Energy storage battery charge state prediction model training and prediction method, device and equipment

Abstract

The invention relates to the technical field of electric energy storage, in particular to a method, a device and equipment for training and predicting a state of charge prediction model of an energy storage battery. The method comprises the following steps: constructing a data set by acquiring current, voltage, temperature and battery state-of-charge values of an energy storage battery at continuous moments; inputting samples in the sample pair into a preset first to-be-trained model for multi-scale feature extraction to obtain multi-scale output features at the current moment; inputting the multi-scale output feature at the current moment into a preset second to-be-trained model for prediction processing to obtain a prediction output result at the current moment; constructing a quantile loss function according to the prediction output result at the current moment, the label corresponding to the sample and a preset quantile regression model; the first to-be-trained model and the second to-be-trained model are adjusted according to the quantile loss function corresponding to the sample pair at each moment, a trained energy storage battery charge state prediction model is obtained, and the operation safety and management efficiency of an energy storage system are improved.

Description

Technical Field

[0001] This invention relates to the field of power energy storage technology, specifically to a method, apparatus, and equipment for training and predicting the state of charge prediction model of energy storage batteries. Background Technology

[0002] Lithium-ion batteries are indispensable for power storage, and battery management systems (BMS) are essential in their operation. BMS monitors battery parameters such as voltage, current, temperature, and state of charge (SOC) in real time, managing and controlling the charging and discharging process to prevent overcharging, over-discharging, and thermal runaway, ensuring the long-term safe and stable operation of lithium-ion batteries. SOC estimation is a crucial component of BMS; however, various factors affect SOC estimation during actual battery operation, leading to uncertainty. Traditional point estimation methods can only provide a single SOC value, making high-precision SOC estimation a significant challenge. Summary of the Invention

[0003] The purpose of this invention is to provide a training and prediction method, apparatus and equipment for predicting the state of charge of energy storage batteries, which solves the problems of the prior art.

[0004] This invention is achieved through the following technical solution:

[0005] In a first aspect, embodiments of the present invention provide a method for training a prediction model for the state of charge of an energy storage battery, comprising:

[0006] Acquire the current, voltage, temperature and state of charge (SOC) values ​​of the energy storage battery at continuous time points, and construct a dataset. Each dataset includes multiple sample pairs corresponding to multiple time points. Each sample pair includes a sample and a label. The sample includes current, voltage and temperature, and the label is the SOC value.

[0007] For the sample pair at the current time, the samples in the sample pair are input into a preset first training model for multi-scale feature extraction to obtain the multi-scale output features at the current time. The first training model includes multiple convolutional kernels and nonlinear activation functions.

[0008] The multi-scale output features at the current moment are input into a preset second training model for prediction processing to obtain the prediction output result at the current moment.

[0009] A quantile loss function is constructed based on the predicted output at the current moment, the label corresponding to the sample, and the preset quantile regression model.

[0010] The first and second models to be trained are adjusted based on the quantile loss function corresponding to the sample pairs at each time step to obtain a trained energy storage battery state of charge prediction model.

[0011] Preferably, the step of inputting samples from the current sample pair into a preset first training model for multi-scale feature extraction to obtain the multi-scale output features at the current time includes:

[0012] The samples in the sample pair are input into the first convolution kernel to obtain the first convolution output, wherein the size of the first convolution kernel is 1×1;

[0013] The samples in the sample pair are input into the second convolution kernel to obtain the second convolution output, wherein the size of the second convolution kernel is 3×3;

[0014] The samples in the sample pair are input into the third convolution kernel to obtain the third convolution output, wherein the size of the third convolution kernel is 5×5;

[0015] The first convolutional output, the second convolutional output, and the third convolutional output are concatenated and fused to obtain multi-scale fused features;

[0016] Substituting the multi-scale fused features into a nonlinear activation function yields the multi-scale output features at the current time.

[0017] Preferably, the first convolution output satisfies:

[0018] ,

[0019] in, For the input sample, This is the kernel weight vector of the first convolutional kernel. is the bias vector of the first convolution kernel.

[0020] Preferably, the second convolution output satisfies:

[0021] ,

[0022] in, These are the intermediate features after the first dimensionality reduction and the ReLU activation function. This is the convolution kernel weight vector corresponding to the first dimension reduction. This is the bias vector corresponding to the first dimension reduction. This is the kernel weight vector of the second convolution kernel. is the bias vector of the second convolution kernel.

[0023] Preferably, the third convolution output satisfies:

[0024] ,

[0025] in, These are the intermediate features after the second dimensionality reduction and the ReLU activation function. This is the convolution kernel weight vector corresponding to the second dimension reduction. This is the bias vector corresponding to the second dimension reduction. The kernel weight vector of the third convolution kernel. is the bias vector of the third convolution kernel.

[0026] Preferably, the step of inputting the multi-scale output features at the current moment into a preset second training model for prediction processing to obtain the prediction output result at the current moment satisfies:

[0027] ,

[0028] ,

[0029] ,

[0030] ,

[0031] in, For the current time, for Multi-scale output features. for Cellular input, for The input gate, for The Gate of Oblivion for The output gate, The tanh activation function is used. for Input pre-activation value, The input weight matrix is ​​used to... Weight parameters projected onto the candidate state space. For the input circular weight matrix, For the input bias vector, for The input gate pre-activation value, Input gate weight matrix, The input gate loop weight matrix is... The input gate bias vector; for The pre-activation value of the forget gate, Forget gate weight matrix The forget gate loop weight matrix, This is the forget gate bias vector; It is the sigmoid activation function. for The output gate pre-activation value, Output gate weight matrix, This is the output gate loop weight matrix. This is the output gate bias vector. for The cell state controls the proportion of new inputs. for The normalized state is used to stabilize the gated output. for The predicted output results.

[0032] Preferably, the quantile loss function constructed based on the prediction output at the current moment, the label corresponding to the sample, and the preset quantile regression model satisfies:

[0033] ,

[0034] ,

[0035] ,

[0036] in, It is a sloping absolute value function. for The corresponding response variable, for The regression coefficient vector derived from the quantiles , The label is the label of the sample pair at the current moment.

[0037] Secondly, embodiments of the present invention provide a method for predicting the state of charge of an energy storage battery, including:

[0038] Acquire current, voltage, and temperature over a period of time;

[0039] The current, voltage, and temperature over a given period of time are input into the energy storage battery state of charge prediction model to obtain the battery state of charge prediction value and confidence interval output by the energy storage battery state of charge prediction model, wherein the energy storage battery state of charge prediction model is trained according to the method of the first aspect.

[0040] Thirdly, embodiments of the present invention provide a training device for a prediction model of the state of charge of an energy storage battery, comprising:

[0041] The dataset module is used to obtain the current, voltage, temperature and state of charge values ​​of the energy storage battery at continuous time points and construct a dataset. Each dataset includes multiple sample pairs corresponding to multiple time points. Each sample pair includes a sample and a label. The sample includes current, voltage and temperature, and the label is the state of charge value of the battery.

[0042] The feature extraction module is used to input the samples in the sample pair at the current time into a preset first training model for multi-scale feature extraction to obtain the multi-scale output features at the current time. The first training model includes multiple convolutional kernels and nonlinear activation functions.

[0043] The prediction module is used to input the multi-scale output features of the current time into a preset second training model for prediction processing, and obtain the prediction output result of the current time.

[0044] The loss function module is used to construct a quantile loss function based on the prediction output at the current time, the label corresponding to the sample, and a preset quantile regression model.

[0045] The adjustment module is used to adjust the first and second models to be trained based on the quantile loss function corresponding to the samples at each time step, so as to obtain a trained energy storage battery state of charge prediction model.

[0046] Fourthly, embodiments of the present invention provide an electronic device, including: at least one processor, at least one memory, and computer program instructions stored in the memory, which, when executed by the processor, implement the method of the first aspect described above.

[0047] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0048] By employing a hybrid architecture combining multi-kernel convolutional neural networks and scalar long short-term memory neural networks, and introducing a quantile regression loss function for model training, point and interval estimations of battery state of charge (SOC) are achieved. This architecture effectively extracts multi-scale time-series features from battery operating current, voltage, and temperature data, providing both accurate SOC estimates and confidence intervals characterizing prediction uncertainties. This combined point and interval estimation capability provides battery management systems with more comprehensive state information, enabling more reliable operational decisions under complex conditions and thus improving the safety and management efficiency of energy storage systems. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0050] Figure 1 This is a flowchart illustrating the training method for the energy storage battery state of charge prediction model provided in this embodiment.

[0051] Figure 2 This is a schematic diagram of the structure of the energy storage battery state of charge prediction model training device provided in this embodiment;

[0052] Figure 3 This is a schematic diagram of the structure of the electronic device provided in this embodiment. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0054] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0055] It should be noted that all actions involving the acquisition of signals, information, or data in this invention are carried out in compliance with the relevant data protection laws and regulations of the locality and with authorization from the owner of the relevant device.

[0056] Example 1

[0057] Please see Figure 1 This invention provides a method for training a prediction model of the state of charge of an energy storage battery, comprising:

[0058] S1. Obtain the current, voltage, temperature and state of charge values ​​of the energy storage battery at continuous time points, and construct a dataset. Each dataset includes multiple sample pairs corresponding to multiple time points. Each sample pair includes a sample and a label. The sample includes current, voltage and temperature, and the label is the state of charge value of the battery.

[0059] Specifically, building the dataset provides a sufficient and structured input-output correspondence for subsequent model training, ensuring that the model can learn the dynamic characteristics of the battery from actual operating data. Time-series data is collected through battery testing platforms or actual operating systems and organized into sample-label pairs according to time steps. For example, the current, voltage, and temperature at a certain moment are used as inputs, and the measured SOC value at the corresponding moment is used as a supervision signal. This builds the data foundation required for supervised learning, thus laying the data foundation for subsequent feature extraction and model training, and effectively supporting the model's generalization ability under actual operating conditions.

[0060] S2. For the sample pair at the current time, input the samples in the sample pair into the preset first training model to perform multi-scale feature extraction to obtain the multi-scale output features at the current time. The first training model includes multiple convolutional kernels and nonlinear activation functions.

[0061] Specifically, the first model to be trained can refer to a multi-kernel convolutional neural network, which is a convolutional structure with multiple parallel convolutional paths. Each path uses convolutional kernels of different sizes, such as 1×1, 3×3, and 5×5, to capture local details and global contextual information in the input signal, respectively. The motivation for adopting this structure is to overcome the problem that a single convolutional kernel has a limited receptive field and is difficult to fully extract the dynamic features of the battery. By setting up multiple convolutional branches in parallel, the input current, voltage, and temperature data are processed by convolution, and then the output features of each branch are fused. Then, a comprehensive multi-scale feature is generated through a nonlinear activation function, thereby preserving both high-frequency details and low-frequency trends at the feature level, enhancing the model's ability to represent the complex dynamics of the battery, and providing richer feature inputs for subsequent sequence modeling.

[0062] S3. Input the multi-scale output features at the current time into a preset second training model for prediction processing to obtain the prediction output result at the current time.

[0063] Specifically, the second model to be trained can be a long short-term memory neural network, which uses memory units and gating structures to model the dynamic process of battery SOC changes over time. Multi-scale features are used as sequence inputs, and information flow is regulated through input gates, forget gates, and output gates. Combined with cell state to maintain long-term information, this enables the prediction of the current SOC value. This process effectively captures the state evolution of the battery under continuous operating conditions, improving the model's accuracy in modeling temporal dynamics.

[0064] In some implementations, S3 involves inputting the multi-scale output features of the current moment into a preset second training model for prediction processing to obtain the prediction output result of the current moment, satisfying the following:

[0065] ,

[0066] ,

[0067] ,

[0068] ,

[0069] in, For the current time, for Multi-scale output features. for Cellular input, for The input gate, for The Gate of Oblivion for The output gate, The tanh activation function is used. for Input pre-activation value, The input weight matrix is ​​used to... Weight parameters projected onto the candidate state space. For the input circular weight matrix, This is the input bias vector. for The input gate pre-activation value, Input gate weight matrix, The input gate loop weight matrix is... The input gate bias vector; for The pre-activation value of the forget gate, Forget gate weight matrix The forget gate loop weight matrix, This is the forget gate bias vector; It is the sigmoid activation function. for The output gate pre-activation value, Output gate weight matrix, This is the output gate loop weight matrix. This is the output gate bias vector. for The cell state controls the proportion of new inputs. for The normalized state is used to stabilize the gated output. for The predicted output results.

[0070] S4. Construct a quantile loss function based on the prediction output at the current moment, the label corresponding to the sample, and the preset quantile regression model;

[0071] Specifically, quantile regression is a statistical method for estimating conditional quantiles, reflecting the range of change of the response variable at different quantiles. Its loss function weights the prediction errors at different quantiles using a skewed absolute value function. Quantile regression extends point estimation to interval estimation, thereby quantifying prediction uncertainty. By setting multiple quantile parameters, the quantile loss between the predicted value and the true label at each quantile is calculated, thus constructing a composite loss function that simultaneously optimizes multiple quantile outputs. This allows the model to not only output point estimates but also provide confidence intervals characterizing uncertainty, enhancing the battery management system's ability to assess abnormal operating conditions and prediction risks.

[0072] In some implementations, S4 involves constructing a quantile loss function based on the predicted output at the current moment, the label corresponding to the sample, and a preset quantile regression model, satisfying the following:

[0073] ,

[0074] ,

[0075] ,

[0076] in, It is a sloping absolute value function. for The corresponding response variable, for The regression coefficient vector derived from the quantiles , The label is the label of the sample pair at the current moment.

[0077] Specifically, the traditional mean squared error loss is replaced by a quantile regression loss function. This function enables the model to learn different locational information in the data distribution by simultaneously optimizing the output corresponding to multiple preset quantiles (e.g., τ=0.1, 0.5, 0.9). For example, when... The median value (point estimate) is output at that time, while 0.1 and 0.9 will output predicted values ​​reflecting the lower and upper bounds of the distribution, respectively. In practice, the model outputs the estimated values ​​at the points (…). While generating a single predicted value for SOC, the system also continuously generates a confidence interval for quantifiable uncertainty. This allows the battery management system to not only obtain a single predicted value for SOC, but also to assess the reliability of the prediction results through the range of intervals, thereby adopting a more conservative safety strategy under extreme operating conditions.

[0078] For example, when When the quantile regression is performed, the output is as follows:

[0079] .

[0080] S5. Adjust the first and second models to be trained according to the quantile loss function corresponding to the samples at each time step to obtain the trained energy storage battery state of charge prediction model.

[0081] Specifically, this step iteratively optimizes the network parameters using the backpropagation algorithm, gradually converging the quantile loss function. Gradient descent algorithms can be employed, updating the weights and biases of each layer's parameters based on the gradient of the loss function. Through multiple rounds of training, the model gradually adapts to the data distribution, ultimately resulting in a hybrid neural network model that can simultaneously output SOC point estimates and confidence intervals. This model, while ensuring the accuracy of point estimation, further provides uncertainty information, contributing to safer decision-making and control in practical battery management systems.

[0082] In some implementations, S2, for the sample pair at the current time, inputs the samples in the sample pair into a preset first training model for multi-scale feature extraction to obtain the multi-scale output features at the current time, including:

[0083] S21. Input the samples in the sample pair into the first convolution kernel to obtain the first convolution output, wherein the size of the first convolution kernel is 1×1;

[0084] Specifically, the 1×1 convolution kernel is mainly used to perform linear combination and dimensionality transformation of the input features. It can fuse and compress information between channels while maintaining the spatial dimension. This step is used to build the basic path for feature extraction. By performing linear projection of the original current, voltage, and temperature data at the channel level, the initial feature transformation is achieved, providing basic data with a unified dimension for the subsequent fusion of features at different scales. This allows for the initial integration of features with relatively low computational overhead.

[0085] S22. Input the samples in the sample pair into the second convolution kernel to obtain the second convolution output, wherein the size of the second convolution kernel is 3×3;

[0086] Specifically, the 3×3 convolutional kernel has a moderate receptive field, enabling it to capture local correlation patterns between adjacent data points in the input sequence. This path is designed to extract local, detailed features of the signal, such as instantaneous voltage fluctuations or short-term temperature trends. By sliding across the input data, this convolutional operation focuses on information interactions within local neighborhoods, thereby capturing features with greater spatial contextuality than 1×1 convolutions, enhancing the model's ability to perceive short-term changes in battery dynamics.

[0087] S23. Input the samples in the sample pair into the third convolution kernel to obtain the third convolution output, wherein the size of the third convolution kernel is 5×5;

[0088] Specifically, the 5×5 convolutional kernel has a larger receptive field, enabling it to perceive broader contextual information and long-term trends in the input data. This branch is introduced to capture slower-changing global features in the signal, such as continuous patterns of current change or overall temperature drift. The larger kernel size allows it to cover longer data segments, thus integrating information over a longer time span, which helps the model understand the evolution of battery state at a macroscopic scale.

[0089] S24. The first convolutional output, the second convolutional output, and the third convolutional output are concatenated and fused to obtain multi-scale fused features;

[0090] Specifically, feature concatenation is an operation that connects feature tensors from different sources along the channel dimension. This fusion operation is performed to integrate features extracted from 1×1, 3×3, and 5×5 convolutional paths at different scales. By concatenating along the channel dimension, the fused features simultaneously contain compressed channel information, local details, and global context, forming a more comprehensive feature representation with a richer information hierarchy, providing comprehensive input information for subsequent sequence modeling modules.

[0091] S25. Substitute the multi-scale fusion features into the nonlinear activation function to obtain the multi-scale output features at the current time.

[0092] Specifically, a nonlinear activation function is a mathematical function that introduces a nonlinear transformation into a neural network, such as ReLU. Applying activation functions aims to perform a nonlinear mapping on the fused linear features, enhancing the model's expressive power. This process allows complex patterns in the fused features, such as the nonlinear coupling relationships between battery parameters, to be better learned and represented, thereby outputting more discriminative multi-scale features and improving the adaptability of subsequent SOC prediction models to complex operating conditions.

[0093] ,

[0094] in The outputs of the merged convolutions are as follows: This represents the final output feature of the first model to be trained.

[0095] In some implementations, the first convolution output satisfies:

[0096] ,

[0097] in, For the input sample, This is the kernel weight vector of the first convolutional kernel. is the bias vector of the first convolution kernel.

[0098] The second convolution output satisfies:

[0099] ,

[0100] in, These are the intermediate features after the first dimensionality reduction and the ReLU activation function. This is the convolution kernel weight vector corresponding to the first dimension reduction. This is the bias vector corresponding to the first dimension reduction. This is the kernel weight vector of the second convolution kernel. is the bias vector of the second convolution kernel.

[0101] The third convolution output satisfies:

[0102] ,

[0103] in, These are the intermediate features after the second dimensionality reduction and the ReLU activation function. This is the convolution kernel weight vector corresponding to the second dimension reduction. This is the bias vector corresponding to the second dimension reduction. The kernel weight vector of the third convolution kernel. is the bias vector of the third convolution kernel.

[0104] In some embodiments, this embodiment also provides a method for predicting the state of charge of an energy storage battery, including:

[0105] Acquire current, voltage, and temperature over a period of time;

[0106] The current, voltage, and temperature over a given period of time are input into the energy storage battery state of charge prediction model to obtain the predicted state of charge value and confidence interval of the battery output by the energy storage battery state of charge prediction model. The energy storage battery state of charge prediction model is trained according to the method described in this embodiment.

[0107] Specifically, the system acquires a sequence of current, voltage, and temperature data generated by the energy storage battery during continuous operation. This time-series data sequence is then input into a pre-trained energy storage battery state-of-charge (SOC) prediction model, which outputs the predicted SOC value and confidence interval for the corresponding time point. The SOC prediction model is a hybrid neural network model trained using the method described in the preceding embodiments. It extracts multi-scale features from the input data through a multi-convolutional kernel structure and utilizes a scalar long short-term memory network with quantile regression capabilities for sequence modeling. When processing the input time-series data, this model can simultaneously output point estimates of the battery SOC and their corresponding confidence intervals, thereby providing state estimation while quantifying the uncertainty of the prediction, offering more comprehensive information support for the battery management system's decision-making.

[0108] Example 2

[0109] Please see Figure 2 This invention provides a training device for a battery state of charge prediction model, comprising:

[0110] Data set module 201 is used to obtain the current, voltage, temperature and state of charge values ​​of the energy storage battery at continuous time points and construct a dataset. Each dataset includes multiple sample pairs corresponding to multiple time points. Each sample pair includes a sample and a label. The sample includes current, voltage and temperature, and the label is the state of charge value of the battery.

[0111] Feature extraction module 202 is used to input the samples in the sample pair at the current time into a preset first training model for multi-scale feature extraction to obtain the multi-scale output features at the current time. The first training model includes multiple convolutional kernels and nonlinear activation functions.

[0112] The prediction module 203 is used to input the multi-scale output features of the current time into a preset second training model for prediction processing to obtain the prediction output result of the current time.

[0113] The loss function module 204 is used to construct a quantile loss function based on the prediction output at the current time, the label corresponding to the sample, and a preset quantile regression model.

[0114] The adjustment module 205 is used to adjust the first training model and the second training model according to the quantile loss function corresponding to the sample pair at each time step, so as to obtain the trained energy storage battery state of charge prediction model.

[0115] It should be noted that each module and unit in the energy storage battery state of charge prediction model training device in this embodiment corresponds one-to-one with each step in the energy storage battery state of charge prediction model training method in the aforementioned embodiment. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned energy storage battery state of charge prediction model training method, and will not be repeated here.

[0116] Example 3

[0117] Please see Figure 3 This embodiment provides an electronic device, including at least one processor 301 and a memory 302. Optionally, the device further includes a communication component 303. The processor 301, memory 302, and communication component 303 are connected via a bus 304.

[0118] In a specific implementation, at least one processor 301 executes computer execution instructions stored in memory 302, causing at least one processor 301 to perform the above-described method.

[0119] The specific implementation process of processor 301 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0120] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0121] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0122] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0123] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0124] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0125] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0126] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0127] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0128] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0129] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0130] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0131] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0132] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. 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 training a prediction model of the state of charge of an energy storage battery, characterized in that, include: Acquire the current, voltage, temperature and state of charge (SOC) values ​​of the energy storage battery at continuous time points, and construct a dataset. Each dataset includes multiple sample pairs corresponding to multiple time points. Each sample pair includes a sample and a label. The sample includes current, voltage and temperature, and the label is the SOC value. For the sample pair at the current time, the samples in the sample pair are input into a preset first training model for multi-scale feature extraction to obtain the multi-scale output features at the current time. The first training model includes multiple convolutional kernels and nonlinear activation functions. The multi-scale output features at the current moment are input into a preset second training model for prediction processing to obtain the prediction output result at the current moment. A quantile loss function is constructed based on the predicted output at the current moment, the label corresponding to the sample, and the preset quantile regression model. The first and second models to be trained are adjusted based on the quantile loss function corresponding to the sample pairs at each time step to obtain a trained energy storage battery state of charge prediction model.

2. The method according to claim 1, characterized in that, For the sample pair at the current time, the samples in the sample pair are input into a preset first training model for multi-scale feature extraction to obtain the multi-scale output features at the current time, including: The samples in the sample pair are input into the first convolution kernel to obtain the first convolution output, wherein the size of the first convolution kernel is 1×1; The samples in the sample pair are input into the second convolution kernel to obtain the second convolution output, wherein the size of the second convolution kernel is 3×3; The samples in the sample pair are input into the third convolution kernel to obtain the third convolution output, wherein the size of the third convolution kernel is 5×5; The first convolutional output, the second convolutional output, and the third convolutional output are concatenated and fused to obtain multi-scale fused features; Substituting the multi-scale fused features into a nonlinear activation function yields the multi-scale output features at the current time.

3. The method according to claim 2, characterized in that, The first convolution output satisfies: ， in, For the input sample, This is the kernel weight vector of the first convolutional kernel. is the bias vector of the first convolution kernel.

4. The method according to claim 2, characterized in that, The second convolution output satisfies: ， in, These are the intermediate features after the first dimensionality reduction and the ReLU activation function. This is the convolution kernel weight vector corresponding to the first dimension reduction. This is the bias vector corresponding to the first dimension reduction. This is the kernel weight vector of the second convolution kernel. is the bias vector of the second convolution kernel.

5. The method according to claim 2, characterized in that, The third convolution output satisfies: ， in, These are the intermediate features after the second dimensionality reduction and the ReLU activation function. This is the convolution kernel weight vector corresponding to the second dimension reduction. This is the bias vector corresponding to the second dimension reduction. The kernel weight vector of the third convolution kernel. is the bias vector of the third convolution kernel.

6. The method according to claim 1, characterized in that, The process involves inputting the multi-scale output features at the current moment into a preset second training model for prediction processing to obtain the prediction output result at the current moment, satisfying the following conditions: ， ， ， ， in, For the current time, for Multi-scale output features. for Cellular input, for The input gate, for The Gate of Oblivion for The output gate, The tanh activation function is used. for Input pre-activation value, The input weight matrix is ​​used to... Weight parameters projected onto the candidate state space. For the input circular weight matrix, For the input bias vector, for The input gate pre-activation value, Input gate weight matrix, The input gate loop weight matrix is... The input gate bias vector; for The pre-activation value of the forget gate, Forget gate weight matrix The forget gate loop weight matrix, This is the forget gate bias vector; It is the sigmoid activation function. for The output gate pre-activation value, Output gate weight matrix, This is the output gate loop weight matrix. This is the output gate bias vector. for The cell state controls the proportion of new inputs. for The normalized state is used to stabilize the gated output. for The predicted output results.

7. The method according to claim 1, characterized in that, The quantile loss function is constructed based on the predicted output at the current moment, the label corresponding to the sample, and the preset quantile regression model, satisfying: ， ， ， in, It is a sloping absolute value function. for The corresponding response variable, for The regression coefficient vector derived from the quantiles , The label is the label of the sample pair at the current moment.

8. A method for predicting the state of charge of an energy storage battery, characterized in that, include: Acquire current, voltage, and temperature over a period of time; The current, voltage, and temperature over a given period of time are input into the energy storage battery state of charge prediction model to obtain the predicted battery state of charge value and confidence interval output by the energy storage battery state of charge prediction model, wherein the energy storage battery state of charge prediction model is trained according to any one of the methods described in claims 1-7.

9. A training device for predicting the state of charge of an energy storage battery, characterized in that, include: The dataset module is used to obtain the current, voltage, temperature and state of charge values ​​of the energy storage battery at continuous time points and construct a dataset. Each dataset includes multiple sample pairs corresponding to multiple time points. Each sample pair includes a sample and a label. The sample includes current, voltage and temperature, and the label is the state of charge value of the battery. The feature extraction module is used to input the samples in the sample pair at the current time into a preset first training model for multi-scale feature extraction to obtain the multi-scale output features at the current time. The first training model includes multiple convolutional kernels and nonlinear activation functions. The prediction module is used to input the multi-scale output features of the current time into a preset second training model for prediction processing, and obtain the prediction output result of the current time. The loss function module is used to construct a quantile loss function based on the prediction output at the current time, the label corresponding to the sample, and a preset quantile regression model. The adjustment module is used to adjust the first and second models to be trained based on the quantile loss function corresponding to the samples at each time step, so as to obtain a trained energy storage battery state of charge prediction model.

10. An electronic device, characterized in that, include: At least one processor, at least one memory, and computer program instructions stored in the memory, which, when executed by the processor, implement the method as described in any one of claims 1-8.

Images