporal convolutional neural network model using the improved detective behavior algorithm. These parameters include the number of convolutional kernels and the kernel size. The improved detective behavior algorithm introduces... The population is initialized using chaotic mapping, an adaptive switching mechanism is used to control the exploration and development phases, the search range is expanded during the exploration phase, Lévy flight is introduced during the development phase, and random operators and reverse learning mechanisms are combined for global optimization. Step 5: Construct a multi-scale loss function that includes a physical regularization term, and train a multi-scale adaptive temporal convolutional neural network model using the optimized parameters; Step 6: Normalize the complete battery operating time series data, and input the normalized data into the trained multi-scale adaptive temporal convolutional neural network model to predict the state of charge.

2. The method for predicting the state of charge of an energy storage battery according to claim 1, characterized in that, The process of obtaining intrinsic mode components by adaptive noise complete set empirical mode decomposition of battery running time series data is as follows: Gaussian white noise with a mean of 0 is added multiple times to the battery running time series data to construct the sequence to be decomposed. Empirical mode decomposition is performed on the sequence after each addition of noise to obtain the first intrinsic mode component and the residual signal; the above operation is repeated until the residual signal is a monotonic function to obtain multiple intrinsic mode components and the final residual.

3. The method for predicting the state of charge of an energy storage battery according to claim 1, characterized in that, The specific steps for building a multi-scale adaptive temporal convolutional neural network model are as follows: Step 3.1: Set the number of residual blocks, the number of convolutional layers in each residual block, and the basic parameters of each convolutional layer; Step 3.2: Calculate the center frequencies of the intrinsic mode components by calculating the adaptive noise complete set empirical mode decomposition. Calculate the dilation factor of the corresponding convolutional layer for dynamic configuration. This allows high-frequency components to correspond to small expansion rates to capture details, while low-frequency components correspond to large expansion rates to capture trends. ； In the formula, Let i be the center frequency of the i-th intrinsic mode component. As the reference frequency, The maximum frequency of all intrinsic modal components. The minimum frequency of all intrinsic modal components. Represents the floor function; Step 3.3: Construct residual blocks, select the number of stacked convolutional layers, select activation functions, and set random deactivation and batch normalization; Step 3.4: Set the length of the input and output sequences, the learning rate, and the optimizer.

4. The method for predicting the state of charge of an energy storage battery according to claim 1, characterized in that, The improved detective behavior algorithm is used to optimize the number and size of convolutional kernels in the TCN model. The specific process is as follows: Step 4 utilizes the improved detective behavior algorithm to optimize the parameters of the multi-scale adaptive temporal convolutional neural network model, specifically including the following sub-steps: Step 4.1: Set the population size and maximum number of iterations; Step 4.2: Initialize the population using Logistic-Tent chaotic mapping, where each individual represents a set of parameters for the number and size of convolutional kernels; Step 4.3: Enter the iterative optimization process, where: According to the switching function The exploration and development phases are controlled. When f(it) > 0.5, the development phase is entered; otherwise, the exploration phase is entered. It is the current iteration number, MaxIt is the maximum iteration number, Rand is a random number in [0,1], α is a constant, and e is the natural constant. During the exploration phase, an update formula that expands the search range is used to update the location; Levi Flight was introduced during the development phase for location updates; The positions of some individuals are updated by incorporating random operators; A reverse learning mechanism is used to generate a reverse individual; if the reverse individual has better fitness, it replaces the original individual. Step 4.4: Determine whether the current iteration has reached the maximum number of iterations. If not, continue iterating; otherwise, output the optimal individual as the optimal value for the number of convolution kernels and the kernel size of the multi-scale adaptive temporal convolutional neural network model.

5. The method for predicting the state of charge of an energy storage battery according to claim 1, characterized in that, The loss function containing the physical regularization term in step 5 includes a data fitting term, a physical boundary constraint term, and an ampere-hour integral consistency term; the physical boundary constraint term is used to constrain the predicted charge state value to be between 0 and 1; the ampere-hour integral consistency term is used to constrain the predicted charge state value at adjacent time points to conform to the ampere-hour integral relationship.

6. The method for predicting the state of charge of an energy storage battery according to claim 1, characterized in that, The formula for the reverse learning mechanism described in step 4 is: ； ； In the formula, For the i-th reverse individual, For the i-th individual, Let be the fitness value of the i-th reverse individual. Let be the fitness value of the i-th individual, and ub and lb be the upper and lower bounds of the search space, respectively. This is the fitness function.

7. The method for predicting the state of charge of an energy storage battery according to claim 4, characterized in that, The multiscale loss function including the physical regularization term is as follows: ； ； In the formula, For multi-scale loss functions, For data fitting terms, For physical boundary constraints, For ampere-hour integral consistency term, and The coefficients are constants. , For the sample size, Let SOC be the value predicted by the TCN model at time t. This corresponds to the actual SOC value; , In the formula, The predicted SOC value for the next time step. The predicted SOC value at the current moment. For sampling current, The sampling time interval, For the battery's rated capacity, is the Coulomb efficiency coefficient.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the energy storage battery charge state prediction method as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the energy storage battery charge state prediction method as described in any one of claims 1 to 7.

Images