A Battery Pack SOH Prediction System and Method Based on Blockchain and Informer Neural Network

Abstract

The present invention provides a battery pack SOH prediction system and method based on blockchain and Informer neural network. The prediction system includes an in-vehicle blockchain platform, a regional private chain, and a consortium blockchain. The consortium blockchain extracts features and labels from the public index data of electric vehicles and their battery packs, preprocesses the dataset of the public index data features, and together with the labels, constitutes a dataset X n , and discards weakly correlated features to obtain a dataset X' n , and uses the dataset X' of other electric vehicles n to train an Informer-based neural network model, recalibrate the trained Informer-based neural network model according to the historical data of the electric vehicle itself, set the prediction sequence length for the recalibrated Informer-based neural network model, and predict the SOH of the battery pack. The present invention can accurately monitor the future health status of the battery pack, which helps to manage the entire life cycle of the battery well.

Description

Technical Field

[0001] The present invention belongs to the field of blockchain technology and health status prediction of battery packs of new energy electric vehicles, and specifically relates to a battery pack SOH prediction system and method based on blockchain and Informer neural network. Background Art

[0002] Lithium-ion battery packs are usually composed of thousands of cells connected in series and parallel, all of which are monitored and managed in real time by the battery management system BMS, which measures the external characteristic data of the battery pack, such as current, voltage, temperature, etc., and completes functions such as state estimation, thermal management, automatic balancing, and fault diagnosis to ensure the safe operation of the battery. The charging and discharging of lithium-ion battery packs is a complex process of converting electrical energy, chemical energy, and thermal energy. It is highly nonlinear and uncertain, and its performance parameters are easily affected by many factors such as environmental conditions, battery aging, and user behavior. This makes it difficult to control the safe and efficient operation of the battery pack. At present, a complete battery management theory and technology system has not yet been formed. Current research on battery SOH prediction is mainly focused on the cell level, and the problem of battery pack SOH prediction in practical applications needs to be solved urgently. Accurate long-sequence prediction of battery SOH is of great significance to ensuring the safety of electric vehicles, and can also provide guarantees for the maintenance and replacement of battery packs. Summary of the invention

[0003] In view of this, the present invention provides a battery pack SOH prediction system and method based on blockchain and Informer neural network.

[0004] The present invention achieves the above technical objectives through the following technical means.

[0005] A battery pack SOH prediction method based on blockchain and Informer neural network, specifically:

[0006] The alliance blockchain extracts features and labels from the public index data of electric vehicles and their battery packs;

[0007] Preprocess the dataset of public index data features and form dataset X together with the labels n , and discard weakly correlated features to obtain the data set X′ n ;

[0008] Using other electric vehicle datasets X′ n Train the Informer-based neural network model;

[0009] The trained Informer-based neural network model is recalibrated according to the historical data of the electric vehicle itself, and the prediction sequence length is set for the recalibrated Informer-based neural network model to predict the SOH of the battery pack.

[0010] Further technical solution, the features include vehicle driving behavior features, battery pack state features, and environmental features of the area where the electric vehicle is located.

[0011] Further technical solution, the label is:

[0012]

[0013] where: y SOH represents the label of the public index data, C′ T represents the result of autoregressive processing on the corrected capacity C T and C T = K T C max K T is the temperature influence factor, C max is the current maximum available capacity of the battery pack, C R is the rated capacity of the battery pack, y N is the SOH value of the Nth charge / discharge segment.

[0014] Even further technical solution, preprocess the dataset of public index data features, including:

[0015] For the SOC in the battery pack state features, perform anomaly detection. When ΔSOC k+1 -ΔSOC k < μ, retain the current SOC value, otherwise discard it, where ΔSOC k is the battery SOC corresponding to the sampling time point k, and μ is the SOC change rate threshold;

[0016] For features other than SOC in the dataset: Discard the columns where the missing values reach more than 80%, and fill those with missing values below 80%; then perform outlier detection on each column of features and delete the rows where the outliers are located; then divide the remaining features into continuous features X c and discrete features X d , and perform normalization to X′ c , one-hot encoding to X′ d ;

[0017] The dataset X n = concat(X′ c , X′ d , y SOH ), concat is the concatenation function, and y SOH represents the label of the public index data.

[0018] Even further technical solution, the weakly correlated features refer to the dataset X nFeatures with a correlation coefficient less than 0.6 in

[0019] In a further technical solution, the neural network model based on Informer includes an embedding layer, a multi-head probabilistic sparse self-attention layer, a self-attention distillation layer, a generative decoder network, and a fully connected layer that communicate in sequence.

[0020] In a further technical solution, the process of recalibrating the trained neural network model based on Informer is as follows: sending token T to the consortium blockchain ID , if token T ID = True, all public indexes belonging to the vehicle itself can be found, and the calibration is successful.

[0021] A battery pack SOH prediction system based on blockchain and Informer neural network, comprising:

[0022] An in-vehicle blockchain platform that stores the private information and public information of each registered electric vehicle. The private information is encrypted and a security index is generated, and the public information directly generates a public index;

[0023] A regional private chain that receives the security index and public index sent by the in-vehicle blockchain platform and stores the security index;

[0024] A consortium blockchain that stores the public indexes sent by the regional private chain, extracts features and labels from the public indexes, trains a neural network model based on Informer using the public indexes of other electric vehicles, and recalibrates the trained neural network model based on Informer using the data of the electric vehicle itself.

[0025] In the above technical solution, the public information includes the vehicle's location, driving behavior, and battery pack data.

[0026] The beneficial effects of the present invention are as follows:

[0027] (1) The blockchain technology adopted by the prediction system of the present invention has high data integrity and security. While paying attention to privacy, it enables the battery data of electric vehicles from different driving conditions and regions to be shared between vehicles; trains a neural network model based on Informer using the public indexes of other electric vehicles, and recalibrates the trained neural network model based on Informer using the data of the electric vehicle itself to obtain a more accurate SOH prediction model;

[0028] (2) The prediction method of the present invention uses an Informer-based neural network model, which has higher accuracy, efficiency, and longer sequence length compared to CNN, LSTM, and Transformer neural networks. It can accurately monitor the health status of the power battery pack in the next week or even longer, which will help manage the entire life cycle of the battery and provide a reference for battery maintenance and replacement. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a flowchart of the SOH prediction of the battery pack based on blockchain and Informer neural network according to the present invention;

[0030] Figure 2 It is a schematic diagram of the neural network structure based on Informer according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] The present invention will be further described below with reference to the drawings and specific embodiments, but the protection scope of the present invention is not limited thereto.

[0032] As Figure 1 shown, the SOH prediction system of the battery pack based on blockchain and Informer neural network according to the present invention includes:

[0033] An on-vehicle blockchain platform. After each electric vehicle is registered on the on-vehicle blockchain platform, it obtains a token T ID . When the electric vehicle travels to area n, a corresponding vehicle record is generated. The on-vehicle blockchain platform stores the original complete information of each registered electric vehicle, including private information (electric vehicle ID, driving route, and smart device login information) and public information (including the area where the vehicle is located, the driving behavior of the vehicle, and battery pack data). The private information is encrypted and a security index is generated, while the public information directly generates a public index; the on-vehicle blockchain platform exchanges information with the area private chain and the consortium blockchain.

[0034] The area private chain receives the security index and public index sent by the on-vehicle blockchain platform, stores the security index, and sends the public index of the electric vehicle to the consortium blockchain.

[0035] The consortium blockchain stores the public index of electric vehicles, extracts relevant features and labels from the public index data of electric vehicles and their battery packs, preprocesses the feature dataset, builds a neural network model based on Informer, trains the neural network model based on Informer using the public indexes of other electric vehicles, recalibrates the trained neural network model based on Informer using the data of the electric vehicle itself, predicts the SOH of the electric vehicle through the recalibrated neural network model based on Informer, and outputs it to the in-vehicle blockchain platform for display.

[0036] As Figure 1 shown, a method for predicting the SOH of a battery pack based on blockchain and Informer neural network specifically includes the following steps:

[0037] S1. The consortium blockchain extracts relevant features and labels from the public index data of electric vehicles and their battery packs

[0038] S1.1. Use Python-related libraries to extract vehicle driving behavior features X D (including vehicle speed, acceleration, depth of discharge, driving mileage) and battery pack status features X B (including voltage, current, discharge rate, temperature, SOC) and environmental features X of the region n where the electric vehicle is located W (including temperature, humidity, air pressure, precipitation, visibility, wind speed) to obtain a dataset D with initially extracted features,

[0039]

[0040] where: i is the i-th feature in the vehicle driving behavior feature X D j is the j-th feature in the battery pack status feature X B k is the k-th feature in the local environmental feature X W N is the N-th charge / discharge segment of the battery pack, i.e., the sequence length, d model is the feature dimension, is the set of real numbers;

[0041] S1.2. Under laboratory conditions, the rated capacity of the current battery pack is measured through a capacity calibration experiment. However, in the actual operation of the vehicle, it is difficult to meet such full charge and discharge calibration conditions in practice, and it is necessary to extract the current maximum available capacity C of the battery pack max :

[0042]

[0043] where SOC(t0) is the state of charge of the battery at the start of discharge, SOC(tk ) is the state of charge of the battery at the end of discharge, I(t) is the current of the battery pack at time t, η is the Coulomb efficiency, and Δt is the sampling interval;

[0044] S1.3, the temperature influence factor corrects the current maximum available capacity C of the battery pack max , the ambient temperature will affect the maximum available capacity of the battery pack. The discharge capacity of the battery pack will decrease significantly in a low-temperature environment and increase in a high-temperature environment. Therefore, the corrected capacity C T = K T C max , where K T is the temperature influence factor. The battery pack corresponding to a certain vehicle model can be placed in experiments at -40°C, -25°C, -15°C, 0°C, 15°C, 25°C, and 40°C to obtain the temperature influence factor K T , and the temperature influence factor K T in the remaining temperature ranges can be obtained by interpolation;

[0045] S1.4, use the regression algorithm to perform autoregressive processing on C T to obtain C′ T , divide it by the rated capacity C of the battery pack R and use it as the label y SOH :

[0046]

[0047] where, y N is the SOH value of the Nth charge / discharge segment;

[0048] The regression algorithm includes logistic regression, decision tree regression, support vector machine regression, and locally weighted linear regression.

[0049] S2, the data collected from real vehicles is difficult to meet the full charge and full discharge test conditions. And due to factors such as sensor failures and BMS sampling accuracy, the federated blockchain is required to preprocess the feature dataset D;

[0050] S2.1, for the SOC in the battery pack state feature X B perform anomaly detection. When ΔSOC k+1 - ΔSOC k < μ, retain the current SOC value, otherwise discard it. Where, ΔSOC k is the battery SOC corresponding to the sampling time point k, and μ is the SOC change rate threshold;

[0051] S2.2. For the features in the feature dataset D other than SOC, the columns where the missing values of the features reach more than 80% are discarded. For the features with missing values below 80%, the missing values are filled using the previous value, the next value, the mean value, the hot deck imputation, the fitting imputation, or the multiple imputation. Then, outlier detection is performed on each column of features, and the rows where the outliers are located are deleted. Further, it is divided into continuous features X c and discrete features X d , and they are respectively normalized to X′ c and one-hot encoded to X′ d .

[0052] S3. Concatenate the dataset X n = concat(X′ c , X′ d , y SOH ), and use correlation analysis to discard the weakly correlated features with a correlation coefficient less than 0.6 to obtain a new dataset X′ n . Take the features of the new dataset X′ n as the input of the neural network model based on Informer.

[0053] S4. Build a neural network model based on Informer, as Figure 2 shown. The Informer neural network includes an embedding layer, a multi-head probabilistic sparse self-attention layer, a self-attention distillation layer, a generative decoder network, and a fully connected layer;

[0054] S4.1. The dataset X′ n is input into the embedding layer, and the embedding layer converts it into an input vector The input vector consists of three parts: the feature scalar the local timestamp PE, and the global timestamp SE;

[0055]

[0056] Among them, α is a factor that balances the size between the scalar mapping and the local / global embedding. If the sequence input has been normalized, then α = 1; m ∈ [1, N], and p is the number of global timestamp types;

[0057] The calculation formula for the local timestamp PE is as follows:

[0058]

[0059]

[0060] Among them, n ∈ [1,..., d model / 2], and pos represents the position information;

[0061] The global timestamp SE uses a fully connected layer to map the input timestamp to 512 dimensions.

[0062] S4.2, The multi-head probabilistic sparse self-attention layer linearly transforms the input vector into three vectors Q, K, and V. The formula is as follows:

[0063]

[0064] where, are the query vector, key vector, and value vector respectively, is the training parameter matrix corresponding to Q, K, and V;

[0065] Furthermore, obtain the output of each multi-head probabilistic sparse self-attention layer:

[0066]

[0067] where, is the sparse matrix of Q, and Softmax is the activation function.

[0068] S4.3, The self-attention distillation layer uses distillation operations to retain dominant features. The process from the q-th layer to the q+1-th layer is as follows:

[0069]

[0070] where, is the output of the multi-head probabilistic sparse self-attention layer of the q+1-th layer, MaxPool is the max pooling function, ELU is the activation function, Conv1d is the one-dimensional convolution function, is the calculation result of the multi-head probabilistic sparse self-attention layer of the q-th layer, including the multi-head attention mechanism.

[0071] S4.4, The generative decoder network generates all prediction outputs at once. The process is as follows:

[0072]

[0073] where, is the input of the generative decoder network, is the start token, is the placeholder for the target sequence, filled with 0, concat is the concatenation function, N y is the length of the prediction sequence, N token is the length of the input sequence.

[0074] S4.5, After passing through the generative decoder network, each position to be predicted in the sequence length has a vector, which is then input into a fully connected layer to obtain the prediction result.

[0075] S5. For the built Informer-based neural network model, it needs to be trained to learn the internal parameters of the model;

[0076] S5.1. Divide the dataset X′ of other electric vehicles into a training set X n and a test set X train ; test

[0077] S5.2. Set the hyperparameters of the Informer-based neural network model. The hyperparameters include the frequency freq of time feature encoding, the prediction sequence length pred_len, Dropout, the learning rate lr, the number of encoder layers e_layers, the number of decoder layers d_layers, the number of attention heads n_heads, the number of stacked encoder layers s_layers, the loss function loss, and the number of Epochs. At the same time, use the backpropagation algorithm to update the internal weights of the Informer-based neural network model, and use the mean square error MSE and the mean relative error MRE as the loss function and evaluation index of the Informer-based neural network model. The formulas are as follows:

[0078]

[0079]

[0080] Among them, is the actual SOH value of the current test sample, is the estimated value of, a ∈ [1, N];

[0081] S5.3. Use the training set X train to train the built Informer-based neural network model. After training, use the test set X test to test the trained model. If the mean relative error MRE < ε, output the trained Informer-based neural network model; otherwise, set the hyperparameters and retrain.

[0082] S6. Re-calibrate the trained Informer-based neural network model according to the historical data of the electric vehicle itself: Send the token T ID to the consortium blockchain. If the token T ID = True, all public indexes belonging to the electric vehicle can be found. If the token T ID = false, it means that the electric vehicle is not registered on the in-vehicle blockchain platform.

[0083] ​S7. Based on the recalibrated Informer-based neural network model, by setting the prediction sequence length pred_len, the prediction length of the model can be adjusted. The recalibrated Informer-based neural network model can be used to estimate the SOH of the battery pack for a relatively long period in the future. On the basis of a certain prediction sequence length, its accuracy and efficiency are much higher than those of neural networks such as CNN and LSTM.

[0084] The described embodiments are the preferred embodiments of the present invention, but the present invention is not limited to the above embodiments. Without departing from the essence of the present invention, any obvious improvements, substitutions or variations that those skilled in the art can make all fall within the protection scope of the present invention.

Claims

1. A method for predicting the state of health (SOH) of a battery pack based on blockchain and Informer neural network, characterized in that: The consortium blockchain extracts features and labels from the public index data of electric vehicles and their battery packs; Preprocess the dataset of common index data features, which together with the labels form the dataset X n , and discard weakly related features to obtain the dataset X' n ; Utilize the dataset X′ of other electric vehicles n Train an Informer-based neural network model; Based on the historical data of the electric vehicle itself, the trained neural network model based on Informer is recalibrated. For the recalibrated neural network model based on Informer, the prediction sequence length is set to predict the SOH of the battery pack; Preprocess the dataset of the public index data features, including: Perform anomaly detection on the SOC in the battery pack state characteristics. When ΔSOC k+1 -ΔSOC k < μ, retain the current SOC value; otherwise, discard it. Here, ΔSOC k is the battery SOC corresponding to the sampling time point k, and μ is the SOC change rate threshold; For features in the dataset other than SOC: Discard the columns where the missing values reach more than 80%, and fill the features with missing values below 80%; then perform outlier detection on each column of features and delete the rows where outliers are located; then divide the remaining features into continuous features X c and discrete features X d , and perform normalization to X' c , one-hot encoding to X' d ; The dataset X n = concat(X′ c , X′ d , y SOH ), where concat is the concatenation function, and y SOH represents the label of the common index data.

2. The method for predicting the SOH of a battery pack according to claim 1, wherein The features include vehicle driving behavior features, battery pack state features, and environmental features of the area where the electric vehicle is located.

3. The method for predicting the SOH of a battery pack according to claim 1, wherein The labels are: Where: y SOH represents the label of the common index data, C′ T represents the result of autoregressive processing on the corrected capacity C T and C T = K T C max , K T is the temperature influence factor, C max is the current maximum available capacity of the battery pack, C R is the rated capacity of the battery pack, y N is the SOH value of the Nth charge / discharge segment.

4. The method for predicting the SOH of a battery pack according to claim 1, characterized in that, The weakly correlated features refer to the dataset X n with a correlation coefficient less than 0.6 for the features.

5. The method for predicting the SOH of a battery pack according to claim 1, wherein, The neural network model based on Informer includes an embedding layer, a multi-head probabilistic sparse self-attention layer, a self-attention distillation layer, a generative decoder network, and a fully connected layer that communicate in sequence.

6. A system for implementing the battery pack SOH prediction method according to any one of claims 1-5, characterized in that, Including: An in-vehicle blockchain platform that stores the private information and public information of each registered electric vehicle. The private information is encrypted and a security index is generated, and the public information directly generates a public index; A regional private chain that receives the security index and public index sent by the in-vehicle blockchain platform and stores the security index; A consortium blockchain that stores the public index sent by the regional private chain, extracts features and labels from the public index, trains a neural network model based on Informer using the public indexes of other electric vehicles, and recalibrates the trained neural network model based on Informer using the data of the electric vehicle itself; The process of recalibrating the trained Informer-based neural network model is as follows: sending token T to the consortium blockchain ID , if token T ID = True, all public indexes belonging to the vehicle can be found, and the calibration is successful.

7. The system according to claim 6, wherein The public information includes the area where the vehicle is located, the vehicle driving behavior, and the battery pack data.

Images