Method and device for determining state of charge of energy storage battery

Abstract

This specification provides a method and device for determining the state of charge of an energy storage battery. First, according to the preset training rules, the source domain data associated with the sample energy storage battery is used to train a target feature representation model based on the Att‑BiLSTM architecture with multi-dimensional feature representation; then, the target feature representation model is used as the initial feature extraction module to construct an initial state of charge prediction model based on the improved DA‑GAN, and the source domain data associated with the sample energy storage battery and the target domain data associated with the target energy storage battery are combined to train a preset state of charge prediction model that meets the requirements. In specific implementation, the operating data of the target energy storage battery in the current preset time period is obtained; the state of charge of the target energy storage battery is determined by processing the operating data in the current preset time period using the preset state of charge prediction model. Thereby, the state of charge of the target energy storage battery can be determined more efficiently and accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of energy storage batteries, and in particular to a method and device for determining the state of charge of an energy storage battery. Background Art

[0002] With the development and promotion of new energy technologies, energy storage batteries are widely used in energy storage systems such as solar and wind power generation and energy storage supporting equipment, energy storage charging stations, etc. due to their excellent performance such as high storage capacity, small size and zero pollution. Among them, accurate estimation of the state of charge SOC (State of Charge) of energy storage batteries is particularly important for the monitoring and management of system performance and safety of energy storage systems.

[0003] Based on the existing methods, since the sample energy storage batteries used in training the model are mostly different from the target energy storage batteries in actual application in terms of battery type, operating conditions, etc., it is easy to make errors when using the prediction model trained based on the sample data of the sample energy storage battery to predict the charge state of the target energy storage battery. In addition, the prediction model trained based on the existing method is relatively less accurate due to the limitations of the model's own structure.

[0004] To address the above problems, no effective solution has been proposed yet. Summary of the invention

[0005] The present specification provides a method and device for determining the state of charge of an energy storage battery, which can train a preset state of charge prediction model with high accuracy and good generalization at a relatively low data processing cost, and then can more efficiently and accurately determine the state of charge of a target energy storage battery based on the above preset state of charge prediction model, effectively reduce data processing errors, and reduce overall data processing costs.

[0006] This specification provides a method for determining the state of charge of an energy storage battery, including:

[0007] Obtaining the operating data of the target energy storage battery during the current preset time period;

[0008] The operation data of the target energy storage battery in the current preset time period is processed by using a preset state of charge prediction model to obtain a corresponding target prediction result; wherein the preset state of charge prediction model is a neural network model based on an improved DA-GAN, which is obtained by training according to preset training rules using target domain data associated with the target energy storage battery and source domain data associated with the sample energy storage battery; the preset state of charge prediction model at least includes a feature extraction module based on the Att-BiLSTM architecture of multi-dimensional feature representation;

[0009] According to the target prediction results, the charge state of the target energy storage battery is determined.

[0010] In one embodiment, the feature extraction module includes at least: a first LSTM layer and a second LSTM layer; wherein the first LSTM layer includes a plurality of first LSTM substructures connected in series in a forward time sequence, and the second LSTM layer includes a plurality of second LSTM substructures connected in series in a reverse time sequence;

[0011] The first LSTM substructure is also connected to a corresponding first attention unit, which is also connected to other first LSTM substructures in the first LSTM layer; the second LSTM substructure is also connected to a corresponding second attention unit, which is also connected to other second LSTM substructures in the second LSTM layer;

[0012] The first LSTM layer is used to extract corresponding process operation features; the second LSTM layer is used to extract corresponding initial operation features.

[0013] In one embodiment, the preset state of charge prediction model further includes: a domain feature splitting module, a domain discrimination module, and a state classification module;

[0014] The feature extraction module is connected to the domain feature splitting module, the domain feature splitting module is connected to the domain discrimination module and the state classification module, and the domain discrimination module is connected to the state classification module;

[0015] The state classification module includes: a first sub-classification network associated with the target domain, and a second sub-classification network associated with the source domain, wherein the first sub-classification network and the second sub-classification network have the same structure and share parameters.

[0016] In one embodiment, the operation data of the target energy storage battery in the current preset time period is processed using a preset state of charge prediction model to obtain a corresponding target prediction result, including:

[0017] Using the feature extraction module to process the operating data of the target energy storage battery in the current preset time period, obtain the initial operating characteristics of the current preset time period, and the process operating characteristics of the current preset time period;

[0018] Using the domain feature splitting module, common features and independent features are split from the initial operation features of the current preset time period and the process operation features of the current preset time period; wherein the independent features include: a first independent feature unique to the target domain and / or a second independent feature unique to the source domain;

[0019] Use the domain discrimination module to process common features and independent features to obtain the corresponding domain discrimination results;

[0020] When the domain discrimination result indicates the target domain, the domain discrimination result, the common feature, and the first independent feature are sent to a first sub-classification network in the state classification module;

[0021] The first sub-classification network is used to process the domain discrimination result, the common feature, and the first independent feature to obtain and output the corresponding target prediction result.

[0022] In one embodiment, the operating data at least includes: voltage, current, and temperature.

[0023] In one embodiment, before obtaining the operating data of the target energy storage battery in the current preset time period, the method further includes:

[0024] According to the preset training rules, an initial feature representation and state prediction model is constructed; wherein the initial feature representation and state prediction model comprises an initial feature representation model and an initial state prediction model connected to each other; the initial feature representation model comprises at least: an initial first LSTM layer, and an initial second LSTM layer;

[0025] Using source domain data associated with the sample energy storage battery, construct a corresponding first training set;

[0026] According to the preset training rules, the initial feature representation and state prediction model are trained using the first training set to obtain the target feature representation and state prediction model that meet the requirements;

[0027] According to the preset training rules, an initial state of charge prediction model based on the improved DA-GAN is constructed; wherein the initial state of charge prediction model includes an initial feature extraction module, an initial domain feature splitting module, an initial domain discrimination module, and an initial state classification module;

[0028] Migrating the parameters of the target feature representation model in the target feature representation and state prediction model to the initial feature extraction module in the initial state of charge prediction model, and migrating the parameters of the target state prediction model in the target feature representation and state prediction model to the initial state classification module in the initial state of charge prediction model; wherein the feature extraction module in the initial state of charge prediction model has the same structure as the target feature representation model, and the state classification module in the initial state of charge prediction model has the same partial structure as the target state prediction model;

[0029] Combining source domain data associated with the sample energy storage battery and target domain data associated with the target energy storage battery to construct a corresponding second training set;

[0030] According to the preset training rules, multiple rounds of model training are performed using the second training set and the initial state of charge prediction model to obtain a state of charge prediction model that meets the preset requirements.

[0031] In one embodiment, according to a preset training rule, multiple rounds of model training are performed using the second training set and the initial state of charge prediction model, including:

[0032] Perform the current round of model training as follows:

[0033] Obtain the state of charge prediction model of the previous round;

[0034] The state of charge prediction model of the previous round is used to process the second training data of the current round to obtain corresponding first-category prediction results and second-category prediction results; wherein the first-category prediction results are state prediction results output by the first sub-classification network, and the second-category prediction results are state prediction results output by the second sub-classification network;

[0035] According to the preset training rules, the function value of the target loss function of the current round is calculated using the second training data of the current round, the first type of prediction results, and the second type of prediction results;

[0036] According to the function value of the target loss function of the current round, the model parameters of the state of charge prediction model of the previous round are adjusted to obtain the state of charge prediction model of the current round.

[0037] In one embodiment, the target loss function includes an error function based on the state of charge of the source domain data and a discriminator loss function based on a generative adversarial network.

[0038] This specification also provides a device for determining the state of charge of an energy storage battery, comprising:

[0039] An acquisition module, used to acquire the operating data of the target energy storage battery during the current preset time period;

[0040] A processing module, used to process the operating data of the target energy storage battery in the current preset time period using a preset state of charge prediction model to obtain a corresponding target prediction result; wherein the preset state of charge prediction model is a neural network model based on an improved DA-GAN obtained by training according to preset training rules using target domain data associated with the target energy storage battery and source domain data associated with the sample energy storage battery; the preset state of charge prediction model at least includes a feature extraction module based on the Att-BiLSTM architecture of multi-dimensional feature representation;

[0041] The determination module is used to determine the charge state of the target energy storage battery according to the target prediction result.

[0042] This specification also provides a computer-readable storage medium on which computer instructions are stored. When the instructions are executed by a processor, the relevant steps of the method for determining the state of charge of the energy storage battery are implemented.

[0043] Based on the method and device for determining the state of charge of the energy storage battery provided in this specification, before specific implementation, the source domain data associated with the sample energy storage battery can be used as the first training set according to the preset training rules to train the target feature representation model of the Att-BiLSTM architecture based on multi-dimensional feature representation after structural improvement, and the corresponding target state prediction model; and the target feature representation model is used as the initial feature extraction module, and the target state prediction model is used as the initial state classification module to construct the initial state of charge prediction model based on the improved DA-GAN; and then the source domain data associated with the sample energy storage battery and the target domain data associated with the target energy storage battery are combined to construct the second training set; and the initial state of charge prediction model is domain-adaptively learned by using the second training set to obtain a preset state of charge prediction model that meets the requirements. In specific implementation, the operating data of the current preset time period of the target energy storage battery can be obtained first; and then the state of charge of the target energy storage battery can be determined by processing the operating data of the current preset time period using the preset state of charge prediction model. In this way, a preset state of charge prediction model with higher accuracy and better generalization can be trained with lower data processing cost, and then based on the preset state of charge prediction model, the state of charge of the target energy storage battery can be determined more efficiently and accurately, effectively reducing data processing errors and lowering the overall data processing cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of this specification, the drawings required for use in the embodiments will be briefly introduced below. The drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0045] Figure 1 It is a flowchart of a method for determining the state of charge of an energy storage battery provided by an embodiment of this specification;

[0046] Figure 2 It is a schematic diagram of an embodiment of a method for determining the state of charge of an energy storage battery provided by an embodiment of this specification, in a scenario example;

[0047] Figure 3 It is a schematic diagram of an embodiment of a method for determining the state of charge of an energy storage battery provided by an embodiment of this specification, in a scenario example;

[0048] Figure 4 It is a schematic diagram of an embodiment of a method for determining the state of charge of an energy storage battery provided by an embodiment of this specification, in a scenario example;

[0049] Figure 5 It is a schematic diagram of an embodiment of a method for determining the state of charge of an energy storage battery provided by an embodiment of this specification, in a scenario example;

[0050] Figure 6 It is a schematic diagram of an embodiment of a method for determining the state of charge of an energy storage battery provided by an embodiment of this specification, in a scenario example;

[0051] Figure 7 It is a schematic diagram of the structure of a server provided by an embodiment of this specification;

[0052] Figure 8 It is a schematic diagram of the structure of a device for determining the state of charge of an energy storage battery provided by an embodiment of this specification;

[0053] Fig. 9 It is a schematic diagram of an embodiment of a method for determining the state of charge of an energy storage battery provided by an embodiment of this specification, in a scenario example;

[0054] Fig.10 It is a schematic diagram of an embodiment of a method for determining the state of charge of an energy storage battery provided by an embodiment of this specification, in a scenario example. DETAILED DESCRIPTION

[0055] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.

[0056] It should be noted that in the embodiments of this specification, certain software, components, models and other existing solutions in the industry may be mentioned, and they should be regarded as exemplary. Their purpose is only to illustrate the feasibility of implementing the technical solution of this application, but it does not mean that the applicant has or will necessarily use the solution.

[0057] Considering the existing methods, on the one hand, due to the limitations of the model architecture, the constructed prediction model often only focuses on the single feature of the energy storage battery, resulting in the inability to comprehensively analyze and process the operation of the energy storage battery, thereby affecting the model accuracy and the generalization of the model, making it easy to make errors when using the model to predict the charge state of the energy storage battery.

[0058] On the other hand, when training prediction models based on existing methods, in most cases, only unlabeled sample data of the target energy storage battery for actual application can be obtained, and the target energy storage battery and the sample energy storage battery for experimental testing have differences in battery type and operating conditions, which leads to the need to first label the sample data of the target energy storage battery; then, using the labeled sample data of the target energy storage battery, adjust the model previously trained based on the sample data of the sample energy storage battery to obtain a prediction model suitable for the target energy storage battery. This will inevitably increase the training cost of the model and affect the overall data processing efficiency.

[0059] This application has noticed the above problems, and combined with the above problems, first of all, it is considered that the model structure can be improved. By introducing the Att-BiLSTM architecture, a feature representation model that can extract operation features of multiple dimensions including initial operation features and process operation features is constructed. Then, the feature representation model is used as the initial feature extraction module and connected to the domain adaptive architecture to construct an initial state of charge prediction model based on the improved DA-GAN. Then, the source domain data associated with the sample energy storage battery with annotations and the target domain data associated with the target energy storage battery without annotations can be combined to perform deep domain adaptive learning on the above initial state of charge prediction model, so as to train and obtain a preset state of charge prediction model based on the improved DA-GAN that meets the requirements and is suitable for the target energy storage battery without the need to annotate the target domain data. Thereby, the overall data processing cost can be effectively reduced, the model accuracy can be improved, and the above preset state of charge prediction model can be used to determine the state of charge of the target energy storage battery more efficiently and accurately.

[0060] See also Figure 1 As shown, the embodiment of this specification provides a method for determining the state of charge of an energy storage battery. In specific implementation, the method may include the following contents:

[0061] S101: Acquire the operating data of the target energy storage battery in the current preset time period.

[0062] S102: Using a preset state of charge prediction model to process the operating data of the target energy storage battery in the current preset time period to obtain a corresponding target prediction result; wherein the preset state of charge prediction model is a neural network model based on an improved DA-GAN, which is trained according to preset training rules using target domain data associated with the target energy storage battery and source domain data associated with the sample energy storage battery; the preset state of charge prediction model at least includes a feature extraction module based on the Att-BiLSTM architecture of multi-dimensional feature representation.

[0063] S103: Determine the state of charge of the target energy storage battery according to the target prediction result.

[0064] The target energy storage battery mentioned above may be a battery actually used in an energy storage system based on energy sources such as solar energy and wind energy.

[0065] The above-mentioned current preset time period may specifically be the most recent 10 minutes, or the most recent day, etc. In specific implementation, a suitable time period may be flexibly set as the preset time period according to the system scale of the energy storage system and the monitoring requirements for the energy storage system.

[0066] The operating data of the target energy storage battery in the current preset time period may specifically include: the operating data of the target energy storage battery at each time point (or moment) measured in the preset time period. Specifically, the operating data at least includes parameters such as the voltage (which can be recorded as V), the current (which can be recorded as I), and the temperature (which can be recorded as T) of the target energy storage battery.

[0067] The above-mentioned preset state of charge prediction model can be specifically understood as a neural network model based on an improved model structure that is trained in advance according to preset training rules, using target domain data associated with the target energy storage battery and source domain data associated with the sample energy storage battery. Based on the above-mentioned preset state of charge prediction model, the corresponding prediction result of the state of charge of the battery can be output by processing the input battery operation data of a preset time period.

[0068] The above-mentioned sample energy storage battery can be specifically understood as an energy storage battery used in an experimental test in advance in an experimental test environment.

[0069] Before specific implementation, a sample energy storage battery can be used to conduct an experimental test in an experimental test environment, and the operating data of the sample energy storage battery during the experimental test and the corresponding state of charge data can be collected as source domain data associated with the sample energy storage battery, which can be recorded as the source domain (or source domain). Furthermore, the operating data of the sample energy storage battery can be annotated according to the state of charge data to obtain the annotated operating data of the sample energy storage battery. Therefore, in this embodiment, the above-mentioned source domain data can specifically be annotated data.

[0070] In specific implementation, during the use of the target energy storage battery, the operating data of the target energy storage battery can be collected and recorded as the target domain data associated with the target energy storage battery, which can be recorded as the target domain (or target field). In this embodiment, the above target domain data can specifically be data without annotations.

[0071] The target energy storage battery and the sample energy storage battery may have different battery types and / or different operating conditions.

[0072] The preset state of charge prediction model may at least include a feature extraction module based on the Att-BiLSTM architecture of multi-dimensional feature representation. The preset feature extraction module may specifically be a feature representation model that uses the source domain data of the sample energy storage battery separately according to preset training rules.

[0073] The Att-BiLSTM (Attention-Based Bidirectional Long Short-Term Memory Networks) mentioned above can be specifically understood as a bidirectional long short-term memory network based on the attention mechanism. Based on this network, special storage units and gating mechanisms can be introduced to effectively capture and process long-term dependencies in sequence data.

[0074] It should be noted that the Att-BiLSTM architecture is introduced and used here because it is possible to establish long-distance dependencies between the forward and reverse time series by using independent LSTM layers in opposite directions, so that the model can more accurately extract the required process operation characteristics and initial operation characteristics; at the same time, by introducing the attention mechanism, the model can more easily focus on features with relatively high value, thereby effectively solving the gradient vanishing problem of the conventional LSTM architecture.

[0075] The above-mentioned DA-GAN (Deep Attention Generative Adversarial Networks) can be specifically understood as a domain adaptive network framework based on generative adversarial networks. Based on this network framework, the differences between the source domain (or source domain) and the target domain (or target domain) can be learned through adaptive learning to achieve the migration of the source domain model to the target domain.

[0076] It should be noted that the DA-GAN architecture is introduced and used here because this type of domain adaptation method can cope with the situation where SOC data in the target domain is missing, reducing data costs; at the same time, DA-GAN is combined with a generative adversarial network. The confrontation between the generator (or encoder) and the discriminator (or decoder) makes the two processes of data generation and domain discrimination complement each other. This application applies and improves the improved DA-GAN framework on the basis of DA-GAN. The application aspects include the domain discrimination module and discriminator loss in the application, and the improvement aspects include the domain feature splitting module and the state classification module, which are more suitable for SOC estimation tasks.

[0077] In specific implementation, the feature extraction module based on the above preset can process the input operation data and extract operation features of multiple different dimensions, such as initial operation features based on local dimensions and process operation features based on global dimensions. Furthermore, the preset state of charge prediction model can use the operation features of multiple different dimensions, such as initial operation features and process operation features, to make a more comprehensive and detailed state of charge prediction.

[0078] In a specific implementation, the operating data of the target energy storage battery during the current preset time period can be used as a model input and input into a preset state of charge prediction model; the model is run, and the operating data of the current preset time period is processed by the model to obtain a corresponding target prediction result as a model output; and the state of charge of the target energy storage battery is determined based on the target prediction result.

[0079] In specific implementation, according to the corresponding monitoring rules, the operating data of the target energy storage battery in the current preset time period can be obtained at each preset time interval (for example, every 2 hours); and the preset battery status prediction model is used to process the operating data of the target energy storage battery in the current preset time period to determine the current charge state of the target energy storage battery.

[0080] Furthermore, it is also possible to judge whether the current state of the target energy storage battery is healthy or whether there is a safety risk based on the current state of charge of the target energy storage battery. If it is determined that the current state of the target energy storage battery is unhealthy or there is a safety risk, a corresponding error message is generated to prompt the user to promptly check and eliminate the safety hazards of the target energy storage battery to ensure the safe and healthy operation of the target energy storage battery.

[0081] Based on the above embodiment, the improved preset state of charge prediction model can be used to accurately and efficiently determine the state of charge of the target energy storage battery by processing the operating data of the target energy storage battery in the current preset time period.

[0082] In some embodiments, after obtaining the operating data of the target energy storage battery for the current preset time period, the operating data for the current preset time period can also be normalized to obtain normalized operating data that is compatible with a preset state of charge prediction model and eliminates errors caused by differences in numerical amplitudes between different operating data.

[0083] After obtaining the operating data of the target energy storage battery in the current preset time period, data noise reduction may be performed on the operating data in the current preset time period to reduce data errors and improve data accuracy.

[0084] After obtaining the operating data of the target energy storage battery in the current preset time period, the operating data of the current preset time period can be converted into corresponding time series data according to the time sequence to adapt to the feature extraction module in the preset state of charge prediction model.

[0085] In some embodiments, see Figure 2 As shown, the above-mentioned preset state of charge prediction model at least includes a feature extraction module based on an LSTM architecture with multi-dimensional feature representation.

[0086] For details, please refer to Figure 3 As shown, the feature extraction module (or bidirectional feature extraction module) may at least include: a first LSTM layer and a second LSTM layer; wherein the first LSTM layer includes a plurality of first LSTM substructures (e.g., LSTM1) connected in series in a forward time sequence, and the second LSTM layer includes a plurality of second LSTM substructures (e.g., LSTM2) connected in series in a reverse time sequence;

[0087] The first LSTM substructure is also connected to a corresponding first attention unit (e.g., Attention1), which is also connected to other first LSTM substructures in the first LSTM layer; the second LSTM substructure is also connected to a corresponding second attention unit (e.g., Attention2), which is also connected to other second LSTM substructures in the second LSTM layer;

[0088] The first LSTM layer is used to extract corresponding process operation features; the second LSTM layer is used to extract corresponding initial operation features.

[0089] Specifically, in the above feature extraction module, the first LSTM layer and the second LSTM layer are independent of each other and in opposite directions. By using the first LSTM layer and the second LSTM layer together, long-distance dependencies between the forward and reverse time series can be established respectively. Specifically, the first LSTM layer can extract process operation features from the forward time series, and the second LSTM layer can reversely infer the initial operation features from the reverse time series. In addition, by setting corresponding attention units in the first LSTM layer and the second LSTM layer, the attention mechanism is introduced, which can make it easier for the model to focus on more valuable features, effectively solving the gradient vanishing problem of the conventional LSTM architecture.

[0090] In addition, based on the attention mechanism, the first LSTM layer can focus and pay attention to the initial operation characteristics based on the local initialization of the operation data during the training process; based on the attention mechanism, the second LSTM layer can focus and pay attention to the process operation characteristics of the global complete process of the operation data during the training process.

[0091] For specific implementation, please refer to Figure 3 As shown, based on the model structure of the first LSTM layer and the second LSTM layer, when the operating data of a preset time period is input into the first LSTM layer and the second LSTM layer respectively, the number of LSTM substructures contained in the first LSTM layer and the second LSTM layer can be determined according to the number of time points of the operating data; and the operating data of multiple time points can be input into the corresponding LSTM substructures in the first LSTM layer and the second LSTM layer in chronological order.

[0092] Specifically, for example, multiple time points can be represented as: t0, t1...tn. Correspondingly, the operation data of multiple time points can be represented as (Vt0, It0, Tt0), (Vt1, It1, Tt1)...(Vtn, Itn, Ttn) respectively.

[0093] In this way, while each LSTM substructure processes the running data of the corresponding time point directly input, it also receives the data of other time points processed and output by the adjacent previous LSTM substructure; and through the forget gate, the data of the previous time point is connected with the data of the corresponding time point currently being processed, so as to realize continuous processing of time series data. In this way, it is possible to mine and obtain abstract features that can reflect the long-term dependency relationship within the preset time period and are applicable to both the source domain and the target domain, and obtain intermediate shared features with better effects.

[0094] Specifically, the first LSTM layer can be used to perform feature processing based on local dimensions on the input operation data of multiple time points based on the attention mechanism, and obtain and output the corresponding initial operation features (or starting features, denoted as P1). At the same time, the second LSTM layer can be used to perform feature processing based on global dimensions on the input operation data of multiple time points based on the attention mechanism, and obtain and output the corresponding process operation features (or process features, denoted as P2).

[0095] Based on the above embodiments, the feature extraction module can be used to extract various types of operation features comprehensively and precisely based on multiple dimensions.

[0096] In some embodiments, see Figure 2 As shown, the preset state of charge prediction model may also include: a domain feature splitting module, a domain discrimination module and a state classification module;

[0097] The feature extraction module is connected to the domain feature splitting module, the domain feature splitting module is connected to the domain discrimination module and the state classification module, and the domain discrimination module is connected to the state classification module;

[0098] The state classification module includes: a first sub-classification network associated with the target domain, and a second sub-classification network associated with the source domain, wherein the first sub-classification network and the second sub-classification network have the same structure and share parameters.

[0099] Specifically, the above status classification module can be found in Fig. 9 As shown, the initial state classification module and the initial feature extraction module can be trained using labeled source domain data.

[0100] Specifically, the domain discrimination module may include at least a binary classification fully connected layer (FC) and a softmax regression layer on different time dimensions. The sub-classification network may include at least a binary classification fully connected layer (FC) and a softmax regression layer on different time dimensions.

[0101] In a specific implementation, the domain feature splitting module is used to split the common features shared by the target domain and the source domain, and the independent features from the initial operation features and the process operation features; wherein the independent features include: a first independent feature unique to the target domain and / or a second independent feature unique to the source domain;

[0102] The domain discrimination module is used to discriminate the domain according to the features separated by the domain feature separation module, and output the corresponding domain discrimination result; and according to the domain discrimination result, send the domain discrimination result, the common features, and the independent features matching the domain discrimination result to the state classification module;

[0103] The state prediction model determines a matching classification sub-network according to the domain discrimination result; and sends the domain discrimination result, common features, and independent features matching the domain discrimination result to the matching sub-classification network for prediction.

[0104] Specifically, the state classification module may include two branches: a first sub-classification network associated with the target domain, and a second sub-classification network associated with the source domain. The first sub-classification network includes at least one fully connected layer (which can be denoted as FC1), and the second sub-classification network includes at least one fully connected layer (which can be denoted as FC2). Specifically, the first sub-classification network can be obtained by adaptive transfer learning training using the second sub-classification network during the model training process, that is, the two sub-classification network parameters are shared.

[0105] In specific implementation, based on the domain discrimination result, when it is determined that it belongs to the source domain, the domain discrimination result, the common feature, and the second independent feature matching the domain discrimination result are sent to the corresponding second sub-classification network associated with the source domain; and the second sub-classification network is used to jointly process the above-mentioned feature data to perform charge state prediction based on the source domain and output the corresponding prediction result.

[0106] On the contrary, based on the domain discrimination result, when it is determined that it belongs to the target domain, the domain discrimination result, the common features, and the first independent features matching the domain discrimination result are sent to the corresponding first sub-classification network associated with the target domain; and the first sub-classification network is used to jointly process the above-mentioned feature data to perform charge state prediction based on the target domain and output the corresponding prediction result.

[0107] Based on the above embodiments, the preset state of charge prediction model which includes the feature extraction module, the field feature splitting module, the field discrimination module and the state classification module can be used to effectively process the input operation data, distinguish different fields, and carry out targeted state of charge prediction related to the field.

[0108] In some embodiments, specifically, the above-mentioned domain discrimination module may include a fully connected layer (which can be denoted as FC) on different time dimensions. The above-mentioned domain discrimination module may also include a softmax layer. Accordingly, based on the above structure, the domain discrimination module can process the multi-dimensional features (including initial operation features and process operation features) split out by the domain feature splitting module separately by utilizing two different connection layers, and adapt them to the multi-dimensional features before splitting; and effectively learn the nonlinear relationship in the above-mentioned features, so as to accurately determine the domain to which the operation data belongs based on the input multi-dimensional features.

[0109] In some embodiments, see Figure 4 As shown, the above-mentioned use of the preset state of charge prediction model to process the operating data of the target energy storage battery in the current preset time period to obtain the corresponding target prediction result may include the following contents during specific implementation:

[0110] S1: using a feature extraction module to process the operation data of the target energy storage battery in the current preset time period, and obtain the initial operation characteristics of the current preset time period and the process operation characteristics of the current preset time period;

[0111] S2: using a domain feature splitting module to split common features and independent features from the initial operation features of the current preset time period and the process operation features of the current preset time period; wherein the independent features include: a first independent feature unique to the target domain and / or a second independent feature unique to the source domain;

[0112] S3: Use the domain discrimination module to process common features and independent features to obtain corresponding domain discrimination results;

[0113] S4: When the domain discrimination result indicates the target domain, the domain discrimination result, the common feature, and the first independent feature are sent to the first sub-classification network in the state classification module;

[0114] S5: Utilize the first sub-classification network to process the domain discrimination result, the common feature, and the first independent feature to obtain and output the corresponding target prediction result.

[0115] In specific implementation, when the domain discrimination result indicates the source domain, the common features and the second independent features in the domain discrimination result can be sent to the second sub-classification network in the state classification module; then the first sub-classification network with the same structure and shared parameters is used to process the common features and the first independent features in the domain discrimination result to obtain and output the corresponding target prediction result.

[0116] Based on the above embodiment, the preset state of charge prediction model can be used to first process the operating data of the current preset time period of the target energy storage battery, and extract the multi-dimensional operating characteristics including the initial operating characteristics of the current preset time period and the process operating characteristics of the current preset time period; then, according to the above multi-dimensional operating characteristics, the corresponding feature splitting is performed, and the corresponding field is accurately determined; then, different fields can be distinguished, and the corresponding multi-dimensional operating characteristics split can be targetedly processed by using a sub-classification network that can process characteristics of different fields, so that the state of charge of the target energy storage battery can be accurately predicted, and a target prediction result with high accuracy and small error can be obtained.

[0117] In some embodiments, the operation data may at least include: voltage, current, temperature, etc. Further, the operation data may also include: other parameters such as temperature curve.

[0118] In some embodiments, see Figure 5 As shown, before obtaining the operating data of the target energy storage battery in the current preset time period, the method may further include the following contents when implemented:

[0119] S1: constructing an initial feature representation and state prediction model according to a preset training rule; wherein the initial feature representation and state prediction model comprises a connected initial feature representation model and an initial state prediction model; the initial feature representation model comprises at least: an initial first LSTM layer, and an initial second LSTM layer;

[0120] S2: Using source domain data associated with the sample energy storage battery, construct a corresponding first training set;

[0121] S3: According to a preset training rule, the initial feature representation and state prediction model are trained using the first training set to obtain a target feature representation and state prediction model that meets the requirements;

[0122] S4: constructing an initial state of charge prediction model based on the improved DA-GAN according to a preset training rule; wherein the initial state of charge prediction model includes an initial feature extraction module, an initial domain feature splitting module, an initial domain discrimination module, and an initial state classification module;

[0123] S5: Migrating the parameters of the target feature representation model in the target feature representation and state prediction model to the initial feature extraction module in the initial state of charge prediction model, and migrating the parameters of the target state prediction model in the target feature representation and state prediction model to the initial state classification module in the initial state of charge prediction model; wherein the feature extraction module in the initial state of charge prediction model has the same structure as the target feature representation model, and the state classification module in the initial state of charge prediction model has the same partial structure as the target state prediction model;

[0124] S6: Combining the source domain data associated with the sample energy storage battery and the target domain data associated with the target energy storage battery to construct a corresponding second training set;

[0125] S7: According to the preset training rules, multiple rounds of model training are performed using the second training set and the initial state of charge prediction model to obtain a preset state of charge prediction model that meets the requirements.

[0126] Based on the above embodiments, the target feature representation model and target state prediction model based on multi-dimensional feature representation that meet the requirements can be first trained using the labeled source domain data according to the preset training rules; then, according to the preset training rules, the parameters of the target feature representation model are migrated to the initial feature extraction module, and the parameters of the target state prediction model are migrated to the initial state classification module, and then connected to the DA-GAN architecture, and other modules in the DA-GAN architecture are modified accordingly to obtain an initial charge state prediction model based on the improved DA-GAN; then, according to the preset training rules, the labeled source domain data and the unlabeled target domain data are jointly used to train the initial charge state prediction model, so that there is no need to waste resources and time to additionally label the target domain data, and a preset state of charge prediction model with high accuracy and good effect can be efficiently trained at a relatively small data processing cost.

[0127] In some embodiments, during specific implementation, the operating data of the sample energy storage battery can be extracted from the source domain data of the sample energy storage battery according to preset training rules as the first training data to construct a first training set; the first training set is then used to train the initial feature representation model to obtain a feature representation model that meets the requirements, which can be recorded as the target feature representation model.

[0128] In specific implementation, the above-trained target feature representation model and target state prediction model can be connected to the DA-GAN architecture according to the preset training rules, and the domain feature splitting module, domain discrimination module, and state classification module in the DA-GAN architecture can be modified accordingly, so as to obtain the initial state of charge prediction model based on the improved DA-GAN architecture.

[0129] Then, the operation data of the preset time period of the sample energy storage battery and the corresponding state of charge data are extracted from the source domain data of the sample energy storage battery; and the operation data of the preset time period of the sample energy storage battery is annotated using the state of charge data to obtain the first type of sample data; at the same time, the operation data of the preset time period of the target energy storage battery is extracted from the target domain data of the target energy storage battery as the second type of sample data. The first type of sample data and the second type of sample data are then combined as the second training data to construct a second training set. The second training set is then used to train the initial state of charge prediction model to perform multiple rounds of domain adaptive transfer learning, and the model parameters are continuously adjusted and optimized to obtain a preset state of charge prediction model based on the improved DA-GAN that meets the requirements.

[0130] When the model is trained, the second training set can be split into a second training subset and a second test subset. Accordingly, the model can be trained with the second training subset first, and then the trained model can be adjusted with the second test subset.

[0131] In some embodiments, the above-mentioned method performs multiple rounds of model training using the second training set and the initial state of charge prediction model according to the preset training rules, and the specific implementation includes:

[0132] See also Figure 6 As shown, the current round of model training can be performed in the following way:

[0133] S1: Obtain the state of charge prediction model of the previous round;

[0134] S2: Process the second training data of the current round using the state of charge prediction model of the previous round to obtain corresponding first-category prediction results and second-category prediction results; wherein the first-category prediction results are state prediction results output by the first sub-classification network, and the second-category prediction results are state prediction results output by the second sub-classification network;

[0135] S3: According to a preset training rule, using the second training data of the current round, the first type of prediction results, and the second type of prediction results, calculate the function value of the target loss function of the current round;

[0136] S4: According to the function value of the target loss function of the current round, the model parameters of the state of charge prediction model of the previous round are adjusted to obtain the state of charge prediction model of the current round.

[0137] In specific implementation, domain-adaptive transfer learning can be achieved by adjusting the model parameters of the state of charge prediction model of the previous round according to the function value of the target loss function of the current round.

[0138] In some embodiments, the target loss function may specifically include an error function based on the state of charge of the source domain data and a discriminator loss function based on a generative adversarial network.

[0139] In specific implementation, based on the above-mentioned target loss function, during the model training process, when the domain discrimination module discriminates the source domain data as the target domain, or discriminates the target domain as the source domain, the discriminator loss function takes effect, and the parameters of the next round of domain discrimination module are adjusted and corrected; when the domain discrimination module discriminates the source domain data as the source domain, the independent features output by the domain feature splitting module are used as the second independent features, and are sent to the second sub-classification network in the state classification module together with the common features. When the state of charge output by the second sub-classification network is inconsistent with the true label, the error function takes effect, and the parameters of the next round of domain feature splitting module are adjusted and corrected.

[0140] In some embodiments, the above target loss function may further include an error function based on the state of charge of the source domain data, wherein the error function may simultaneously integrate the model mean square error and the MMD loss.

[0141] Among them, the above-mentioned MMD (Max mean discrepancy) loss can be specifically understood as a loss function used to measure the distance between two different but related distributions.

[0142] Correspondingly, using the function value of the above objective loss function can guide the model to perform domain adaptive transfer learning relatively more effectively.

[0143] According to the above method, after the current round of model training is completed, it can be detected whether the model training end condition is met. For example, it can be detected whether the function value of the target loss function of the current round is less than or equal to a preset loss function threshold.

[0144] When it is determined that the function value of the target loss function of the current round is less than or equal to the preset loss function threshold, the model training can be determined to be completed, and the state of charge prediction model of the current round is determined as a preset state of charge prediction model that meets the requirements.

[0145] On the contrary, when it is determined that the function value of the target loss function of the current round is greater than the preset loss function threshold, the above method can be repeated to perform the next round of model training until the function value of the target loss function is less than or equal to the preset loss function threshold.

[0146] Based on the above embodiment, multiple rounds of model training may be performed using the second training data set according to preset training rules to obtain a preset state of charge prediction model that meets the requirements.

[0147] In some embodiments, the above-mentioned calculation of the function value of the target loss function of the current round according to the preset training rules using the second training data of the current round, the first type of prediction results, and the second type of prediction results may include the following contents during specific implementation:

[0148] S1: Determine the mean square error of the first type of model based on the second training data and the first type of prediction results of the current round; determine the mean square error of the second type of model based on the second training data and the second type of prediction results of the current round;

[0149] S2: Determine the model mean square error of the previous round of state of charge prediction model based on the first type model mean square error and the second type model mean square error;

[0150] S3: Calculate the MMD loss of the state of charge prediction model of the previous round according to the second training data of the current round, the first type of prediction results, and the second type of prediction results;

[0151] S4: Calculate the function value of the target loss function of the current round according to the model mean square error of the state of charge prediction model of the previous round and the MMD loss of the state of charge prediction model of the previous round.

[0152] In specific implementation, the function value of the target loss function of the current round can be calculated according to the following formula:

[0153]

[0154] Among them, L DANN Represents the function value of the target loss function of the current round, L MSE represents the model mean square error of the state of charge prediction model in the previous round, γ represents the first weight coefficient, and λ MMD represents the MMD loss of the state of charge prediction model in the previous round, q s represents the source domain, q t Indicates the target domain.

[0155] In specific implementation, the model mean square error of the state of charge prediction model of the previous round can be calculated according to the following formula:

[0156]

[0157] Among them, ω1 represents the second weight coefficient, ω2 represents the third weight coefficient, represents the mean square error of the first type model, represents the mean square error of the second type model.

[0158] Based on the above embodiments, the function value of the target loss function with better effect can be accurately determined, and then the model parameters can be effectively adjusted based on the above target loss function to obtain a preset state of charge prediction model that meets the requirements.

[0159] In some embodiments, after determining the charge state of the target energy storage battery according to the target prediction result, the method may further include the following contents: judging whether the target energy storage battery has a safety risk according to the current charge state of the target energy storage battery; when it is determined that the target energy storage battery has a safety risk, querying a preset state template according to the current charge state of the target energy storage battery to determine the risk type and risk level of the target energy storage battery; querying a preset battery maintenance processing strategy set according to the risk type and risk level of the target energy storage battery to determine a matching target maintenance processing strategy; and then performing corresponding maintenance processing on the target energy storage battery according to the target maintenance processing strategy to eliminate safety risks and protect the healthy and safe operation of the target energy storage battery.

[0160] The preset battery maintenance processing strategy set includes a plurality of preset maintenance processing strategies, and each preset maintenance processing strategy corresponds to at least one combination of risk type and risk level.

[0161] Before specific implementation, a large number of battery maintenance processing records for different risk situations can be obtained; and then the battery maintenance processing records are clustered to construct a preset battery maintenance processing strategy set.

[0162] As can be seen from the above, based on the method for determining the state of charge of the energy storage battery provided by the embodiment of this specification, before specific implementation, the source domain data associated with the sample energy storage battery can be used as the first training set according to the preset training rules to train the target feature representation model of the Att-BiLSTM architecture based on multi-dimensional feature representation after structural improvement, and the corresponding target state prediction model; and the target feature representation model is used as the initial feature extraction module, and the target state prediction model is used as the initial state classification module to construct the initial state of charge prediction model based on the improved DA-GAN; and then the source domain data associated with the sample energy storage battery and the target domain data associated with the target energy storage battery are combined to construct a second training set; and the initial state of charge prediction model is domain-adaptively learned by using the second training set to obtain a preset state of charge prediction model that meets the requirements. In specific implementation, the operating data of the current preset time period of the target energy storage battery can be first obtained; and then the state of charge of the target energy storage battery is determined by processing the operating data of the current preset time period using the preset state of charge prediction model. In this way, a preset state of charge prediction model with higher accuracy and better generalization can be trained with lower data processing cost, and then based on the preset state of charge prediction model, the state of charge of the target energy storage battery can be determined more efficiently and accurately, effectively reducing data processing errors and reducing overall data processing costs.

[0163] This specification also provides a data processing method, which may include the following contents during implementation:

[0164] S1: constructing an initial feature representation and state prediction model according to a preset training rule; wherein the initial feature representation and state prediction model comprises a connected initial feature representation model and an initial state prediction model; the initial feature representation model comprises at least: an initial first LSTM layer, and an initial second LSTM layer;

[0165] S2: Using source domain data associated with the sample energy storage battery, construct a corresponding first training set;

[0166] S3: According to a preset training rule, the initial feature representation and state prediction model are trained using the first training set to obtain a target feature representation and state prediction model that meets the requirements;

[0167] S4: constructing an initial state of charge prediction model based on the improved DA-GAN according to a preset training rule; wherein the initial state of charge prediction model includes an initial feature extraction module, an initial domain feature splitting module, an initial domain discrimination module, and an initial state classification module;

[0168] S5: Migrating the parameters of the target feature representation model in the target feature representation and state prediction model to the initial feature extraction module in the initial state of charge prediction model, and migrating the parameters of the target state prediction model in the target feature representation and state prediction model to the initial state classification module in the initial state of charge prediction model; wherein the feature extraction module in the initial state of charge prediction model has the same structure as the target feature representation model, and the state classification module in the initial state of charge prediction model has the same partial structure as the target state prediction model;

[0169] S6: Combining the source domain data associated with the sample energy storage battery and the target domain data associated with the target energy storage battery to construct a corresponding second training set;

[0170] S7: According to the preset training rules, multiple rounds of model training are performed using the second training set and the initial state of charge prediction model to obtain a preset state of charge prediction model that meets the requirements.

[0171] Therefore, with low data processing cost and through domain adaptive deep learning, a preset state of charge prediction model with high accuracy and good effect for the target energy storage battery suitable for practical applications can be efficiently trained.

[0172] This specification embodiment provides a server, referring to Figure 7The server includes a network communication port 701, a processor 702 and a memory 703, and the above structures are connected through internal cables so that each structure can perform specific data interaction.

[0173] The network communication port 701 can be used to obtain the operation data of the target energy storage battery in the current preset time period.

[0174] The processor 702 can be specifically used to process the operating data of the target energy storage battery in the current preset time period using a preset state of charge prediction model to obtain a corresponding target prediction result; wherein the preset state of charge prediction model is a neural network model based on an improved DA-GAN, which is obtained by training according to preset training rules using target domain data associated with the target energy storage battery and source domain data associated with the sample energy storage battery; the preset state of charge prediction model at least includes a feature extraction module based on the Att-BiLSTM architecture of multi-dimensional feature representation; and the state of charge of the target energy storage battery is determined according to the target prediction result.

[0175] The memory 703 may be specifically used to store corresponding instruction programs, as well as operating data of the target energy storage battery in the current preset time period, target prediction results and other data.

[0176] Based on the above method, the relevant structural performance of the server can be effectively utilized, the data processing speed of the electronic device can be improved, and the determination data processing of the charge state of the energy storage battery can be efficiently realized.

[0177] In this embodiment, the network communication port 701 can be a virtual port that is bound to different communication protocols so that different data can be sent or received. For example, the network communication port can be a port responsible for web data communication, a port responsible for FTP data communication, or a port responsible for email data communication. In addition, the network communication port can also be a physical communication interface or communication chip. For example, it can be a wireless mobile network communication chip, such as GSM, CDMA, etc.; it can also be a Wifi chip; it can also be a Bluetooth chip.

[0178] In this embodiment, the processor 702 may be implemented in any appropriate manner. For example, the processor may take the form of a microprocessor or processor and a computer-readable medium storing a computer-readable program code (such as software or firmware) executable by the (micro)processor, a logic gate, a switch, an application specific integrated circuit (ASIC), a programmable logic controller, and an embedded microcontroller, etc. This specification does not limit this.

[0179] In this embodiment, the memory 703 may include multiple levels. In a digital system, anything that can store binary data can be a memory; in an integrated circuit, a circuit with a storage function that has no physical form is also called a memory, such as RAM, FIFO, etc.; in a system, a storage device with a physical form is also called a memory, such as a memory stick, TF card, etc.

[0180] The embodiment of the present specification also provides a computer-readable storage medium based on the above-mentioned method for determining the state of charge of the energy storage battery, wherein the computer-readable storage medium stores computer program instructions, and when the computer program instructions are executed, the following are implemented: obtaining the operating data of the target energy storage battery in the current preset time period; using a preset state of charge prediction model to process the operating data of the target energy storage battery in the current preset time period to obtain a corresponding target prediction result; wherein the preset state of charge prediction model is a neural network model based on an improved DA-GAN obtained by training according to preset training rules using target domain data associated with the target energy storage battery and source domain data associated with the sample energy storage battery; the preset state of charge prediction model at least includes a feature extraction module based on the Att-BiLSTM architecture of multi-dimensional feature representation; and determining the state of charge of the target energy storage battery according to the target prediction result.

[0181] In this embodiment, the storage medium includes, but is not limited to, a random access memory (RAM), a read-only memory (ROM), a cache, a hard disk (HDD), or a memory card. The memory may be used to store computer program instructions. The network communication unit may be an interface for network connection communication set in accordance with the standard specified by the communication protocol.

[0182] In this embodiment, the functions and effects specifically implemented by the program instructions stored in the computer-readable storage medium can be explained in comparison with other implementations and will not be described in detail here.

[0183] The embodiments of the present specification also provide a computer program product, which at least includes a computer program, and when the computer program is executed by a processor, the following method steps are implemented: obtaining the operating data of the target energy storage battery in the current preset time period; using a preset state of charge prediction model to process the operating data of the target energy storage battery in the current preset time period to obtain a corresponding target prediction result; wherein the preset state of charge prediction model is a neural network model based on an improved DA-GAN, which is obtained by training according to preset training rules using target domain data associated with the target energy storage battery and source domain data associated with the sample energy storage battery; the preset state of charge prediction model at least includes a feature extraction module based on the Att-BiLSTM architecture of multi-dimensional feature representation; and according to the target prediction result, determining the state of charge of the target energy storage battery.

[0184] See also Figure 8 As shown, the embodiment of this specification also provides a device for determining the state of charge of an energy storage battery, which may specifically include the following structural modules:

[0185] The acquisition module 801 may be specifically used to acquire the operation data of the target energy storage battery in the current preset time period;

[0186] The processing module 802 can be specifically used to process the operation data of the target energy storage battery in the current preset time period using a preset state of charge prediction model to obtain a corresponding target prediction result; wherein the preset state of charge prediction model is a neural network model based on an improved DA-GAN obtained by training according to preset training rules using target domain data associated with the target energy storage battery and source domain data associated with the sample energy storage battery; the preset state of charge prediction model at least includes a feature extraction module based on the Att-BiLSTM architecture of multi-dimensional feature representation;

[0187] The determination module 803 may be specifically configured to determine the state of charge of the target energy storage battery according to the target prediction result.

[0188] In some embodiments, the feature extraction module includes at least: a first LSTM layer and a second LSTM layer; wherein the first LSTM layer includes a plurality of first LSTM substructures connected in series in a forward time sequence, and the second LSTM layer includes a plurality of second LSTM substructures connected in series in a reverse time sequence;

[0189] The first LSTM substructure is also connected to a corresponding first attention unit, which is also connected to other first LSTM substructures in the first LSTM layer; the second LSTM substructure is also connected to a corresponding second attention unit, which is also connected to other second LSTM substructures in the second LSTM layer;

[0190] The first LSTM layer is used to extract corresponding process operation features; the second LSTM layer is used to extract corresponding initial operation features.

[0191] In some embodiments, the preset state of charge prediction model may further include: a domain feature splitting module, a domain discrimination module, and a state classification module;

[0192] The feature extraction module is connected to the domain feature splitting module, the domain feature splitting module is connected to the domain discrimination module and the state classification module, and the domain discrimination module is connected to the state classification module;

[0193] The state classification module includes: a first sub-classification network associated with the target domain, and a second sub-classification network associated with the source domain, wherein the first sub-classification network and the second sub-classification network have the same structure and share parameters.

[0194] In some embodiments, when the above-mentioned processing module 803 is specifically implemented, the operation data of the current preset time period of the target energy storage battery can be processed by using the preset state of charge prediction model in the following manner to obtain the corresponding target prediction result: the operation data of the current preset time period of the target energy storage battery is processed by using the feature extraction module to obtain the initial operation characteristics of the current preset time period and the process operation characteristics of the current preset time period; the common characteristics and independent characteristics are separated from the initial operation characteristics of the current preset time period and the process operation characteristics of the current preset time period by using the domain feature splitting module; wherein the independent characteristics include: a first independent feature unique to the target domain and / or a second independent feature unique to the source domain; the common characteristics and independent characteristics are processed by using the domain discrimination module to obtain the corresponding domain discrimination result; when the domain discrimination result indicates the target domain, the domain discrimination result, the common characteristics, and the first independent characteristics are sent to the first sub-classification network in the state classification module; the first sub-classification network is used to process the domain discrimination result, the common characteristics, and the first independent characteristics to obtain and output the corresponding target prediction result.

[0195] In some embodiments, the operating data may include at least: voltage, current, temperature, etc.

[0196] In some embodiments, before obtaining the operating data of the target energy storage battery for the current preset time period, the device can also be used to: construct an initial feature representation and state prediction model according to a preset training rule; wherein the initial feature representation and state prediction model includes a connected initial feature representation model and an initial state prediction model; the initial feature representation model includes at least: an initial first LSTM layer, and an initial second LSTM layer; use the source domain data associated with the sample energy storage battery to construct a corresponding first training set; according to the preset training rules, use the first training set to train the initial feature representation and state prediction model to obtain a target feature representation and state prediction model that meets the requirements; according to the preset training rules, construct an initial state of charge prediction model based on the improved DA-GAN; wherein the initial state of charge prediction model includes an initial feature extraction module, an initial domain extraction module, and an initial domain extraction module. Domain feature splitting module, initial domain discrimination module, initial state classification module; migrating the parameters of the target feature representation model in the target feature representation and state prediction model to the initial feature extraction module in the initial state of charge prediction model, migrating the parameters of the target state prediction model in the target feature representation and state prediction model to the initial state classification module in the initial state of charge prediction model; wherein, the feature extraction module in the initial state of charge prediction model has the same structure as the target feature representation model, and the state classification module in the initial state of charge prediction model has the same partial structure as the target state prediction model; combining the source domain data associated with the sample energy storage battery and the target domain data associated with the target energy storage battery to construct a corresponding second training set; according to the preset training rules, using the second training set and the initial state of charge prediction model, multiple rounds of model training are performed to obtain a state of charge prediction model that meets the preset requirements.

[0197] In some embodiments, when the device is implemented, the model training of the current round in multiple rounds of model training can be performed in the following manner: obtain the state of charge prediction model of the previous round; use the state of charge prediction model of the previous round to process the second training data of the current round to obtain corresponding first-category prediction results and second-category prediction results; wherein the first-category prediction results are state prediction results output by the first subclassification network, and the second-category prediction results are state prediction results output by the second subclassification network; according to preset training rules, use the second training data of the current round, the first-category prediction results, and the second-category prediction results to calculate the function value of the target loss function of the current round; according to the function value of the target loss function of the current round, adjust the model parameters of the state of charge prediction model of the previous round to obtain the state of charge prediction model of the current round.

[0198] In some embodiments, the target loss function may include an error function based on the state of charge of the source domain data and a discriminator loss function based on a generative adversarial network.

[0199] Based on the above-mentioned target loss function, during the model training process, when the domain discrimination module discriminates the source domain data as the target domain, or discriminates the target domain as the source domain, the discriminator loss function takes effect, and the parameters of the next round of domain discrimination module are adjusted and corrected; when the domain discrimination module discriminates the source domain data as the source domain, the independent features output by the domain feature splitting module are used as the second independent features, and are sent to the second sub-classification network in the state classification module together with the common features. When the state of charge output by the second sub-classification network is inconsistent with the true label, the error function takes effect, and the parameters of the next round of domain feature splitting module are adjusted and corrected.

[0200] It should be noted that the units, devices or modules described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. For the convenience of description, the above devices are described separately by functions divided into various modules. Of course, when implementing this specification, the functions of each module can be implemented in the same or more software and / or hardware, or the modules that implement the same function can be implemented by a combination of multiple sub-modules or sub-units. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0201] It can be seen from the above that the device for determining the state of charge of an energy storage battery provided in the embodiments of this specification can train a preset state of charge prediction model with high accuracy and good generalization at a relatively low data processing cost, and then can efficiently and accurately determine the state of charge of the target energy storage battery based on the preset state of charge prediction model, effectively reduce data processing errors, and reduce the overall data processing cost.

[0202] In a specific scenario example, the method for determining the state of charge of an energy storage battery provided in this specification can be applied to estimate the state of charge of an energy storage lithium battery based on feature splitting of the source domain and the target domain. The specific implementation process may include the following contents.

[0203] In this scenario example, considering that there are often many limitations when estimating the state of charge of energy storage lithium batteries based on existing methods: 1. PCA dimensionality reduction relies on linear transformation and may not be able to handle nonlinear relationships well; 2. Traditional LSTM networks have long memory loss problems (gradient disappearance); 3. The MMD function is based on Euclidean distance and may not be able to capture the structural differences between data well when dealing with nonlinear problems. Since SOC estimation is a nonlinear fitting of multiple characteristics in principle, and the data distribution has a nonlinear structure, the features after PCA dimensionality reduction cannot fully reveal the intrinsic structure of the original multidimensional data, and it is easy to accumulate errors to the LSTM network, thereby affecting the prediction; similarly, for such nonlinear problems, the MMD function may not accurately measure the difference between the source domain and the target domain. In addition, in the SOC estimation task with long-distance dependence, the LSTM network with the gradient disappearance problem is not suitable as a feature extractor and label (SOC) predictor, which also has adverse effects on model migration and domain adaptation. In addition, the above methods rarely consider the influence of the SOC value at the starting moment on the subsequent SOC estimation, but the classic ampere-hour integration method in the traditional method clearly depends on the initial SOC value, so the importance of the initial value and process data cannot be ignored.

[0204] In this scenario example, we consider a method for estimating the state of charge of energy storage lithium batteries based on feature splitting of the source domain and the target domain. We use the bidirectional long short-term memory network (Att-BiLSTM) combined with the attention mechanism to extract the initial features and process features of the laboratory data in the source domain (or source domain) and the energy storage power station operating data in the target domain (or target domain). We introduce a domain feature splitting module into the domain adaptation framework to separate the common features and characteristic features of the source domain and the target domain. Based on the idea of ​​different branches in different domains, we distinguish the differences between laboratory and operating batteries while achieving domain adaptation. This method can not only effectively handle the data distribution differences between the source domain and the target domain in the absence of target data labels, but also improve the accuracy and generalization ability of the state of charge estimation of energy storage lithium batteries, and has better performance and applicability in practical applications. In turn, we can effectively handle the differences between laboratory and operating batteries, achieve better domain adaptation capabilities, and thus improve the performance of the model in the target domain. In addition, when the SOC data in the target domain is missing, the model is trained using only the labeled data in the source domain and the unlabeled data in the target domain, thereby reducing the cost and complexity of labeling the operating data of the energy storage power station.

[0205] Before specific implementation, a multidimensional feature extraction model combined with attention (e.g., a feature representation model) can be established to simultaneously extract the starting features (e.g., initial operation features) and process features (e.g., process operation features) of a battery operation data segment (e.g., operation data of a preset time period), and obtain a feature representation model of the source domain data by training on the source domain data.

[0206] Specifically, a bidirectional feature extraction module (Att-BiLSTM) combined with attention (e.g., feature extraction module, feature representation model) can be first established to simultaneously extract the starting features and process features of the battery operation data segment; and combined with the SOC label prediction module (e.g., state classification module, state prediction model), feature learning and representation of source domain data can be achieved through training on source domain data. Then, an improved framework of the domain adaptation method based on generative adversarial network (DA-GAN) is built, including the feature extraction module, label prediction module, domain feature splitting module, and domain discrimination module after the above training parameter migration. Through training on source domain data and target domain data, domain commonality migration and feature adaptation (obtaining a preset state of charge prediction model) can be achieved.

[0207] In specific implementation, the domain adaptation framework based on generative adversarial network (DA-GAN) can be used to achieve domain commonality migration and feature adaptation. The model mainly consists of four parts: feature extraction module, domain feature splitting module, domain discrimination module, and label prediction module. Its core lies in the domain feature splitting module and the data flow of different branches in different fields.

[0208] Among them, the feature extraction module is used to receive source domain and target domain data, and is responsible for extracting features from the received data. This module and its parameters are migrated from the Att-BiLSTM module trained in the source domain;

[0209] The domain feature splitting module further splits the features to extract domain-invariant features (common features) related to SOC estimation and source / target domain independent features (characteristic features) that are essentially irrelevant to SOC estimation;

[0210] The domain discrimination module is used to determine whether the input features come from the source domain or the target domain, thereby determining the data flow direction, that is, which domain's characteristic features need to be sent to the label prediction module;

[0211] The label prediction module is used to receive the independent characteristic features of the two domains, the shared common features, and the results of the domain discrimination module (source domain / target domain). It is responsible for integrating the features of the corresponding domains and completing the prediction of the model label (SOC value).

[0212] During the training process of the model, the framework roughly follows the discriminator loss and adversarial loss of the Generative Adversarial Network (GAN).

[0213] In the specific implementation, for the feature extraction module combined with attention (or called bidirectional feature extraction module), please refer to Figure 3 As shown in the figure. Since the battery SOC represents the state of charge of the battery (i.e., the percentage of power), its change pattern and the correlation with voltage, current, temperature, etc. are long-term changes, so the introduction of LSTM makes it easier to learn the long-term change relationship.

[0214] The estimation of battery SOC depends on the initial value and process data, both of which are indispensable. However, for the actual operating data obtained (usually a certain fragment), it is not only difficult to obtain the label of the target moment (i.e., the true SOC value), but also difficult to know the true SOC value at the starting moment of the fragment. The previous battery SOC migration estimation methods based on domain adaptation have failed to fully consider the influence of the initial value. Therefore, the present invention believes that it is necessary to process the initial features and process features at the same time. Since LSTM is more suitable for processing a single task, the present invention constructs a bidirectional neural network BiLSTM to extract the starting features and global features of the battery data fragment at the same time. The constructed feature extraction model consists of two independent LSTM structures in opposite directions, which respectively establish long-distance dependencies between forward and reverse time series, where the LSTM1 structure extracts process features from the forward sequence of the time series, and the LSTM2 structure reverses the starting features from the reverse sequence of the time series, and introduces the attention mechanism in both, so that the model is more likely to focus on more valuable features and solve the gradient disappearance problem of traditional LSTM.

[0215] Among them, this module can be called Att-BiLSTM, which is used for feature representation and feature extraction. Its input is the original battery voltage, current, and temperature data, and its output is the bidirectional feature of interest. The purpose of this process is to process the original input data into feature data that cannot directly determine the data domain, so as to achieve reasonable "abstraction".

[0216] After the model is trained on the source domain data to learn the feature representation, the frozen parameters are migrated to the feature extraction module of the domain adaptation framework.

[0217] For the specific implementation, please refer to Fig. 9 As shown in the figure. The SOC label prediction module consists of a fully connected layer (FC) and a Softmax regression layer on different time dimensions, which is used to predict the SOC value. The Att-BiLSTM bidirectional feature extraction module is connected in series with this module and trained on the source domain data to minimize the following model error (mean square error MSE):

[0218]

[0219] After the model is trained on the source domain data to learn the relevant parameters of the SOC regression function, the frozen parameters are migrated to the label prediction module of the domain adaptation framework.

[0220] In specific implementation, for the domain feature splitting module, the above-mentioned domain feature splitting module can split the features obtained from Att-BiLSTM, extract domain-invariant features (common features) related to SOC estimation, and source domain / target domain independent features (characteristic features) that are essentially irrelevant to SOC estimation.

[0221] In specific implementation, the domain discrimination module is used to determine whether the input features come from the source domain or the target domain, and decides which domain's characteristic features need to be sent to the label prediction module. This module consists of a binary fully connected layer (FC) and a softmax layer, and the objective function is the discriminator loss commonly used in generative adversarial networks.

[0222] Based on the above considerations, please refer to Fig.10 As shown, when performing charge state estimation, the following steps may be included:

[0223] Step 1: Data preparation. Collect and prepare energy storage lithium battery data in the source domain and target domain. These data include current, voltage, and temperature. The source domain data is the battery operation data measured in the laboratory, including current, voltage, and temperature curves, and the corresponding real SOC curve can be clearly known; the target domain data is the current, voltage, and temperature curves collected during the online operation of the energy storage battery, and there is no real SOC data.

[0224] Step 2: Multi-feature extraction model training: By training the Att-BiLSTM model, the feature representation and state prediction of the source domain data are learned.

[0225] Step 3: DA-GAM framework training. The parameters of step 2 are transferred to the initial module of the DA-GAN framework and trained using data from the source and target domains.

[0226] Step 4: State of charge estimation: Use the trained model to estimate the state of the new energy storage lithium battery data.

[0227] Through the above scenario examples, it is verified that the method for determining the state of charge of the energy storage battery provided in this specification has the following advantages: 1) The attention BiLSTM structure is adopted to extract process features through the forward dependency of the time series, and the starting features are extracted through the reverse dependency, fully considering the importance of the initial value, and comprehensively analyzing the battery operation. 2) The common features and independent features of the source domain / target domain are extracted, and the feature changes and differences between the source domain and the target domain are effectively learned, so as to achieve better domain adaptation capabilities and improve the performance of the model in the target domain. 3) No labeled data of the target domain is required, and only unlabeled data of the source domain and the target domain are used for training, which reasonably reduces the labeling cost of the target energy storage battery data and the application cost of the model, and at the same time improves the generalization ability of the energy storage lithium battery state of charge estimation model. 4) It is suitable for processing situations where there are large differences or complexities between different operating condition data sets, and can effectively deal with data distribution changes and domain transfer problems in actual energy storage power stations.

[0228] Although the present specification provides method operation steps as described in the embodiments or flow charts, more or less operation steps may be included based on conventional or non-creative means. The order of steps listed in the embodiments is only one way of executing the order of many steps, and does not represent a unique execution order. When the device or client product in practice is executed, it can be executed in sequence or in parallel according to the method shown in the embodiments or the drawings (for example, a parallel processor or a multi-threaded processing environment, or even a distributed data processing environment). The term "include", "include" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, product or device including a series of elements includes not only those elements, but also includes other elements that are not explicitly listed, or also includes elements inherent to such a process, method, product or device. In the absence of more restrictions, it is not excluded that there are other identical or equivalent elements in the process, method, product or device including the elements. The first, second, etc. words are used to represent the name, and do not represent any particular order.

[0229] Those skilled in the art also know that, in addition to implementing the controller in a purely computer-readable program code, the controller can be made to implement the same function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, such a controller can be considered as a hardware component, and the devices for implementing various functions included therein can also be considered as structures within the hardware component. Or even, the devices for implementing various functions can be considered as both software modules for implementing the method and structures within the hardware component.

[0230] This specification may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, classes, etc. that perform specific tasks or implement specific abstract data types. This specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules may be located in local and remote computer-readable storage media including storage devices.

[0231] Through the description of the above embodiments, it can be known that those skilled in the art can clearly understand that the present specification can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present specification can essentially be embodied in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a mobile terminal, a server, or a network device, etc.) to execute the methods described in each embodiment of the present specification or some parts of the embodiments.

[0232] Although the present specification is described through embodiments, those skilled in the art will appreciate that there are many modifications and changes to the present specification without departing from the spirit of the present specification, and it is intended that the appended claims include these modifications and changes without departing from the spirit of the present specification.

Claims

1. A method for determining the state of charge of an energy storage battery, characterized in that: include: Obtaining the operating data of the target energy storage battery during the current preset time period; The operation data of the target energy storage battery in the current preset time period is processed by using a preset state of charge prediction model to obtain a corresponding target prediction result; wherein the preset state of charge prediction model is a neural network model based on an improved DA-GAN obtained by training according to preset training rules using target domain data associated with the target energy storage battery and source domain data associated with the sample energy storage battery; the preset state of charge prediction model at least includes a feature extraction module based on the Att-BiLSTM architecture of multi-dimensional feature representation; According to the target prediction results, the charge state of the target energy storage battery is determined; Before obtaining the operating data of the target energy storage battery in the current preset time period, the method also includes: constructing an initial feature representation and state prediction model according to a preset training rule; wherein the initial feature representation and state prediction model includes a connected initial feature representation model and an initial state prediction model; the initial feature representation model includes at least: an initial first LSTM layer, and an initial second LSTM layer; using source domain data associated with the sample energy storage battery, constructing a corresponding first training set; according to the preset training rules, using the first training set to train the initial feature representation and state prediction model to obtain a target feature representation and state prediction model that meets the requirements; according to the preset training rules, constructing an initial state of charge prediction model based on the improved DA-GAN; wherein the initial state of charge prediction model includes an initial feature extraction module, an initial domain feature splitting module , an initial domain discrimination module, an initial state classification module; migrating the parameters of the target feature representation model in the target feature representation and state prediction model to the initial feature extraction module in the initial state of charge prediction model, and migrating the parameters of the target state prediction model in the target feature representation and state prediction model to the initial state classification module in the initial state of charge prediction model; wherein, the feature extraction module in the initial state of charge prediction model has the same structure as the target feature representation model, and the state classification module in the initial state of charge prediction model has the same partial structure as the target state prediction model; combining the source domain data associated with the sample energy storage battery and the target domain data associated with the target energy storage battery to construct a corresponding second training set; according to the preset training rules, using the second training set and the initial state of charge prediction model, multiple rounds of model training are performed to obtain a state of charge prediction model that meets the preset requirements.

2. The method according to claim 1, characterized in that The feature extraction module at least includes: a first LSTM layer and a second LSTM layer; wherein the first LSTM layer includes a plurality of first LSTM substructures connected in series in a forward time sequence, and the second LSTM layer includes a plurality of second LSTM substructures connected in series in a reverse time sequence; The first LSTM substructure is also connected to a corresponding first attention unit, which is also connected to other first LSTM substructures in the first LSTM layer; the second LSTM substructure is also connected to a corresponding second attention unit, which is also connected to other second LSTM substructures in the second LSTM layer; The first LSTM layer is used to extract corresponding process operation features; the second LSTM layer is used to extract corresponding initial operation features.

3. The method according to claim 2, characterized in that The preset state of charge prediction model also includes: a domain feature splitting module, a domain discrimination module, and a state classification module; The feature extraction module is connected to the domain feature splitting module, the domain feature splitting module is connected to the domain discrimination module and the state classification module, and the domain discrimination module is connected to the state classification module; The state classification module includes: a first sub-classification network associated with the target domain, and a second sub-classification network associated with the source domain, wherein the first sub-classification network and the second sub-classification network have the same structure and share parameters.

4. The method according to claim 3, characterized in that The preset state of charge prediction model is used to process the operating data of the target energy storage battery in the current preset time period to obtain the corresponding target prediction result, including: Using the feature extraction module to process the operating data of the target energy storage battery in the current preset time period, obtain the initial operating characteristics of the current preset time period, and the process operating characteristics of the current preset time period; Using the domain feature splitting module, common features and independent features are split from the initial operation features of the current preset time period and the process operation features of the current preset time period; wherein the independent features include: a first independent feature unique to the target domain and / or a second independent feature unique to the source domain; Use the domain discrimination module to process common features and independent features to obtain the corresponding domain discrimination results; When the domain discrimination result indicates the target domain, the domain discrimination result, the common feature, and the first independent feature are sent to a first sub-classification network in the state classification module; The first sub-classification network is used to process the domain discrimination result, the common feature, and the first independent feature to obtain and output the corresponding target prediction result.

5. The method according to claim 1, characterized in that The operating data at least includes: voltage, current, and temperature.

6. The method according to claim 1, characterized in that According to the preset training rules, multiple rounds of model training are performed using the second training set and the initial state of charge prediction model, including: Perform the current round of model training as follows: Obtain the state of charge prediction model of the previous round; The state of charge prediction model of the previous round is used to process the second training data of the current round to obtain corresponding first-category prediction results and second-category prediction results; wherein the first-category prediction results are state prediction results output by the first sub-classification network, and the second-category prediction results are state prediction results output by the second sub-classification network; According to the preset training rules, the function value of the target loss function of the current round is calculated using the second training data of the current round, the first type of prediction results, and the second type of prediction results; According to the function value of the target loss function of the current round, the model parameters of the state of charge prediction model of the previous round are adjusted to obtain the state of charge prediction model of the current round.

7. The method according to claim 6, characterized in that The target loss function includes an error function based on the state of charge of the source domain data and a discriminator loss function based on a generative adversarial network.

8. A device for determining the state of charge of an energy storage battery, characterized in that: include: An acquisition module, used to acquire the operating data of the target energy storage battery during the current preset time period; A processing module, used to process the operating data of the target energy storage battery in the current preset time period using a preset state of charge prediction model to obtain a corresponding target prediction result; wherein the preset state of charge prediction model is a neural network model based on an improved DA-GAN obtained by training according to preset training rules using target domain data associated with the target energy storage battery and source domain data associated with the sample energy storage battery; the preset state of charge prediction model at least includes a feature extraction module based on the Att-BiLSTM architecture of multi-dimensional feature representation; A determination module is used to determine the charge state of the target energy storage battery according to the target prediction result; Before obtaining the operating data of the target energy storage battery in the current preset time period, the device is also used to: construct an initial feature representation and state prediction model according to a preset training rule; wherein the initial feature representation and state prediction model includes a connected initial feature representation model and an initial state prediction model; the initial feature representation model at least includes: an initial first LSTM layer, and an initial second LSTM layer; use the source domain data associated with the sample energy storage battery to construct a corresponding first training set; according to the preset training rules, use the first training set to train the initial feature representation and state prediction model to obtain a target feature representation and state prediction model that meets the requirements; according to the preset training rules, construct an initial state of charge prediction model based on the improved DA-GAN; wherein the initial state of charge prediction model includes an initial feature extraction module, an initial domain feature splitting module , an initial domain discrimination module, an initial state classification module; migrating the parameters of the target feature representation model in the target feature representation and state prediction model to the initial feature extraction module in the initial state of charge prediction model, and migrating the parameters of the target state prediction model in the target feature representation and state prediction model to the initial state classification module in the initial state of charge prediction model; wherein, the feature extraction module in the initial state of charge prediction model has the same structure as the target feature representation model, and the state classification module in the initial state of charge prediction model has the same partial structure as the target state prediction model; combining the source domain data associated with the sample energy storage battery and the target domain data associated with the target energy storage battery to construct a corresponding second training set; according to the preset training rules, using the second training set and the initial state of charge prediction model, multiple rounds of model training are performed to obtain a state of charge prediction model that meets the preset requirements.

9. A computer-readable storage medium, characterized in that: Computer instructions are stored thereon, and when the instructions are executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images