A method and device for predicting the life of energy storage batteries based on improved TFT

Abstract

The present invention discloses a method and device for predicting the life of an energy storage battery based on an improved TFT. The method extracts set time series data from a time series database of an energy storage power station of a set scale; performs missing value supplementation processing on the set time series data by a spline interpolation method, performs standardization processing by a z-score strategy, performs abnormal point detection by a class of support vector machines, and obtains abnormal point data; performs replacement processing by a weighted moving average strategy, and obtains target time series data; performs feature capture on periodic data by a time convolution network and a long short-term memory network, and obtains periodic time series data features; performs feature capture on non-periodic data by a multi-layer perceptron, and obtains non-periodic time series data features; inputs data features into an improved TFT prediction model for calculation processing, and outputs prediction results. The present invention can effectively process the time dependence of long time series data and capture complex nonlinear relationships in battery data time series.

Description

Technical Field

[0001] The present invention relates to the technical field of energy storage battery analysis, and in particular to a method and device for predicting the life of an energy storage battery based on an improved TFT. Background Art

[0002] The main problems in current data-driven battery life prediction methods are the difficulty in accurately capturing and modeling the dynamic characteristics of complex changes in battery performance over time, as well as the limitations in processing multi-source heterogeneous data. These problems include:

[0003] First, there is insufficient handling of dynamic complexity: many traditional prediction models fail to effectively capture long-term and short-term time dependencies, resulting in lower accuracy in predicting battery life.

[0004] The second is the difficulty in integrating heterogeneous data: In the prediction of battery capacity attenuation, it is necessary to consider a variety of influencing factors (such as temperature, charging and discharging frequency, etc.). These factors often come from different data streams, and traditional models are difficult to effectively integrate these heterogeneous data sources.

[0005] Third, the periodic and non-periodic factors are not effectively distinguished: Battery life decay is a complex process, which is affected by a variety of periodic (such as seasonal changes) and non-periodic (such as long-term use loss) factors. Existing models often fail to fully distinguish between these two types of factors, resulting in prediction results that cannot accurately reflect the actual performance of batteries under different conditions.

[0006] Fourth, data noise and outliers are not handled well: Battery performance data often contains noise and outliers, such as equipment failures, measurement errors, etc. The existence of these data greatly reduces the accuracy and reliability of model predictions. Existing prediction methods often do not work well when processing these non-standard data and cannot effectively identify and eliminate noise and anomalies.

[0007] Therefore, how to invent a method for predicting the life of energy storage batteries and improve the accuracy of battery life prediction is an urgent problem to be solved. Summary of the invention

[0008] To this end, the present invention provides a method and device for predicting the life of an energy storage battery based on an improved TFT. By adopting the improved TFT model as the main framework of the battery life prediction algorithm, the time dependence of long time series data can be effectively processed and the complex nonlinear relationship in the battery data time series can be captured. By separating and processing periodic and non-periodic attenuation factors, the life changes of batteries under different environments and usage conditions can be more accurately identified and simulated. Through the enhanced outlier and noise data processing mechanism, while improving the prediction accuracy of the model, the robustness of the model in practical applications is also enhanced.

[0009] In order to achieve the above object, the present invention provides the following technical solution: a method for predicting the life of an energy storage battery based on an improved TFT, comprising:

[0010] Extracting set time series data from a time series database of a storage power station of a set scale, wherein the set time series data is used to improve the TFT prediction model for training;

[0011] The missing values ​​of the set time series data are supplemented by the spline interpolation method to obtain the supplemented set time series data; the supplemented set time series data are standardized by the z-score strategy to obtain standardized time series data; the standardized time series data are detected for abnormal points by a class of support vector machines to obtain abnormal point data; the abnormal point data are replaced by the weighted moving average strategy to obtain target time series data;

[0012] Capturing the features of the periodic data in the target time series data through a time convolution network and a long short-term memory network to obtain the features of the periodic time series data;

[0013] Capturing features of the non-periodic data in the target time series data by using a multi-layer perceptron to obtain features of the non-periodic time series data;

[0014] The periodic time series data features and the non-periodic time series data features are input into the improved TFT prediction model; calculation processing is performed through the improved TFT prediction model to output a prediction result.

[0015] As a preferred solution of the energy storage battery life prediction method based on improved TFT, in the process of extracting the set time series data from the time series database of the energy storage power station of a set scale, the set time series data includes: battery temperature, voltage, current, number of charge and discharge times, external ambient temperature, charge and discharge cycle, battery initial capacity, battery capacity attenuation data and actual battery service life.

[0016] As a preferred solution of the energy storage battery life prediction method based on improved TFT, in the process of supplementing the missing values ​​of the set time series data by the spline interpolation method to obtain the supplemented set time series data, the mathematical expression of the spline interpolation method is:

[0017] ;

[0018] In the formula, is the interpolation function, which means The function value in the interval, and are the horizontal coordinates of adjacent known data points respectively; , , , is the interpolation coefficient and is the model hyperparameter.

[0019] As a preferred solution of the energy storage battery life prediction method based on improved TFT, in the process of obtaining standardized time series data by performing standardization processing through the z-score strategy, the mathematical expression of the z-score strategy is:

[0020] ;

[0021] In the formula, is the original data matrix; and are the mean and standard deviation of the data respectively.

[0022] As a preferred solution of the energy storage battery life prediction method based on improved TFT, in the process of replacing the abnormal point data by the weighted moving average strategy to obtain the target time series data, the mathematical expression of the weighted moving average strategy is:

[0023] ;

[0024] In the formula, For time point The data value of is the weight; is the number of neighboring points considered.

[0025] As a preferred solution of the energy storage battery life prediction method based on improved TFT, in the process of capturing the features of the periodic data in the target time series data through the time convolution network and the long short-term memory network, the calculation formula of the time convolution network is:

[0026] ;

[0027] In the formula, is the periodic data input; is the number of filters; is the size of the convolution kernel; It is an activation function used to enhance the nonlinear ability of the network;

[0028] The calculation formula of the long short-term memory network is:

[0029] ;

[0030] In the formula, The time steps The hidden state and cell state of .

[0031] As a preferred solution of the energy storage battery life prediction method based on improved TFT, in the process of capturing the features of the non-periodic data in the target time series data through the multi-layer perceptron, the calculation formula of the multi-layer perceptron is:

[0032] ;

[0033] In the formula, X is the input non-periodic data matrix; and , are the weights and biases of each layer respectively.

[0034] As a preferred solution of a method for predicting the life of an energy storage battery based on an improved TFT, the steps of calculating and processing the improved TFT prediction model are as follows:

[0035] Through the feature selection layer, the input periodic time series data features and the non-periodic time series data features are screened to obtain the features that contribute most to the battery life prediction;

[0036] Through the multi-head self-attention mechanism and gated recursive unit in the core computing layer, the screened features are calculated and processed to output the prediction results.

[0037] As a preferred solution of the energy storage battery life prediction method based on improved TFT, the prediction result output by the improved TFT prediction model is an interval prediction, which indicates that the battery will have a certain life span in the future. The predicted life expectancy value for each day within ; the interval representation of the prediction result is:

[0038] ;

[0039] In the formula, For in time The minimum predicted life span value; For in time The highest predicted lifespan value.

[0040] The present invention also provides a device for predicting the life of an energy storage battery based on an improved TFT, based on the above method for predicting the life of an energy storage battery based on an improved TFT, comprising:

[0041] A time series data extraction module is used to extract set time series data from a time series database of a set scale of energy storage power station, and the set time series data is used to improve the TFT prediction model for training;

[0042] A time series data preprocessing module is used to supplement the missing values ​​of the set time series data by spline interpolation method to obtain the supplemented set time series data; standardize the supplemented set time series data by z-score strategy to obtain standardized time series data; perform outlier detection on the standardized time series data by a class of support vector machines to obtain outlier data; replace the outlier data by weighted moving average strategy to obtain target time series data;

[0043] A periodic time series data feature acquisition module is used to capture the features of the periodic data in the target time series data through a time convolution network and a long short-term memory network to obtain the periodic time series data features;

[0044] A non-periodic time series data feature acquisition module, used to capture the features of the non-periodic data in the target time series data through a multi-layer perceptron to obtain the features of the non-periodic time series data;

[0045] The improved TFT prediction model prediction module is used to input the periodic time series data characteristics and the non-periodic time series data characteristics into the improved TFT prediction model; perform calculation processing through the improved TFT prediction model and output a prediction result.

[0046] As a preferred solution of the energy storage battery life prediction device based on improved TFT, in the timing data extraction module, in the process of extracting the set timing data from the timing database of the energy storage power station of a set scale, the set timing data includes: battery temperature, voltage, current, number of charge and discharge times, external ambient temperature, charge and discharge cycle, initial capacity of the battery, battery capacity attenuation data and actual service life of the battery.

[0047] As a preferred solution of the energy storage battery life prediction device based on improved TFT, in the time series data preprocessing module, in the process of supplementing the missing values ​​of the set time series data by the spline interpolation method to obtain the supplemented set time series data, the mathematical expression of the spline interpolation method is:

[0048] ;

[0049] In the formula, is the interpolation function, which means The function value in the interval, and are the horizontal coordinates of adjacent known data points respectively; , , , is the interpolation coefficient and is the model hyperparameter.

[0050] As a preferred solution of the energy storage battery life prediction device based on improved TFT, in the time series data preprocessing module, in the process of obtaining standardized time series data by performing standardization processing through the z-score strategy, the mathematical expression of the z-score strategy is:

[0051] ;

[0052] In the formula, is the original data matrix; and are the mean and standard deviation of the data respectively.

[0053] As a preferred solution of the energy storage battery life prediction device based on improved TFT, in the time series data preprocessing module, in the process of replacing the abnormal point data by the weighted moving average strategy to obtain the target time series data, the mathematical expression of the weighted moving average strategy is:

[0054] ;

[0055] In the formula, For time point The data value of is the weight; is the number of neighboring points considered.

[0056] As a preferred solution of the energy storage battery life prediction device based on improved TFT, in the periodic time series data feature acquisition module, in the process of capturing the features of the periodic data in the target time series data through the time convolution network and the long short-term memory network, the calculation formula of the time convolution network is:

[0057] ;

[0058] In the formula, is the periodic data input; is the number of filters; is the size of the convolution kernel; It is an activation function used to enhance the nonlinear ability of the network;

[0059] The calculation formula of the long short-term memory network is:

[0060] ;

[0061] In the formula, The time steps The hidden state and cell state of .

[0062] As a preferred solution of the energy storage battery life prediction device based on improved TFT, in the non-periodic time series data feature acquisition module, in the process of capturing the features of the non-periodic data in the target time series data through the multi-layer perceptron, the calculation formula of the multi-layer perceptron is:

[0063] ;

[0064] In the formula, X is the input non-periodic data matrix; and , are the weights and biases of each layer respectively.

[0065] As a preferred solution of the energy storage battery life prediction device based on improved TFT, in the improved TFT prediction model prediction module, the prediction processing submodule includes:

[0066] A feature selection submodule, used to screen the input periodic time series data features and the non-periodic time series data features through a feature selection layer to obtain the features that contribute most to battery life prediction;

[0067] The core computing submodule is used to calculate and process the selected features through the multi-head self-attention mechanism and gated recursive unit in the core computing layer and output the prediction results.

[0068] As a preferred solution of the energy storage battery life prediction device based on improved TFT, in the prediction module of the improved TFT prediction model, the prediction result output by the improved TFT prediction model is an interval prediction, and the interval prediction indicates that the battery will have a certain life span in the future. The predicted life expectancy value for each day within ; the interval representation of the prediction result is:

[0069] ;

[0070] In the formula, For in time The minimum predicted life span value; For in time The highest predicted lifespan value.

[0071] The present invention has the following advantages: extracting set time series data from a time series database of a set scale of energy storage power station, the set time series data is used to improve the TFT prediction model for training; supplementing missing values ​​of the set time series data by spline interpolation method to obtain the supplemented set time series data; standardizing the supplemented set time series data by z-score strategy to obtain standardized time series data; performing abnormal point detection on the standardized time series data by a class of support vector machines to obtain abnormal point data; replacing the abnormal point data by weighted moving average strategy to obtain target time series data; capturing features of periodic data in the target time series data by time convolution network and long short-term memory network to obtain periodic time series data features; capturing features of non-periodic data in the target time series data by multi-layer perceptron to obtain non-periodic time series data features; inputting the periodic time series data features and the non-periodic time series data features into the improved TFT prediction model; performing calculation processing by the improved TFT prediction model to output prediction results. The present invention adopts an improved TFT model as the main framework of the battery life prediction algorithm. Since the battery life prediction contains a large amount of time series data, the advantages of TFT in time series prediction are fully utilized. The multi-head self-attention mechanism and gated recursive unit contained in the TFT model can effectively process the time dependence of long time series data and capture the complex nonlinear relationship in the battery data time series. The present invention innovatively introduces a solution for separating periodic and non-periodic components, and adopts two independent processing components, one for processing periodic change characteristics and the other for processing non-periodic change characteristics, so that the model can more accurately identify and simulate the life changes of batteries under different environments and usage conditions. The present invention introduces advanced data cleaning and anomaly detection algorithms to effectively identify and process noise and outliers in the input data. This not only improves the prediction accuracy of the model, but also enhances the robustness of the model in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the implementation methods or the description of the prior art. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.

[0073] The structures, proportions, sizes, etc. illustrated in this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with the technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantial technical significance. Any structural modification, change in proportion or adjustment of size shall still fall within the scope of the technical contents disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention.

[0074] Figure 1 A schematic flow chart of a method for predicting the life of an energy storage battery based on an improved TFT provided in Example 1 of the present invention;

[0075] Figure 2 A schematic diagram of a specific implementation process of a method for predicting the life of an energy storage battery based on an improved TFT provided in Example 1 of the present invention;

[0076] Figure 3 This is a schematic diagram of the architecture of an energy storage battery life prediction device based on improved TFT provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0077] The following is a description of the implementation of the present invention by specific embodiments. People familiar with the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0078] Example 1

[0079] See also Figure 1 and Figure 2 Embodiment 1 of the present invention provides a method for predicting the life of an energy storage battery based on an improved TFT, comprising the following steps:

[0080] S1. Extracting set time series data from a time series database of a storage power station of a set scale, wherein the set time series data is used to improve the TFT prediction model for training;

[0081] S2, supplement the missing values ​​of the set time series data by spline interpolation method to obtain the supplemented set time series data; standardize the supplemented set time series data by z-score strategy to obtain standardized time series data; perform outlier detection on the standardized time series data by a class of support vector machines to obtain outlier data; replace the outlier data by weighted moving average strategy to obtain target time series data;

[0082] S3. Capturing the features of the periodic data in the target time series data through a time convolution network and a long short-term memory network to obtain the features of the periodic time series data;

[0083] S4. Capturing features of the non-periodic data in the target time series data by using a multi-layer perceptron to obtain features of the non-periodic time series data;

[0084] S5. Input the periodic time series data features and the non-periodic time series data features into the improved TFT prediction model; perform calculation processing through the improved TFT prediction model and output a prediction result.

[0085] In this embodiment, in step S1, set time series data is extracted from a time series database of a set scale of energy storage power station, and the set time series data is used to improve the TFT prediction model for training;

[0086] Specifically, the data required for model training is extracted from the time series database of a 50MW / 100MWH energy storage power station, mainly including: battery temperature, voltage, current, charge and discharge times, external ambient temperature, charge and discharge cycle, battery initial capacity, etc. The data is stored with minute-level timestamps and a time span of 3 years, as well as battery capacity attenuation data (expressed as a percentage) and the actual battery service life (days).

[0087] In this embodiment, in step S2, the missing value supplementation process is performed on the set time series data by using the spline interpolation method to obtain the supplemented set time series data;

[0088] Specifically, the cubic spline interpolation method is used to supplement the missing values. The mathematical expression of the spline interpolation method is:

[0089] ;

[0090] In the formula, is the interpolation function, which means The function value in the interval, and are the horizontal coordinates of adjacent known data points respectively; , , , is the interpolation coefficient and is the model hyperparameter.

[0091] Among them, the interpolation coefficient , , , Determine method:

[0092] 1. Continuity requirement: The interpolation function is continuous at the connection of each interval, that is, and ,in and They are and The known function value at .

[0093] 2. First-order derivative continuity: The first-order derivative of the interpolation function is required to be continuous at the connection point, that is, and continuous.

[0094] 3. Second-order derivative continuity: The second-order derivative of the interpolation function is required to be continuous at the connection point, that is, and continuous.

[0095] In the present invention, the interpolation coefficients are obtained by solving the following linear equations:

[0096] ;

[0097] Since all data is stored at the minute level, that is, the time interval is 1 minute, .

[0098] and The estimation is based on the data of the adjacent data points before and after, using the difference method:

[0099] ;

[0100] In this embodiment, the supplemented set time series data is standardized by a z-score strategy to obtain standardized time series data;

[0101] Specifically, the z-score strategy is used for standardization to transform the data into a form with zero mean and unit variance, which is convenient for subsequent data analysis and processing.

[0102] The mathematical expression of the z-score strategy is:

[0103] ;

[0104] In the formula, is the original data matrix; and are the mean and standard deviation of the data respectively.

[0105] In this embodiment, outlier detection is performed on the standardized time series data by a type of support vector machine to obtain outlier data;

[0106] Specifically, since one-class support vector machine (One-Class SVM) performs better in processing data with complex feature space, this method is used for outlier detection. The radial basis function (RBF) is selected as the model kernel function to find the minimum hypersphere, and the points outside the hypersphere are judged as outliers.

[0107] In this embodiment, the abnormal point data is replaced by a weighted moving average strategy to obtain target time series data;

[0108] Specifically, for the detected abnormal points, a weighted moving average strategy is used to replace the abnormal points to maintain the smoothness and continuity of the time series data. The mathematical expression of the weighted moving average strategy is:

[0109] ;

[0110] In the formula, For time point The data value of is the weight; is the number of neighboring points considered.

[0111] In this embodiment, in step S3, the periodic data in the target time series data is feature captured by a time convolution network and a long short-term memory network to obtain periodic time series data features;

[0112] Specifically, for periodic data such as seasonal data, holiday data, and battery charging and discharging cycles, the periodic time series data features are captured and processed by combining the temporal convolutional network (TCN) and the long short-term memory network (LSTM).

[0113] First, feature extraction is performed through the TCN layer to capture short-term time dependencies and help the model understand the data changes in different time periods. The calculation formula of the time convolution network is:

[0114] ;

[0115] In the formula, is the periodic data input; is the number of filters; is the size of the convolution kernel; It is an activation function used to enhance the nonlinear ability of the network.

[0116] Next is the LSTM layer, which receives the output of the TCN layer to capture the long-term temporal dependencies in the periodic timing components. The calculation formula of the long short-term memory network is:

[0117] ;

[0118] In the formula, The time steps The hidden state and cell state of .

[0119] In this embodiment, in step S4, a multi-layer perceptron is used to capture the features of the non-periodic data in the target time series data to obtain the features of the non-periodic time series data;

[0120] Specifically, for non-periodic data related to batteries, such as the cumulative usage time and charge and discharge status of the battery, a multi-layer perceptron (MLP) is used to process:

[0121] The non-periodic feature components are processed using two fully connected layers, and MLP is used to capture data that does not have obvious periodic patterns, such as the long-term attenuation trend of the battery. The calculation formula of MLP is:

[0122] ;

[0123] In the formula, X is the input non-periodic data matrix; and , are the weights and biases of each layer respectively.

[0124] In this embodiment, in step S5, the periodic time series data features and the non-periodic time series data features are input into the improved TFT prediction model; calculation processing is performed through the improved TFT prediction model to output a prediction result.

[0125] The steps of calculating and processing the improved TFT prediction model are as follows:

[0126] S51, screening the input periodic time series data features and the non-periodic time series data features through a feature selection layer to obtain features that contribute most to battery life prediction;

[0127] Specifically, after feature extraction by the periodic feature processing module and the non-periodic feature processing module, the obtained features enter the feature selection layer. The main task of the feature selection layer is to select the features that contribute most to the battery life prediction from these extracted features and filter out those redundant or irrelevant features, thereby improving the prediction accuracy and efficiency of the model.

[0128] Assume that the feature representation output by the periodic feature extraction module is The feature representation of the output of the non-periodic feature extraction module is Then the input to the feature selection layer is the result of these feature extractions:

[0129] ;

[0130] First, encode the input into a context vector. The context vector generation formula is:

[0131] ;

[0132] In the formula, is the output of the feature extraction layer, is the corresponding weight matrix, is the bias vector and ReLU is the activation function.

[0133] Context vector Further used for feature selection weights , these weights determine the importance of each feature in the model prediction:

[0134] ;

[0135] In the formula, is the hidden state of the previous time step. The gating mechanism of the GRU model is used to control the information flow and dynamically generate weights. .

[0136] Based on the generated weight , the model weights the input features at each time step to decide which features should be paid more attention to and which features should be suppressed:

[0137] ;

[0138] In the formula, It is the feature set after feature selection.

[0139] S52: Through the multi-head self-attention mechanism and gated recursive unit in the core computing layer, the screened features are calculated and processed to output the prediction results.

[0140] Specifically, after the data passes through the feature selection layer, it enters the model calculation core layer. The core calculation layer is responsible for processing the optimized feature data and generating the final prediction results. It includes two parts: multi-head self-attention mechanism and gated recursive unit.

[0141] Among them, the multi-head self-attention mechanism is processed as follows:

[0142] Through linear transformation, the input features are represented Mapped to the query vector (Query), key vector (Key) and value vector (Value):

[0143] ;

[0144] In the formula, , , are the weight matrices for query, key, and value respectively.

[0145] The attention weights are then calculated by taking the dot product of the query vector and the key vector:

[0146] ;

[0147] In the formula, is the dimension of the key vector, which is used to scale the dot product result to maintain numerical stability. The Softmax function normalizes the dot product result to a weight distribution, and then the weighted value vector .

[0148] The multi-head self-attention mechanism divides the query, key, and value vectors into multiple heads and performs the above attention calculations in parallel. The calculation formula for each head is the same as that for a single head, and the final results are combined through splicing and linear transformation:

[0149] ;

[0150] In the formula, Represents the linear transformation matrix of the output.

[0151] Among them, the gated recursive unit is processed as follows:

[0152] The gated recurrent unit is responsible for transferring and updating information between time steps. GRN controls the flow of information through its internal gating mechanism, thereby effectively skipping unnecessary network components when processing long time series data.

[0153] Two gating mechanisms are used: reset gate and update gate. The above two determine the influence of the information in the previous time step on the current time step.

[0154] ;

[0155] ;

[0156] In the formula, It is the reset gate. It is the update gate. represents the Sigmoid activation function, and is the corresponding weight matrix, and is the bias vector.

[0157] Based on the output of the reset gate, GRN calculates candidate hidden states :

[0158] ;

[0159] GRN output status It is the final feature representation of the current time step, which contains the information filtered by the gating mechanism and is used for subsequent processing of the model.

[0160] In this embodiment, after being processed by the core computing layer, the model will generate an interval prediction, indicating the battery's The predicted life expectancy for each day.

[0161] In the prediction stage of the model, the feature representation after feature selection and core computing layer processing Used to generate a future time period In order to reflect the confidence of the model prediction, the output results will be expressed in the form of intervals, providing the upper and lower limits of the predicted life expectancy for each day.

[0162]

[0163] In the formula, It's in time The minimum predicted life span is It's in time The highest predicted lifespan value.

[0164] In a possible embodiment, a specific energy storage battery life prediction example is provided as follows:

[0165] 1. Experimental Environment

[0166] 1. Hardware environment:

[0167] The experiment was conducted on a high-performance computing server with the following hardware configuration:

[0168] Processor (CPU): Intel Xeon E5-2698 v4 @ 2.20GHz, 20 cores

[0169] Memory (RAM): 256 GB DDR4

[0170] Graphics Processing Unit (GPU): NVIDIA Tesla V100 32GB, 4 cards

[0171] Storage: 2 TB SSD

[0172] 2. Software environment:

[0173] The following software and libraries were used in the experiment:

[0174] Operating system: Ubuntu 20.04

[0175] Python version: Python 3.8

[0176] Deep learning framework: PaddlePaddle 2.0

[0177] Other libraries: NumPy 1.19, Pandas 1.2, scikit-learn0.24, Matplotlib 3.3

[0178] Development tools: Pycharm 2021.1

[0179] 3. Dataset:

[0180] The experiment used a data set from a real power station, which includes the following:

[0181] Time series data: temperature, voltage, current, number of charge and discharge times, external ambient temperature, charge and discharge cycle, initial capacity of the battery, etc., recorded once every minute, with a time span of 3 years.

[0182] Battery life-related data: includes battery capacity decay data (expressed as a percentage) and actual battery life (days).

[0183] 4. Model configuration:

[0184] The main hyper-parameter settings of each comparison model are as follows:

[0185] Linear Regression: Default Parameters

[0186] SVM: kernel function is RBF, C=1.0

[0187] LSTM: Two-layer LSTM, 50 hidden units per layer, dropout=0.2

[0188] DeepAR: 3-layer LSTM, 100 hidden units per layer, dropout=0.2

[0189] ConvTrans: 3 layers of convolution, 64 filters per layer, and kernel size 3

[0190] Standard TFT model: 2 layers of TCN, 2 layers of LSTM, 1 layer of self-attention layer, 128 hidden units in each layer.

[0191] 2. Evaluation Indicators

[0192] 1. Mean Square Error (MSE):

[0193] MSE is very sensitive to the sum of squares of prediction errors, especially for large prediction errors. MSE reflects the overall prediction error of the model. The smaller the value, the more accurate the model's prediction.

[0194] ;

[0195] In the formula, is the true value, is the predicted value, is the sample size.

[0196] 2. Mean absolute error MAE:

[0197] MAE is the absolute value of each prediction error, and will not amplify large prediction errors. It reflects the average deviation of the model from the predicted value. The smaller the value, the more accurate the model's prediction.

[0198] ;

[0199] In the formula, is the true value, is the predicted value, is the sample size.

[0200] 3. R-squared value :

[0201] The value indicates the explanatory power of the model. The range is between 0 and 1. The closer the value is to 1, the better the model can explain the changes in the data.

[0202] ;

[0203] In the formula, is the true value, is the predicted value, Represents the mean of the true values.

[0204] 4. Prediction accuracy Acc:

[0205] Define an error threshold ,when , it is believed that the value is predicted accurately.

[0206] ;

[0207] 5. Training time t:

[0208] Record the time from when the model starts predicting to when the prediction is completed. The shorter the time, the lower the time complexity of the model and the more efficient the model.

[0209] 3. Experimental Data

[0210] The experimental data are shown in Table 1.

[0211] Table 1 Experimental data

[0212]

[0213] Among them, IBL-Pred (Improved Battery Life Prediction) is the method proposed in the present invention; the bold values ​​in the table are the optimal values ​​in the current evaluation indicators.

[0214] In summary, the present invention extracts set time series data from a time series database of a set scale of energy storage power station, and the set time series data is used to improve the TFT prediction model for training; the set time series data is supplemented with missing values ​​by spline interpolation method to obtain the supplemented set time series data; the supplemented set time series data is standardized by z-score strategy to obtain standardized time series data; the standardized time series data is detected for abnormal points by a class of support vector machines to obtain abnormal point data; the abnormal point data is replaced by weighted moving average strategy to obtain target time series data; the periodic data in the target time series data is feature captured by time convolution network and long short-term memory network to obtain periodic time series data features; the non-periodic data in the target time series data is feature captured by multi-layer perceptron to obtain non-periodic time series data features; the periodic time series data features and the non-periodic time series data features are input into the improved TFT prediction model; the improved TFT prediction model is used for calculation and processing to output the prediction result. The present invention adopts an improved TFT model as the main framework of the battery life prediction algorithm. Since the battery life prediction contains a large amount of time series data, the advantages of TFT in time series prediction are fully utilized. The multi-head self-attention mechanism and gated recursive unit contained in the TFT model can effectively process the time dependence of long time series data and capture the complex nonlinear relationship in the battery data time series. The present invention innovatively introduces a solution for separating periodic and non-periodic components, and adopts two independent processing components, one for processing periodic change characteristics and the other for processing non-periodic change characteristics, so that the model can more accurately identify and simulate the life changes of batteries under different environments and usage conditions. The present invention introduces advanced data cleaning and anomaly detection algorithms to effectively identify and process noise and outliers in the input data. This not only improves the prediction accuracy of the model, but also enhances the robustness of the model in practical applications.

[0215] It should be noted that the method of the embodiment of the present disclosure can be performed by a single device, such as a computer or a server. The method of the present embodiment can also be applied in a distributed scenario and completed by multiple devices cooperating with each other. In the case of such a distributed scenario, one of the multiple devices can only perform one or more steps in the method of the embodiment of the present disclosure, and the multiple devices will interact with each other to complete the described method.

[0216] It should be noted that the above describes some embodiments of the present disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the above embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0217] Example 2

[0218] See also Figure 3 Embodiment 2 of the present invention further provides a device for predicting the life of an energy storage battery based on an improved TFT, comprising:

[0219] A time series data extraction module 001 is used to extract set time series data from a time series database of a set scale of energy storage power station, and the set time series data is used to improve the TFT prediction model for training;

[0220] The time series data preprocessing module 002 is used to supplement the missing values ​​of the set time series data by spline interpolation method to obtain the supplemented set time series data; standardize the supplemented set time series data by z-score strategy to obtain standardized time series data; perform outlier detection on the standardized time series data by a class of support vector machines to obtain outlier data; replace the outlier data by weighted moving average strategy to obtain target time series data;

[0221] The periodic time series data feature acquisition module 003 is used to capture the features of the periodic data in the target time series data through a time convolution network and a long short-term memory network to obtain the periodic time series data features;

[0222] The non-periodic time series data feature acquisition module 004 is used to capture the features of the non-periodic data in the target time series data through a multi-layer perceptron to obtain the features of the non-periodic time series data;

[0223] The improved TFT prediction model prediction module 005 is used to input the periodic time series data characteristics and the non-periodic time series data characteristics into the improved TFT prediction model; perform calculation processing through the improved TFT prediction model, and output a prediction result.

[0224] In this embodiment, in the time series data extraction module 001, in the process of extracting the set time series data from the time series database of the energy storage power station of a set scale, the set time series data includes: battery temperature, voltage, current, number of charge and discharge times, external ambient temperature, charge and discharge cycle, initial capacity of the battery, battery capacity attenuation data and actual service life of the battery.

[0225] In this embodiment, in the time series data preprocessing module 002, in the process of supplementing the missing values ​​of the set time series data by the spline interpolation method to obtain the supplemented set time series data, the mathematical expression of the spline interpolation method is:

[0226] ;

[0227] In the formula, is the interpolation function, which means The function value in the interval, and are the horizontal coordinates of adjacent known data points respectively; , , , is the interpolation coefficient and is the model hyperparameter.

[0228] In this embodiment, in the time series data preprocessing module 002, in the process of performing standardization processing by the z-score strategy to obtain standardized time series data, the mathematical expression of the z-score strategy is:

[0229] ;

[0230] In the formula, is the original data matrix; and are the mean and standard deviation of the data respectively.

[0231] In this embodiment, in the time series data preprocessing module 002, in the process of replacing the abnormal point data by the weighted moving average strategy to obtain the target time series data, the mathematical expression of the weighted moving average strategy is:

[0232] ;

[0233] In the formula, For time point The data value of is the weight; is the number of neighboring points considered.

[0234] In this embodiment, in the periodic time series data feature acquisition module 003, in the process of capturing the features of the periodic data in the target time series data through the time convolution network and the long short-term memory network, the calculation formula of the time convolution network is:

[0235] ;

[0236] In the formula, is the periodic data input; is the number of filters; is the size of the convolution kernel; It is an activation function used to enhance the nonlinear ability of the network;

[0237] The calculation formula of the long short-term memory network is:

[0238] ;

[0239] In the formula, The time steps The hidden state and cell state of .

[0240] In this embodiment, in the non-periodic time series data feature acquisition module 004, in the process of capturing the features of the non-periodic data in the target time series data through the multi-layer perceptron, the calculation formula of the multi-layer perceptron is:

[0241] ;

[0242] In the formula, X is the input non-periodic data matrix; and , are the weights and biases of each layer respectively.

[0243] In this embodiment, in the improved TFT prediction model prediction module 005, the prediction processing submodule includes:

[0244] A feature selection submodule 051 is used to screen the input periodic time series data features and the non-periodic time series data features through a feature selection layer to obtain the features that contribute most to battery life prediction;

[0245] The core computing submodule 052 is used to calculate and process the selected features through the multi-head self-attention mechanism and gated recursive unit in the core computing layer, and output the prediction results.

[0246] In this embodiment, in the improved TFT prediction model prediction module 005, the prediction result output by the improved TFT prediction model is an interval prediction, which indicates that the battery The predicted life expectancy value for each day within ; the interval representation of the prediction result is:

[0247] ;

[0248] In the formula, For in time The minimum predicted life span value; For in time The highest predicted lifespan value.

[0249] It should be noted that the information interaction, execution process and other contents between the modules of the above-mentioned system are based on the same concept as the method embodiment in Example 1 of the present application, and the technical effects they bring are the same as those of the method embodiment of the present application. For specific contents, please refer to the description in the method embodiment shown above in the present application, and will not be repeated here.

[0250] Example 3

[0251] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium, in which a program code for a method for predicting the life of an energy storage battery based on an improved TFT is stored, and the program code includes instructions for executing the method for predicting the life of an energy storage battery based on an improved TFT of embodiment 1 or any possible implementation thereof.

[0252] Computer-readable storage media can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0253] Example 4

[0254] Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor;

[0255] The processor and the memory communicate with each other via a bus; the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute a method for predicting the life of an energy storage battery based on an improved TFT according to Example 1 or any possible implementation thereof.

[0256] Specifically, the processor can be implemented by hardware or by software. When implemented by hardware, the processor can be a logic circuit, an integrated circuit, etc.; when implemented by software, the processor can be a general-purpose processor implemented by reading software codes stored in a memory. The memory can be integrated in the processor or can be located outside the processor and exist independently.

[0257] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable systems. The computer instructions can be stored in a computer-readable storage medium, or transmitted from a computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transmitted from a website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center.

[0258] Obviously, those skilled in the art should understand that the above modules or steps of the present invention can be implemented by a general computing system, they can be concentrated on a single computing system, or distributed on a network composed of multiple computing systems, and optionally, they can be implemented by a program code executable by a computing system, so that they can be stored in a storage system and executed by the computing system, and in some cases, the steps shown or described can be executed in a different order than here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.

[0259] Although the present invention has been described in detail above by general description and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made to the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all belong to the scope of protection claimed by the present invention.

Claims

1. A method for predicting the life of an energy storage battery based on improved TFT, characterized in that: include: Extracting set time series data from a time series database of a storage power station of a set scale, wherein the set time series data is used to improve the TFT prediction model for training; The missing values ​​of the set time series data are supplemented by the spline interpolation method to obtain the supplemented set time series data; the supplemented set time series data are standardized by the z-score strategy to obtain standardized time series data; the standardized time series data are detected for abnormal points by a class of support vector machines to obtain abnormal point data; the abnormal point data are replaced by the weighted moving average strategy to obtain target time series data; Capturing the features of the periodic data in the target time series data through a time convolution network and a long short-term memory network to obtain the features of the periodic time series data; Capturing features of the non-periodic data in the target time series data by using a multi-layer perceptron to obtain features of the non-periodic time series data; Inputting the periodic time series data features and the non-periodic time series data features into the improved TFT prediction model; performing calculation processing through the improved TFT prediction model and outputting a prediction result; In the process of extracting the set time series data from the time series database of the energy storage power station of the set scale, the set time series data includes: battery temperature, voltage, current, charge and discharge times, external environment temperature, charge and discharge cycle, battery initial capacity, battery capacity attenuation data and actual battery service life; The steps of calculating and processing the improved TFT prediction model are as follows: Through the feature selection layer, the input periodic time series data features and the non-periodic time series data features are screened to obtain the features that contribute most to the battery life prediction; Through the multi-head self-attention mechanism and gated recursive unit in the core computing layer, the selected features are calculated and processed to output the prediction results; The prediction result output by the improved TFT prediction model is an interval prediction, which represents the predicted life value of the battery every day in the future optional time period t; the interval representation of the prediction result is: In the formula, is the minimum predicted life value at time t; is the highest predicted lifetime value at time t.

2. The method for predicting the life of an energy storage battery based on improved TFT according to claim 1, characterized in that: In the process of supplementing the missing values ​​of the set time series data by the spline interpolation method to obtain the supplemented set time series data, the mathematical expression of the spline interpolation method is: S(x)=a i +b i (x-x i )+c i (x-x i ) 2 +d i (x-x i ) 3 for x∈[x i ,x i+1 ] In the formula, S(x) is the interpolation function, which means that in x∈[x i ,x i+1 ]The function value in the interval, x i and x i+1 are the horizontal coordinates of adjacent known data points; a i 、b i 、c i d i is the interpolation coefficient and is the model hyperparameter.

3. The method for predicting the life of an energy storage battery based on improved TFT according to claim 2, characterized in that: In the process of obtaining standardized time series data by performing standardization processing through the z-score strategy, the mathematical expression of the z-score strategy is: Where X is the original data matrix; μ and σ are the mean and standard deviation of the data, respectively.

4. The method for predicting the life of an energy storage battery based on improved TFT according to claim 3, characterized in that: In the process of replacing the abnormal point data by the weighted moving average strategy to obtain the target time series data, the mathematical expression of the weighted moving average strategy is: In the formula, x t+j is the data value at time point t+j; w j is the weight; k is the number of neighboring points considered.

5. The method for predicting the life of an energy storage battery based on improved TFT according to claim 4, characterized in that: In the process of capturing the features of the periodic data in the target time series data through the temporal convolutional network and the long short-term memory network, the calculation formula of the temporal convolutional network is: TCN(X)=ReLU(Conv1D(X`,filters=f,kernel size =k`)) In the formula, X' is the input periodic data; f is the number of filters; k' is the size of the convolution kernel; ReLU(·) is the activation function used to enhance the nonlinear ability of the network; The calculation formula of the long short-term memory network is: h t ,c t =LSTM(TCN(X t ),h t-1 ,c t-1 ) In the formula, h t ,c t are the hidden state and cell state at time step t, respectively.

6. The method for predicting the life of an energy storage battery based on improved TFT according to claim 5, characterized in that: In the process of capturing features of non-periodic data in the target time series data by the multi-layer perceptron, the calculation formula of the multi-layer perceptron is: y=ReLU(W2·(ReLU(W1·X``+b1))+b2) Where X`` is the input non-periodic data matrix; W1, W2 and b1, b2 are the weights and biases of each layer respectively.

7. A device for predicting the life of an energy storage battery based on improved TFT, using a method for predicting the life of an energy storage battery based on improved TFT according to any one of claims 1 to 6, characterized in that: include: A time series data extraction module is used to extract set time series data from a time series database of a set scale of energy storage power station, and the set time series data is used to improve the TFT prediction model for training; A time series data preprocessing module is used to supplement the missing values ​​of the set time series data by spline interpolation method to obtain the supplemented set time series data; standardize the supplemented set time series data by z-score strategy to obtain standardized time series data; perform outlier detection on the standardized time series data by a class of support vector machines to obtain outlier data; replace the outlier data by weighted moving average strategy to obtain target time series data; A periodic time series data feature acquisition module is used to capture the features of the periodic data in the target time series data through a time convolution network and a long short-term memory network to obtain the periodic time series data features; A non-periodic time series data feature acquisition module, used to capture the features of the non-periodic data in the target time series data through a multi-layer perceptron to obtain the features of the non-periodic time series data; The improved TFT prediction model prediction module is used to input the periodic time series data characteristics and the non-periodic time series data characteristics into the improved TFT prediction model; perform calculation processing through the improved TFT prediction model and output a prediction result.

Images