A deep learning prediction method and device for battery system voltage extreme difference

Abstract

The present invention belongs to the technical field of energy storage batteries, and discloses a deep learning prediction method and device for voltage extreme difference of a battery system; the method comprises: collecting battery operation data of a battery cluster; the battery operation data comprises single cell extreme difference data, battery cluster charge and discharge power, and battery cluster total voltage data; the battery operation data is input into a pre-trained deep learning prediction model for prediction, and a prediction result of voltage extreme difference of the battery cluster within a preset time in the future is obtained; wherein the pre-trained deep learning prediction model is a BiGRU‑Transformer‑LSTM model. The present invention adopts a hybrid model of three different architectures of BiGRU‑Transformer‑LSTM, which can effectively predict the change of battery voltage extreme difference, and can predict the change of voltage extreme difference in the future period of time based on the current voltage and power information, and has a voltage consistency evaluation function; the method is helpful to rationally plan the charging and discharging strategy of the battery pack, extend the battery life, and improve the safety and reliability of the battery.

Description

Technical Field

[0001] The present invention belongs to the technical field of energy storage batteries, and in particular relates to a deep learning prediction method and device for voltage extreme difference of a battery system. Background Art

[0002] Nowadays, battery systems are widely used in electric vehicles, energy storage power stations and other systems, and their safe and reliable operation has attracted much attention. In the actual charging and discharging process, the multiple single cells that make up the battery system will have voltage differences due to individual differences such as manufacturing process and ambient temperature. The difference between the maximum voltage and the minimum voltage of all single cells in a battery pack at a certain moment is called the voltage range.

[0003] Technically speaking, the voltage extreme difference of the battery system can be obtained by the Battery Management System (BMS). If the number of single cells is large, the acquisition of the voltage extreme difference will be overly dependent on the software and hardware system, and will inevitably be affected by noise and interference signals. At this time, if the voltage extreme difference of the battery system is measured to be beyond the limit, and then the balancing, protection and other actions are performed, it will be far from meeting the requirements of system protection. Furthermore, the actual BMS voltage balancing function (including active and passive balancing strategies) is not frequently executed, but generally performed at a specific stage, such as the end of charging. Frequent excessive fluctuations and over-limits of the battery voltage extreme difference will pose a threat to the safe and reliable operation of the battery system. Therefore, predicting and evaluating the voltage extreme difference fluctuations of the battery system to predict risks in advance and control them can effectively ensure the safe operation of the battery system.

[0004] The prediction of voltage extreme difference is inseparable from the prediction of battery system single cell voltage. However, there are many factors that affect the change of single cell voltage, including charge and discharge power, total voltage, temperature, impedance, SOC (State of Charge, SOC), SOH (State of Health, SOH), etc. Therefore, it is very difficult to predict through traditional modeling methods. Summary of the invention

[0005] The purpose of the present invention is to provide a deep learning prediction method and device for battery system voltage extreme difference, so as to solve the technical problem that the existing battery system lacks prediction of the voltage difference fluctuation of a large number of single cells, thereby improving the battery system's ability to evaluate and predict voltage consistency.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a deep learning prediction method for battery system voltage extreme difference, comprising:

[0008] Collecting battery operation data of the battery cluster; the battery operation data includes single cell extreme difference data, battery cluster charge and discharge power and battery cluster total voltage data;

[0009] Input the battery operation data into a pre-trained deep learning prediction model for prediction, and obtain a voltage range prediction result of the battery cluster within a preset time in the future;

[0010] Wherein, the pre-trained deep learning prediction model is a BiGRU-Transformer-LSTM model;

[0011] The BiGRU-Transformer-LSTM model includes a BiGRU layer, a Transformer layer, an LSTM layer, a fully connected layer and an activation function connected in sequence;

[0012] The BiGRU-Transformer-LSTM model includes a BiGRU layer, a Transformer layer, an LSTM layer, a fully connected layer and an activation function connected in sequence;

[0013] The BiGRU layer is used to perform forward and reverse encoding on the input time series data, capture bidirectional dependency information, and output a feature sequence;

[0014] The Transformer layer is used to receive the output of the BiGRU layer, use the multi-head self-attention mechanism to perform global feature modeling on all positions of the time series, and output the feature sequence after global feature modeling;

[0015] The LSTM layer is a double-layer LSTM layer, including the first LSTM layer and the second LSTM layer. The first LSTM layer is used to further extract the time dependency of the feature sequence after global feature modeling, and output the time step features after the first LSTM layer processing. The second LSTM layer is used to receive the output of the first LSTM layer and generate the final output features in combination with the long-term dependency features.

[0016] Fully connected layer and activation function: used to map the final output features of the LSTM layer into prediction results.

[0017] A further improvement of the present invention is that the pre-trained deep learning prediction model is obtained by training through the following steps:

[0018] Acquire a historical data set of a battery cluster of an energy storage power station, the historical data set including: single cell extreme difference data, battery cluster charge and discharge power, and battery cluster total voltage data; divide the historical data set into N parts in chronological order to obtain N historical data subsets;

[0019] N-1 of the N historical data subsets are used as training sets, and the remaining 1 is used as validation and test sets; an identical initial deep learning prediction model is established for each of the N-1 historical data subsets in the training set;

[0020] Based on each historical data subset in the training set, a corresponding initial deep learning prediction model is iteratively trained to obtain N-1 trained initial deep learning prediction models;

[0021] Each network parameter in the N-1 trained initial deep learning prediction models is weighted averaged according to the performance of the validation set, and the weight is determined inversely by the prediction error of each trained initial deep learning prediction model on the validation set; the parameter value after calculating the weighted average is assigned to the deep learning prediction model to generate a pre-trained deep learning prediction model.

[0022] A further improvement of the present invention is that both the first layer LSTM and the second layer LSTM include a gating mechanism.

[0023] A further improvement of the present invention is that the BiGRU-Transformer-LSTM model further includes:

[0024] The dynamic smoothing layer is used to add dynamic smoothing processing after the fully connected layer and the activation function to further eliminate noise in the prediction results and optimize the stability of the prediction curve to obtain the final prediction results.

[0025] A further improvement of the present invention is that: the activation function uses the LeakyReLU activation function.

[0026] A further improvement of the present invention is that the BiGRU layer is specifically used to receive input time series data and perform forward and reverse encoding to capture the bidirectional dependency information of the time series and output the feature sequence ; The input time series data is X={x1,x2,…,x T}, where T represents the length of the time step, x t is the input feature vector at the tth time step;

[0027]

[0028] Among them, h t is the bidirectional feature representation generated at time step t;

[0029] The Transformer layer is specifically used to receive the feature sequence H output by the BiGRU layer BiGRU , use the multi-head self-attention mechanism to perform global feature modeling on all positions of the time series, and output the feature sequence after global feature modeling ;

[0030]

[0031] in, represents the global feature representation of time step t after being processed by the Transformer layer;

[0032] The first layer of LSTM is used to receive the feature sequence of the Transformer layer. , and further extract the time dependency of the feature sequence after global feature modeling, and output the time step feature L1 after the first layer of LSTM processing;

[0033]

[0034] The second layer LSTM is specifically used to receive the output L1 of the first layer LSTM and combine it with the long-term dependency features to generate the final output feature L2;

[0035] .

[0036] A further improvement of the present invention is that the prediction result is generated by a fully connected layer and a LeakyReLU activation function:

[0037] .

[0038] In a second aspect, the present invention provides a deep learning prediction device for battery system voltage extreme difference, comprising:

[0039] A collection module is used to collect battery operation data of the battery cluster; the battery operation data includes single cell extreme difference data, battery cluster charge and discharge power and battery cluster total voltage data;

[0040] A prediction module is used to input the battery operation data into a pre-trained deep learning prediction model for prediction, and obtain a voltage range prediction result of the battery cluster within a preset time in the future;

[0041] Wherein, the pre-trained deep learning prediction model is a BiGRU-Transformer-LSTM model;

[0042] The BiGRU-Transformer-LSTM model includes a BiGRU layer, a Transformer layer, an LSTM layer, a fully connected layer and an activation function connected in sequence;

[0043] The BiGRU-Transformer-LSTM model includes a BiGRU layer, a Transformer layer, an LSTM layer, a fully connected layer and an activation function connected in sequence;

[0044] The BiGRU layer is used to perform forward and reverse encoding on the input time series data, capture bidirectional dependency information, and output a feature sequence;

[0045] The Transformer layer is used to receive the output of the BiGRU layer, use the multi-head self-attention mechanism to perform global feature modeling on all positions of the time series, and output the feature sequence after global feature modeling;

[0046] The LSTM layer is a double-layer LSTM layer, including the first LSTM layer and the second LSTM layer. The first LSTM layer is used to further extract the time dependency of the feature sequence after global feature modeling, and output the time step features after the first LSTM layer processing. The second LSTM layer is used to receive the output of the first LSTM layer and generate the final output features in combination with the long-term dependency features.

[0047] Fully connected layer and activation function: used to map the final output features of the LSTM layer into prediction results.

[0048] A further improvement of the present invention is that the pre-trained deep learning prediction model is obtained by training through the following steps:

[0049] Acquire a historical data set of a battery cluster of an energy storage power station, the historical data set including: single cell extreme difference data, battery cluster charge and discharge power, and battery cluster total voltage data; divide the historical data set into N parts in chronological order to obtain N historical data subsets;

[0050] N-1 of the N historical data subsets are used as training sets, and the remaining 1 is used as validation and test sets; an identical initial deep learning prediction model is established for each of the N-1 historical data subsets in the training set;

[0051] Based on each historical data subset in the training set, a corresponding initial deep learning prediction model is iteratively trained to obtain N-1 trained initial deep learning prediction models;

[0052] Each network parameter in the N-1 trained initial deep learning prediction models is weighted averaged according to the performance of the validation set, and the weight is determined inversely by the prediction error of each trained initial deep learning prediction model on the validation set; the parameter value after calculating the weighted average is assigned to the deep learning prediction model to generate a pre-trained deep learning prediction model.

[0053] A further improvement of the present invention is that both the first layer LSTM and the second layer LSTM include a gating mechanism.

[0054] A further improvement of the present invention is that the BiGRU-Transformer-LSTM model further includes:

[0055] The dynamic smoothing layer is used to add dynamic smoothing processing after the fully connected layer and the activation function to further eliminate noise in the prediction results and optimize the stability of the prediction curve to obtain the final prediction results.

[0056] A further improvement of the present invention is that: the activation function uses the LeakyReLU activation function.

[0057] A further improvement of the present invention is that the BiGRU layer is specifically used to receive input time series data and perform forward and reverse encoding to capture the bidirectional dependency information of the time series and output the feature sequence ; The input time series data is X={x1,x2,…,x T}, where T represents the length of the time step, x t is the input feature vector at the tth time step;

[0058]

[0059] Among them, h t is the bidirectional feature representation generated at time step t;

[0060] The Transformer layer is specifically used to receive the feature sequence H output by the BiGRU layer BiGRU , use the multi-head self-attention mechanism to perform global feature modeling on all positions of the time series, and output the feature sequence after global feature modeling ;

[0061]

[0062] in, represents the global feature representation of time step t after being processed by the Transformer layer;

[0063] The first layer of LSTM is used to receive the feature sequence of the Transformer layer. , and further extract the time dependency of the feature sequence after global feature modeling, and output the time step feature L1 after the first layer of LSTM processing;

[0064]

[0065] The second layer LSTM is specifically used to receive the output L1 of the first layer LSTM and combine it with the long-term dependency features to generate the final output feature L2;

[0066] .

[0067] A further improvement of the present invention is that the prediction result is generated by a fully connected layer and a LeakyReLU activation function:

[0068] .

[0069] In a third aspect, the present invention provides an electronic device comprising a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the deep learning prediction method for battery system voltage extreme difference.

[0070] In a fourth aspect, the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores at least one instruction, and when the at least one instruction is executed by a processor, the deep learning prediction method for battery system voltage extreme difference is implemented.

[0071] Compared with the prior art, the present invention has the following beneficial effects:

[0072] In view of the voltage extreme fluctuations of a large number of single cells in multiple battery clusters in large-capacity energy storage power stations, if they frequently exceed the limit, the battery life will be shortened, causing reliability and safety problems in battery use; the present invention provides a deep learning prediction method for the voltage extreme difference of a battery system, including: collecting battery operation data of a battery cluster; the battery operation data includes single cell extreme difference data, battery cluster charge and discharge power, and battery cluster total voltage data; the battery operation data is input into a pre-trained deep learning prediction model for prediction, and the voltage extreme difference prediction result of the battery cluster within a preset time in the future is obtained; wherein the pre-trained deep learning prediction model is a BiGRU-Transformer-LSTM model. The present invention adopts a hybrid model of three different architectures of BiGRU-Transformer-LSTM, which can effectively predict the changes in battery voltage extreme difference, and can predict the voltage extreme difference changes in the future period of time based on the current voltage and power information, and has a voltage consistency evaluation function.

[0073] The present invention provides a deep learning prediction method for battery system voltage extreme difference, which makes full use of deep learning technology and makes improvements to achieve accurate prediction of battery voltage extreme difference, and makes it have strong generalization ability, and can perform online deduction prediction of voltage extreme difference and voltage consistency analysis. The method helps to rationally plan the charging and discharging strategy of the battery pack, prolong the battery life, and improve the safety and reliability of the battery. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings:

[0075] Figure 1 A schematic flow chart of a deep learning prediction method for voltage extreme difference of a battery system according to an embodiment of the present invention;

[0076] Figure 2 A schematic diagram of the training process of a BiGRU-Transformer-LSTM model according to an embodiment of the present invention;

[0077] Figure 3 A schematic flow chart of a deep learning prediction method for voltage extreme difference of a battery system according to another embodiment of the present invention;

[0078] Figure 4 A comparison chart of prediction results of different neural network models in the present invention;

[0079] Figure 5 This is a schematic diagram of the structure of a deep learning prediction device for battery system voltage extreme difference according to an embodiment of the present invention;

[0080] Figure 6 The figure is a schematic structural diagram of an electronic device of the present invention. DETAILED DESCRIPTION

[0081] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that the embodiments and features in the embodiments of the present invention can be combined with each other without conflict.

[0082] The following detailed description is an exemplary description, which is intended to provide further detailed description of the present invention. Unless otherwise specified, all technical terms used in the present invention have the same meaning as those generally understood by those skilled in the art to which the present invention belongs. The terms used in the present invention are only for describing specific embodiments, and are not intended to limit the exemplary embodiments according to the present invention.

[0083] Data-driven prediction technology has become the preferred method for voltage extreme difference prediction. The most common data-driven technology is deep learning (DL). The powerful learning ability of deep learning enables it to better process time series data and understand the dynamic characteristics of batteries. It has been successfully applied to the field of energy storage battery status prediction, including SOH, battery remaining useful life (RUL), SOC and other aspects. However, deep learning algorithms still have some problems in battery status prediction, such as prediction stability, generalization ability, and portability, and online applications are difficult to achieve.

[0084] In response to the problem of voltage extreme difference in battery systems, an embodiment of the present invention provides a deep learning prediction method for voltage extreme difference in battery systems. By preprocessing the battery system operation data and taking voltage extreme difference as the prediction target, feature engineering is constructed. The data features are divided into N parts, of which N-1 parts are used for model training and the remaining 1 part is used for verification and testing. In each round of training, a deep learning prediction model is constructed, and the optimal parameters of the model are obtained through an optimization algorithm. The corresponding voltage extreme difference prediction model library is trained to achieve online and multi-step prediction through intelligent selection and calling of model files, and a comprehensive evaluation of the voltage consistency state of the battery system can be performed. The present invention designs a corresponding prediction device, including modules such as data acquisition and transmission, data preprocessing and storage, and data analysis and state evaluation. The present invention has high accuracy and generalization ability for the prediction of voltage extreme difference, which helps to extend the service life of various energy storage systems and improve safety and reliability.

[0085] The embodiment of the present invention provides a deep learning prediction method for battery system voltage extreme difference, which belongs to the technical field of energy storage batteries and includes the following steps:

[0086] 1. Obtain historical data of the battery pack (cluster); the historical data at least includes: charge and discharge power, total voltage and voltage range; perform data preprocessing on the historical data of the battery pack (cluster), create feature engineering around the target value, and establish a deep learning prediction model with the charge and discharge power and total voltage of the battery pack (cluster) as input and the voltage range as the prediction target.

[0087] 2. Divide the battery pack (cluster) historical data into N parts according to the operating data characteristics (including historical charge and discharge power, total voltage, and time cycle characteristics), of which N-1 parts are used for deep learning prediction model training, and the last part is used as a verification and test set. N rounds of training are performed to obtain N model files. At the same time, in the N rounds of model training, the network model parameters are continuously adjusted to achieve the optimal prediction effect, and the corresponding model is saved.

[0088] 3. Use N models that can fully reflect different data characteristics to make target predictions, and according to the current measured voltage range information, continuously correct the model input and other adjustable parameters to achieve rolling optimization of the prediction model parameters, and finally achieve the prediction of the voltage range in the future. On this basis, combine the historical voltage range data, standard deviation and target prediction value to build a comprehensive scoring mechanism to achieve the voltage consistency evaluation of the battery cluster. This method has good prediction accuracy and generalization ability, and can quickly and accurately predict the voltage range for decision-making reference, which helps to extend the service life of the energy storage system and improve safety and reliability.

[0089] Specifically, see Figure 1 and Figure 2As shown, an embodiment of the present invention provides a deep learning prediction method for battery system voltage extreme difference, comprising the following steps:

[0090] Step 1: preprocess historical data and divide it into N parts: obtain a historical data set of the battery cluster of the energy storage power station, the historical data set includes: single cell extreme difference data, battery cluster charge and discharge power and battery cluster total voltage data; and preprocess the acquired single cell extreme difference, battery cluster charge and discharge power and battery cluster total voltage data of the battery cluster of the energy storage power station to obtain a preprocessed historical data set.

[0091] In a specific embodiment, the preprocessing can be implemented in the following exemplary manner: removing duplicate data, processing missing values, detecting and processing outliers; however, it should be understood by those skilled in the art that other methods can also be used to preprocess the raw data;

[0092] In a specific embodiment, deduplication of data may be implemented in the following exemplary manner: identifying duplicate records by timestamp and historical data value, and deleting the duplicate data;

[0093] In a specific implementation, processing missing values ​​may be implemented in the following exemplary manner: for missing data points, linear interpolation is used to fill them;

[0094] In a specific embodiment, the detection and processing of outliers can be implemented in the following exemplary manner: applying the Z-score method to identify numerical outliers, and adjusting or deleting the outliers to reduce noise and potential misleading analysis.

[0095] In a specific implementation, considering the time series characteristics of the feature quantity, the present invention divides the pre-processed historical data into N parts in chronological order to ensure that each piece of data is continuous in time.

[0096] Step 2: Prediction model construction: N-1 of the N historical data subsets are used as training sets for prediction model training, and the last one is used as a validation and test set. An initial deep learning prediction model is established for each of the N-1 historical data subsets in the training set to obtain N-1 identical initial deep learning prediction models.

[0097] Step 3: Model training and verification: Iteratively train a corresponding initial deep learning prediction model based on each historical data subset in the training set, until the preset maximum number of training times or the loss function converges, complete the training, and obtain the trained initial deep learning prediction model after verification using the verification and test sets; use N-1 historical data subsets to train a total of N-1 trained initial deep learning prediction models;

[0098] Step 4: Model transfer learning and model parameter self-update: Take the weighted average of each network parameter in the N-1 trained initial deep learning prediction models according to the performance of the validation set, and the weight is determined inversely by the prediction error of each model on the validation set. Calculate the parameter value after weighted average and assign it to the deep learning prediction model to generate a pre-trained deep learning prediction model.

[0099] Using the latest battery operation data of the energy storage power station battery cluster (single battery extreme difference data, battery cluster charge and discharge power and battery cluster total voltage data), the pre-trained deep learning prediction model is optimized through the transfer learning mechanism, and the final prediction model structure and weight are dynamically adjusted to adapt to the timing characteristics of voltage extreme difference fluctuations. On this basis, combined with the operation data characteristics (including historical charge and discharge power, total voltage, and time cycle characteristics), a multi-step prediction of future voltage extreme difference values ​​is completed, and finally the rolling update of the model and the dynamic evaluation of the battery system voltage consistency status are realized.

[0100] The present invention can effectively improve the generalization ability of the model through weighted averaging. Because the relationship between the characteristic quantity of the time series and the prediction target will change with the change of complex working conditions. Therefore, the N-fold crossover and weighted averaging method adopted can improve the generalization ability of the model.

[0101] In a specific implementation, the deep learning prediction model of the present invention is implemented using the following exemplary model: the BiGRU-Transformer-LSTM model is used; the BiGRU-Transformer-LSTM model is not a single, standardized deep learning model, but a hybrid model combining three different architectures of Bidirectional Gated Recurrent Unit (BiGRU), Transformer model (Transformer) and Long Short-Term Memory (LSTM) proposed by the present invention to solve the technical problem of the present invention. This hybrid model is designed to take advantage of the advantages of each component to solve a specific task; the following is an introduction to these three components and how they may be used in combination:

[0102] BiGRU is a variant of GRU (Gated Recurrent Unit) that introduces a bidirectional structure to better capture the bidirectional dependencies of sequence data. It consists of two independent GRU units, one that processes data forward in the time series and the other that processes data in the reverse direction of the time series. Through this bidirectional structure, BiGRU can capture both forward and backward information of sequence data, thereby enhancing the understanding and prediction of patterns in the sequence.

[0103] Transformer is a model based on the self-attention mechanism. It abandons the cyclic structure of the recurrent neural network (RNN) and relies entirely on the self-attention mechanism to process sequence data. Transformer can simultaneously focus on all positions in the sequence to capture global information and effectively capture long-term dependencies. In addition, Transformer has strong parallelization capabilities, which can significantly increase the training speed of the model. This has enabled Transformer to achieve remarkable results in tasks such as natural language processing (such as machine translation).

[0104] LSTM is a special recurrent neural network (RNN) that aims to solve the gradient vanishing and gradient exploding problems encountered by traditional RNNs when processing long sequences. LSTM controls the flow of information by introducing a gating mechanism (including input gate, forget gate, and output gate) to effectively capture long-term dependencies. This makes LSTM perform better when processing long sequence data.

[0105] In a specific embodiment, the deep learning prediction model of the present invention combines three architectures: BiGRU, Transformer, and LSTM to construct a hierarchical time series prediction model. First, the BiGRU layer extracts local information of the sequence through forward and reverse encoding to capture bidirectional dependencies; then, the Transformer layer uses a multi-head self-attention mechanism to generate a global feature representation, which can process all positions in the sequence in parallel and capture long-range dependencies; then, the two-layer LSTM layer further processes global features, captures long-term dependencies, and enhances the memory capacity and robustness of the model. Next, the fully connected layer maps the output of the LSTM layer to the target prediction space, while enhancing the nonlinear expression ability through the LeakyReLU activation function. Finally, the dynamic smoothing layer (using the moving average method) smoothes the output to eliminate short-term fluctuations in the prediction and provide more stable and accurate prediction results. Through this multi-level, comprehensive architecture, the model can better capture multidimensional information in time series data and improve prediction accuracy and robustness.

[0106] In one specific embodiment, the BiGRU-Transformer-LSTM model includes the following layers connected in sequence:

[0107] 1) BiGRU layer: used to perform forward and reverse encoding on the input time series data to capture bidirectional dependency information and output feature sequences. The output feature sequence also contains the features of the before and after relationships of the time series, improving the ability to understand the sequence pattern. The feature sequence output by the BiGRU layer encodes the local temporal dependencies of the time series data, providing contextual information of the time series data for Transformer and LSTM. As a bidirectional recurrent neural network (RNN), the BiGRU layer helps capture past and future dependencies through forward and reverse processing, while Transformer and LSTM are responsible for capturing global features and long-term dependencies, respectively. The entire process forms a hierarchical and progressive feature extraction and processing flow, thereby enhancing the model's ability to predict time series data.

[0108] 2) Transformer layer: used to receive the output of the BiGRU layer, use the multi-head self-attention mechanism to model the global features of all positions in the time series, and output the Transformer global features. The self-attention mechanism captures long-range dependencies and stabilizes feature expressions through residual connections and normalization. The powerful parallel computing capability significantly improves the training efficiency and enhances the feature expression capability of complex data. The output of the Transformer layer is the global feature representation after modeling the features of all positions in the time series through the multi-head self-attention mechanism. Finally, a tensor of (batch_size, sequence_length, hidden_size_gru * 2) dimensions is generated. This output incorporates global context information, allowing the model to capture the long-term dependencies and global relationships between different time steps in the time series in the features of each time step. This process enables the model to more fully understand the complex patterns in time series data and provide richer input features for subsequent LSTM layers.

[0109] 3) LSTM layer: The LSTM layer is a double-layer LSTM layer, including the first LSTM layer and the second LSTM layer; the first LSTM layer is used to further extract the temporal dependency of the Transformer global features. The second LSTM layer is used to combine long-term dependency features to generate smooth and memory-capable output features. Each LSTM layer contains a gating mechanism to enhance the robustness to noise and complex patterns. The role of the first LSTM layer is to further extract and model the local temporal dependencies of the time series from the global features output by the Transformer. It not only focuses on short-term changes and dynamics in the sequence, but also captures the gradually evolving trends, providing more detailed and orderly feature information for the subsequent second-layer LSTM, and ultimately helping to generate more accurate output features. Long-term dependency features refer to the need for the model to understand and remember patterns and trends that appear over a longer time span. Usually, this dependency goes beyond the scope of local information and requires the model to be able to recall and integrate information in long time series. In this model, the formation of long-term dependency features mainly depends on the following aspects:

[0110] Transformer's global features: Although Transformer focuses on global feature modeling, its multi-head self-attention mechanism can capture long-range dependencies between different time steps, that is, relationships over long time spans. These global features contain patterns and regularities in time series that may span a long period of time, such as seasonal changes, long-term trends, etc.

[0111] Local dependency processing of the first LSTM layer: The first LSTM layer captures the gradual evolution of information by processing the local dependencies of the time series (i.e., short-term and medium-term dynamic changes), but it does not overemphasize long-term trends. The feature sequence it outputs is a gradual transition from local dependencies to medium-term dependencies, and these features themselves may contain some short-term to medium-term intrinsic relationships.

[0112] Long-term memory capability of the second LSTM layer: The second LSTM layer receives the output features of the first LSTM layer and continues to process them, starting to focus on dependencies over a longer time span. LSTM has strong long-term memory capabilities, especially in its cell states and hidden states, and is able to retain and update information accumulated over time. This information can be long-term periodic trends in time series, long-term change patterns, slowly evolving fluctuations, etc.

[0113] 4) Fully connected layer and activation function: used to map the final output of the LSTM layer to the prediction result. The activation function uses the LeakyReLU activation function to handle nonlinear relationships and improve the model's adaptability to diverse inputs.

[0114] 5) Dynamic smoothing layer: It is used to add dynamic smoothing processing after the prediction output layer (fully connected layer and activation function) to further eliminate noise in the prediction results and optimize the stability of the prediction curve to obtain the final prediction results.

[0115] In a specific embodiment:

[0116] 1. Output of BiGRU layer: BiGRU layer receives input time series data and performs forward and reverse encoding to capture the bidirectional dependency information of the time series. Assume that the input data is a time series X={x1,x2,…,x T}, where T represents the length of the time step, x t is the input feature vector at the tth time step. The BiGRU layer will perform t Output a feature representation h t After forward and reverse processing, the feature sequence output by BiGRU is yes:

[0117]

[0118] Among them, h t It is a bidirectional feature representation generated at time step t, which may be a vector with dimension d BiGRU , indicating that the BiGRU layer captures the bidirectional dependency of this time step.

[0119] 2. Transformer layer output: The Transformer layer receives the feature sequence output by the BiGRU layer , using the multi-head self-attention mechanism to model global features at all positions in the time series and generate global feature representations. Through the self-attention mechanism, the Transformer layer can capture the relationship between each time step and other time steps and generate new feature representations. Assume that the output of the Transformer layer is , which will be the feature sequence after global modeling .

[0120]

[0121] in, Represents the global feature representation of time step t after processing by the Transformer layer, with dimension d Transformer , which represents the global context information of this time step.

[0122] 3. Output of the first LSTM layer: The first LSTM layer receives the output of the Transformer layer , and further extract the time dependency of the feature sequence after global feature modeling, and output the time step feature after the first layer LSTM processing. Assume that the output of the first layer LSTM is L1, where L1 is the time step feature representation after the first layer LSTM processing.

[0123]

[0124] Specifically, the first layer of LSTM uses the historical information of time steps to capture more complex temporal dependencies. Therefore, the output L1 of the first layer of LSTM is still a time series feature matrix with a shape of (T, d LSTM ), where d LSTM is the output dimension of the LSTM unit.

[0125] 4. Output of the second LSTM layer: The second LSTM layer receives the output L1 of the first LSTM layer and combines it with the long-term dependency features to generate the final output features. The output of the second LSTM layer is L2, which contains further processing of the first LSTM features and has stronger long-term dependency capabilities.

[0126]

[0127] Similarly, the output L2 of the second layer LSTM is also a time series feature matrix with a shape of (T, d LSTM ), which represents the feature representation after being processed by two layers of LSTM.

[0128] 5. Final output (after fully connected layer): After being processed by the LSTM layer, there is usually a fully connected layer to map to the final predicted value. The final output is generated by a fully connected layer and a LeakyReLU activation function, and is further optimized by a moving average smoothing layer.

[0129]

[0130] Among them, Y represents the final output feature, which is usually a scalar value and represents the prediction result of voltage extreme difference fluctuation.

[0131] Summarize:

[0132] 1. Output of BiGRU layer: , represents the bidirectional features at each time step.

[0133] 2. Output of Transformer layer: , which captures the global context information and enhances the modeling of long-term.

[0134] 3. Output of the first layer of LSTM: , which extracts the local time-dependent features of the time series.

[0135] 4. Output of the second layer LSTM: , which combines long-term information and further enhances feature expression.

[0136] 5. Final output: , is the final prediction result of the model.

[0137] The output of each layer is an enhancement of the processing results of the previous layer, helping the model better capture and model local and global features and long-term dependencies in time series.

[0138] See also Figure 3 As shown, an embodiment of the present invention provides a deep learning prediction method for battery system voltage extreme difference, including:

[0139] S1, collecting battery operation data of the battery cluster; the battery operation data includes single cell extreme difference data, battery cluster charge and discharge power and battery cluster total voltage data;

[0140] S2, inputting the battery operation data into a pre-trained deep learning prediction model for prediction, and obtaining a voltage range prediction result of the battery cluster within a preset time in the future;

[0141] Among them, the pre-trained deep learning prediction model is a BiGRU-Transformer-LSTM model.

[0142] In a specific embodiment, the pre-trained deep learning prediction model is trained by the following steps: obtaining a historical data set of a battery cluster of an energy storage power station, the historical data set including: single cell extreme difference data, battery cluster charge and discharge power, and battery cluster total voltage data; dividing the historical data set into N parts in chronological order to obtain N historical data subsets;

[0143] N-1 of the N historical data subsets are used as training sets, and the remaining 1 is used as validation and test sets; an identical initial deep learning prediction model is established for each of the N-1 historical data subsets in the training set;

[0144] Iteratively train a corresponding initial deep learning prediction model based on each historical data subset in the training set to obtain a trained initial deep learning prediction model; use N-1 historical data subsets to train a total of N-1 trained initial deep learning prediction models;

[0145] Each network parameter in the N-1 trained initial deep learning prediction models is weighted averaged according to the performance of the validation set, and the weight is determined inversely by the prediction error of each trained initial deep learning prediction model on the validation set; the parameter value after calculating the weighted average is assigned to the deep learning prediction model to generate a pre-trained deep learning prediction model.

[0146] Weighted average: weighted average is performed based on the performance of each model on the validation set (such as validation loss), rather than a simple arithmetic average:

[0147]

[0148]

[0149] in, is the parameter value after weighted average, is the network parameter weight of the i-th initial deep learning prediction model, is the network parameter of the i-th initial deep learning prediction model, is the prediction error of the i-th initial deep learning prediction model.

[0150] In a specific implementation, the BiGRU-Transformer-LSTM model includes sequentially connecting a BiGRU layer, a Transformer layer, an LSTM layer, a fully connected layer, and an activation function.

[0151] In a specific embodiment,

[0152] The BiGRU layer is used to perform forward and reverse encoding on the input time series data, capture bidirectional dependency information, and output a feature sequence;

[0153] The Transformer layer is used to receive the output of the BiGRU layer, use the multi-head self-attention mechanism to perform global feature modeling on all positions of the time series, and output the feature sequence after global feature modeling;

[0154] The LSTM layer is a double-layer LSTM layer, including the first LSTM layer and the second LSTM layer. The first LSTM layer is used to further extract the time dependency of the feature sequence after global feature modeling, and output the time step features after the first LSTM layer processing. The second LSTM layer is used to receive the output of the first LSTM layer and generate the final output features in combination with the long-term dependency features.

[0155] Fully connected layer and activation function: used to map the final output features of the LSTM layer into prediction results.

[0156] In a specific implementation, both the first LSTM layer and the second LSTM layer include a gating mechanism.

[0157] In a specific implementation, the BiGRU-Transformer-LSTM model further includes:

[0158] The dynamic smoothing layer is used to add dynamic smoothing processing after the fully connected layer and the activation function to further eliminate noise in the prediction results and optimize the stability of the prediction curve to obtain the final prediction results.

[0159] In a specific implementation, the activation function uses the LeakyReLU activation function.

[0160] The present invention proposes a BiGRU-Transformer-LSTM model. Although BiGRU, Transformer and LSTM each have their own unique advantages, using one of the models alone may not fully utilize all the information of the data in the face of complex tasks. Therefore, the present invention innovatively combines these three models to construct a hybrid model to fully utilize their respective advantages.

[0161] The present invention performs simulations and compares the prediction effects of the BiGRU-Transformer-LSTM model of the present invention and the existing LSTM, BiGRU-Transformer, CNN-LSTM-AM, and BiGRU-Transformer-LSTM models on predicting future voltage extreme differences. See Table 1 for details:

[0162] Table 1 Evaluation of model prediction effect

[0163]

[0164] It can be seen from Table 1 that the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) of the BiGRU-Transformer-LSTM model of the present invention are all the smallest. 2 maximum.

[0165] Please refer to Figure 4 As shown in FIG. 1 , a comparison chart of voltage extreme difference predictions under different methods is shown; a comprehensive comparison shows that the BiGRU-Transformer-LSTM model proposed in the present invention has a better prediction effect.

[0166] Step 5: Voltage consistency evaluation: Based on the final prediction results, evaluate the current battery voltage consistency status and optimize the charging and discharging strategies of each battery cluster.

[0167] In a specific implementation, the voltage consistency assessment specifically includes:

[0168] Voltage extreme value deviation rate β: The voltage extreme value of all single cells at each sampling point is divided by the average voltage, which can reflect problems such as battery aging and insufficient capacity.

[0169]

[0170] In the formula, k is the sampling time, u is the single cell voltage, is the average voltage. express k The difference between the maximum single cell voltage and the minimum single cell voltage in the battery cluster at a certain moment (the single cell here is not fixed, and is calculated based on the current maximum voltage). , Respectively k The maximum value and minimum value of the single cell voltage in the battery cluster at that moment.

[0171] Overall voltage deviation rate δ: At each sampling point, the standard deviation of the voltage of all single cells can reflect the overall performance of the battery cluster.

[0172]

[0173] In the formula, i is the battery number, n is the number of single cells in the battery cluster, Number i The single cell voltage, is the average voltage of the single cell.

[0174] Comprehensive scoring system: According to the scores of β and δ (refer to Table 2), a comprehensive score is given in a 1:1 ratio, corresponding to the four levels of "good (≥80), medium (≥70 and <80), passing (≥60 and <70), and unsatisfactory (<60)".

[0175] Table 2 Index scoring reference standards

[0176]

[0177] Note: The scoring standards in Table 2 are reference values ​​and can be adjusted according to the actual system. The weights of the two indicators can be adjusted appropriately. For example, the β weight can be increased to focus on the voltage limit problem. The present invention solves the voltage range problem. The voltage range refers to the difference between the maximum and minimum values ​​of the single cell voltage at a certain moment of each single cell. Therefore, the voltage range is a key indicator in voltage consistency analysis.

[0178] By analyzing the cumulative voltage extremes of battery cells within a certain period of time and combining a large amount of data for comparative analysis, the frequency of abnormal voltage states in battery cells can be obtained to determine whether the battery cells have entered abnormal working states due to long-term operation.

[0179] The voltage difference between battery cells , battery cell cumulative extreme difference voltage , voltage standard deviation Characterizes the voltage consistency index. The calculation method is as follows:

[0180]

[0181]

[0182]

[0183] Where: , They are t The highest and lowest voltage values ​​among all monomers in the sample data at the moment; , Numbered i d The battery cells in the statistical period T Maximum and minimum voltage values ​​within; Number i d The battery cells are t Voltage at the moment; Number i d The battery cells in the statistical period T The average voltage value within n for t The number of samples at a time.

[0184] The voltage range represents the maximum voltage difference between different battery cells at the same time, reflecting the voltage consistency between battery cells; the cumulative voltage range represents the voltage fluctuation of a single battery cell over a period of time, reflecting the voltage consistency of the battery cell after running for a period of time; the voltage standard deviation reflects the discreteness of the voltage between battery cells.

[0185] The embodiment of the present invention provides a deep learning prediction method for battery system voltage extreme difference, which can predict the battery voltage extreme difference fluctuation in multiple battery clusters in an energy storage power station and perform voltage consistency evaluation, thereby minimizing the adverse effects on the safety and life of the battery cluster caused by voltage inconsistency problems such as voltage over-limit, and improving the safety and reliability of the energy storage power station.

[0186] This method combines multiple groups of cross-validation and model transfer learning to improve the generalization ability and prediction accuracy of the model, so that the model has a fast prediction function. This method takes into account the fact that the voltage difference can be directly measured, and can adjust the model parameters and continuously optimize the model based on the feedback of the current voltage difference value and the predicted difference value.

[0187] See also Figure 5 As shown, an embodiment of the present invention provides a deep learning prediction device for battery system voltage extreme difference, comprising:

[0188] A collection module is used to collect battery operation data of the battery cluster; the battery operation data includes single cell extreme difference data, battery cluster charge and discharge power and battery cluster total voltage data;

[0189] A prediction module is used to input the battery operation data into a pre-trained deep learning prediction model for prediction, and obtain a voltage range prediction result of the battery cluster within a preset time in the future;

[0190] Among them, the pre-trained deep learning prediction model is a BiGRU-Transformer-LSTM model.

[0191] In a specific embodiment, the pre-trained deep learning prediction model is trained by the following steps: obtaining a historical data set of a battery cluster of an energy storage power station, the historical data set including: single cell extreme difference data, battery cluster charge and discharge power, and battery cluster total voltage data; dividing the historical data set into N parts in chronological order to obtain N historical data subsets;

[0192] N-1 of the N historical data subsets are used as training sets, and the remaining 1 is used as validation and test sets; an identical initial deep learning prediction model is established for each of the N-1 historical data subsets in the training set;

[0193] Iteratively train a corresponding initial deep learning prediction model based on each historical data subset in the training set to obtain a trained initial deep learning prediction model; use N-1 historical data subsets to train a total of N-1 trained initial deep learning prediction models;

[0194] Each network parameter in the N-1 trained initial deep learning prediction models is weighted averaged according to the performance of the validation set, and the weight is determined inversely by the prediction error of each trained initial deep learning prediction model on the validation set; the parameter value after calculating the weighted average is assigned to the deep learning prediction model to generate a pre-trained deep learning prediction model.

[0195] In a specific implementation, the BiGRU-Transformer-LSTM model includes sequentially connecting a BiGRU layer, a Transformer layer, an LSTM layer, a fully connected layer, and an activation function.

[0196] In a specific embodiment,

[0197] The BiGRU layer is used to perform forward and reverse encoding on the input time series data, capture bidirectional dependency information, and output a feature sequence;

[0198] The Transformer layer is used to receive the output of the BiGRU layer, use the multi-head self-attention mechanism to perform global feature modeling on all positions of the time series, and output the feature sequence after global feature modeling;

[0199] The LSTM layer is a double-layer LSTM layer, including the first LSTM layer and the second LSTM layer. The first LSTM layer is used to further extract the time dependency of the feature sequence after global feature modeling, and output the time step features after the first LSTM layer processing. The second LSTM layer is used to receive the output of the first LSTM layer and generate the final output features in combination with the long-term dependency features.

[0200] Fully connected layer and activation function: used to map the final output features of the LSTM layer into prediction results.

[0201] In a specific implementation, both the first LSTM layer and the second LSTM layer include a gating mechanism.

[0202] In a specific implementation, the BiGRU-Transformer-LSTM model further includes:

[0203] The dynamic smoothing layer is used to add dynamic smoothing processing after the fully connected layer and the activation function to further eliminate noise in the prediction results and optimize the stability of the prediction curve to obtain the final prediction results.

[0204] In a specific implementation, the activation function uses the LeakyReLU activation function.

[0205] See also Figure 6 As shown, an embodiment of the present invention provides an electronic device 100 for implementing a deep learning prediction method for battery system voltage extreme difference; the electronic device 100 includes a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on the at least one processor 102, and at least one communication bus 104.

[0206] The memory 101 can be used to store the computer program 103, and the processor 102 implements the steps of a deep learning prediction method for a battery system voltage extreme difference described in the embodiment by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101. The memory 101 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area may store data (such as audio data) created according to the use of the electronic device 100, etc. In addition, the memory 101 may include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other non-volatile solid-state storage devices.

[0207] The at least one processor 102 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor 102 may be a microprocessor or any conventional processor, etc. The processor 102 is the control center of the electronic device 100, and uses various interfaces and lines to connect various parts of the entire electronic device 100.

[0208] The memory 101 in the electronic device 100 stores a plurality of instructions to implement a deep learning prediction method for a battery system voltage extreme difference, and the processor 102 can execute the plurality of instructions to implement:

[0209] Collecting battery operation data of the battery cluster; the battery operation data includes single cell extreme difference data, battery cluster charge and discharge power and battery cluster total voltage data;

[0210] Input the battery operation data into a pre-trained deep learning prediction model for prediction, and obtain a voltage range prediction result of the battery cluster within a preset time in the future;

[0211] Among them, the pre-trained deep learning prediction model is a BiGRU-Transformer-LSTM model.

[0212] If the module / unit integrated in the electronic device 100 is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the process in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device that can carry the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory and read-only memory (ROM, Read-Only Memory).

[0213] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0214] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0215] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0216] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0217] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A deep learning prediction method for battery system voltage extreme difference, characterized in that: include: Collect battery operation data of battery cluster; The battery operation data includes single cell extreme difference data, battery cluster charge and discharge power and battery cluster total voltage data; Input the battery operation data into a pre-trained deep learning prediction model for prediction, and obtain a voltage range prediction result of the battery cluster within a preset time in the future; Wherein, the pre-trained deep learning prediction model is a BiGRU-Transformer-LSTM model; The BiGRU-Transformer-LSTM model includes a BiGRU layer, a Transformer layer, an LSTM layer, a fully connected layer and an activation function connected in sequence; The BiGRU layer is used to perform forward and reverse encoding on the input time series data, capture bidirectional dependency information, and output a feature sequence; The Transformer layer is used to receive the output of the BiGRU layer, use the multi-head self-attention mechanism to perform global feature modeling on all positions of the time series, and output the feature sequence after global feature modeling; The LSTM layer is a double-layer LSTM layer, including the first LSTM layer and the second LSTM layer. The first LSTM layer is used to further extract the time dependency of the feature sequence after global feature modeling, and output the time step features after the first LSTM layer processing. The second LSTM layer is used to receive the output of the first LSTM layer and generate the final output features in combination with the long-term dependency features. The fully connected layer and activation function are used to map the final output features of the LSTM layer into prediction results.

2. A deep learning prediction method for battery system voltage extreme difference according to claim 1, characterized in that: The pre-trained deep learning prediction model is trained by the following steps: Obtain a historical data set of a battery cluster of an energy storage power station, the historical data set including: single cell extreme difference data, battery cluster charge and discharge power, and battery cluster total voltage data; divide the historical data set into N parts in chronological order to obtain N historical data subsets; N is a positive integer, and N≥3; N-1 of the N historical data subsets are used as training sets, and the remaining 1 is used as validation and test sets; an identical initial deep learning prediction model is established for each of the N-1 historical data subsets in the training set; Based on each historical data subset in the training set, a corresponding initial deep learning prediction model is iteratively trained to obtain N-1 trained initial deep learning prediction models; Each network parameter in the N-1 trained initial deep learning prediction models is weighted averaged according to the performance of the validation set, and the weight is determined inversely by the prediction error of each trained initial deep learning prediction model on the validation set; the parameter value after calculating the weighted average is assigned to the deep learning prediction model to generate a pre-trained deep learning prediction model.

3. A deep learning prediction method for battery system voltage extreme difference according to claim 1, characterized in that: Both the first and second LSTM layers contain gating mechanisms.

4. A deep learning prediction method for battery system voltage extreme difference according to claim 1, characterized in that: The BiGRU-Transformer-LSTM model also includes: The dynamic smoothing layer is used to add dynamic smoothing processing after the fully connected layer and the activation function to further eliminate noise in the prediction results and optimize the stability of the prediction curve to obtain the final prediction results.

5. The deep learning prediction method for battery system voltage extreme difference according to claim 1 is characterized in that: The activation function uses the LeakyReLU activation function.

6. A deep learning prediction device for battery system voltage extreme difference, characterized in that: include: A collection module, used to collect battery operation data of the battery cluster; The battery operation data includes single cell extreme difference data, battery cluster charge and discharge power and battery cluster total voltage data; A prediction module is used to input the battery operation data into a pre-trained deep learning prediction model for prediction, and obtain a voltage range prediction result of the battery cluster within a preset time in the future; Wherein, the pre-trained deep learning prediction model is a BiGRU-Transformer-LSTM model; The BiGRU-Transformer-LSTM model includes a BiGRU layer, a Transformer layer, an LSTM layer, a fully connected layer and an activation function connected in sequence; The BiGRU layer is used to perform forward and reverse encoding on the input time series data, capture bidirectional dependency information, and output a feature sequence; The Transformer layer is used to receive the output of the BiGRU layer, use the multi-head self-attention mechanism to perform global feature modeling on all positions of the time series, and output the feature sequence after global feature modeling; The LSTM layer is a double-layer LSTM layer, including the first LSTM layer and the second LSTM layer. The first LSTM layer is used to further extract the time dependency of the feature sequence after global feature modeling, and output the time step features after the first LSTM layer processing. The second LSTM layer is used to receive the output of the first LSTM layer and generate the final output features in combination with the long-term dependency features. The fully connected layer and activation function are used to map the final output features of the LSTM layer into prediction results.

7. A deep learning prediction device for battery system voltage extreme difference according to claim 6, characterized in that: The pre-trained deep learning prediction model is trained by the following steps: Acquire a historical data set of a battery cluster of an energy storage power station, the historical data set including: single cell extreme difference data, battery cluster charge and discharge power, and battery cluster total voltage data; divide the historical data set into N parts in chronological order to obtain N historical data subsets; N-1 of the N historical data subsets are used as training sets, and the remaining 1 is used as validation and test sets; an identical initial deep learning prediction model is established for each of the N-1 historical data subsets in the training set; Based on each historical data subset in the training set, a corresponding initial deep learning prediction model is iteratively trained to obtain N-1 trained initial deep learning prediction models; Each network parameter in the N-1 trained initial deep learning prediction models is weighted averaged according to the performance of the validation set, and the weight is determined inversely by the prediction error of each trained initial deep learning prediction model on the validation set; the parameter value after calculating the weighted average is assigned to the deep learning prediction model to generate a pre-trained deep learning prediction model.

8. A deep learning prediction device for battery system voltage extreme difference according to claim 6, characterized in that: Both the first and second LSTM layers contain gating mechanisms.

9. A deep learning prediction device for battery system voltage extreme difference according to claim 6, characterized in that: The BiGRU-Transformer-LSTM model also includes: The dynamic smoothing layer is used to add dynamic smoothing processing after the fully connected layer and the activation function to further eliminate noise in the prediction results and optimize the stability of the prediction curve to obtain the final prediction results.

10. A deep learning prediction device for battery system voltage extreme difference according to claim 6, characterized in that: The activation function uses the LeakyReLU activation function.

11. An electronic device, characterized in that: It comprises a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement a deep learning prediction method for battery system voltage extreme difference as described in any one of claims 1 to 5.

12. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one instruction, and when the at least one instruction is executed by the processor, it implements the deep learning prediction method for battery system voltage extreme difference as described in any one of claims 1 to 5.

Images