Battery early warning method based on PRform time sequence prediction model
By using the PRformer time series prediction model in battery early warning, combined with the pyramid-shaped recursive neural network embedded with PRE and Transformer encoder, the problem of poor long-term time series prediction performance in the existing technology is solved, and an accurate warning of battery abnormalities is achieved.
Patent Information
- Application Number
- CN202510051296.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-13
- Publication Date
- 2025-06-10
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The prior art has poor performance in long-time series prediction, and it is difficult to accurately warning battery abnormalities through long-time series prediction.
Using a battery early warning method based on the PRformer time series prediction model, the PRE and Transformer encoder are embedded through a pyramid-shaped recursive neural network to learn the relationship between multi-scale time characteristics and multivariables, and time series prediction and anomaly detection are performed.
It improves the performance of long-term series prediction, can accurately warning battery abnormalities, and enhances the abnormality detection capability during battery operation.
Smart Images
Figure CN120122005A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to battery anomaly warning, and specifically to a battery warning method based on a PRformer time series prediction model. Background Art
[0002] During the abnormal warning in the battery operation process, it is usually modeled according to the changes of battery parameters such as voltage, current, and temperature in the time dimension during the charge and discharge process, and the models used are basically prediction models for time series data, such as time series neural network models (LSTM, GRU, etc.). However, whether these models use single variables or multiple variables as inputs, they do not consider the problem of the mutual relationship and influence between features, and have poor performance in long-time series prediction. Summary of the Invention
[0003] (1) Technical Problems to be Solved
[0004] In view of the above-mentioned disadvantages of the prior art, the present invention provides a battery warning method based on a PRformer time series prediction model, which can effectively overcome the defect that it is difficult to accurately warn battery anomalies through long-time series prediction existing in the prior art.
[0005] (2) Technical Solutions
[0006] To achieve the above object, the present invention is realized through the following technical solutions:
[0007] A battery warning method based on a PRformer time series prediction model includes the following steps:
[0008] S1. Collect multi-dimensional time series data on battery parameters during the battery operation process;
[0009] S2. Construct a PRformer time series prediction model, and use the PRformer time series prediction model to perform time series prediction on the multi-dimensional time series data to obtain the prediction result of the multi-dimensional time series data of the battery parameters in the future time period;
[0010] S3. Perform anomaly detection on the prediction result of the multi-dimensional time series data through data analysis, and perform anomaly warning according to the anomaly detection result.
[0011] Preferably, the collection of the multi-dimensional time series data on battery parameters during the battery operation process in S1 includes:
[0012] S11. Collect multi-dimensional time series data on battery parameters during the battery operation process according to a preset collection frequency;
[0013] S12. Fill in missing values, smooth noisy data, remove outliers, resolve inconsistencies, handle redundancies caused by data integration, and filter data with charging durations that do not meet a specific time for the multi-dimensional time series data;
[0014] S13. Normalize or standardize the multi-dimensional time series data;
[0015] Among them, the battery parameters include parameters such as voltage, current, temperature, and capacity.
[0016] Preferably, the normalization or standardization processing of the multi-dimensional time series data in S13 includes:
[0017] Use the following formula to normalize the data x in the multi-dimensional time series data X i for normalization:
[0018]
[0019] where x i ' is the normalized result of the data x i , and min(X) and max(X) are the minimum and maximum values in the multi-dimensional time series data X respectively;
[0020] Use the following formula to standardize the data x in the multi-dimensional time series data X i for standardization:
[0021]
[0022] where x i ” is the standardized result of the data x i , mean(X) is the average value of the multi-dimensional time series data X, and std(X) is the standard deviation of the multi-dimensional time series data X.
[0023] Preferably, in S2, constructing a PRformer time series prediction model includes:
[0024] The PRformer time series prediction model includes a Pyramid Recurrent Neural Network Embedding PRE, a Transformer encoder, and a time series data linear projection layer;
[0025] The Pyramid Recurrent Neural Network Embedding PRE combines a pyramid structure with a Recurrent Neural Network (RNN) to learn multi-scale time features to preserve the time order. Its input is a univariate time series data sequence, and the output is the multi-scale time series representation of this univariate time series data sequence;
[0026] The Transformer encoder models the relationships between variables, treats the multi-scale time series representations corresponding to each single-variable time series data sequence as a token, and uses the self-attention mechanism to calculate and fuse the correlations in the variable dimension;
[0027] The time series data linear projection layer performs time series prediction on the multi-dimensional time series data according to the output result of the Transformer encoder, and obtains the prediction result of the multi-dimensional time series data of the battery parameters in the future time period.
[0028] Preferably, the Pyramidal Recursive Neural Network Embedding (PRE) includes a pyramidal temporal convolutional module, an upsample module, and a recurrent neural network module;
[0029] The pyramidal temporal convolutional module is used to learn the multi-scale convolutional features of the single-variable time series data sequence, and is modeled by using a multi-layer Convolutional Neural Network (CNN). For each layer of the CNN, different granularity features are extracted through different kernel sizes and stride sizes, and the length of the single-variable time series data sequence is compressed through the CNN to facilitate the processing by the recurrent neural network module;
[0030] The pyramidal temporal convolutional module is a key part in constructing the bottom-up path of the pyramidal structure. The bottom-up path gradually constructs the periodic features of the single-variable time series data sequence through one-dimensional convolution, so as to capture the changes on different time scales. The formula is expressed as follows:
[0031] y l =Conv1D(y l-1 ,kernel=K l ,stride=K l );
[0032] where, y l is the output sequence of the l-th layer, y l-1 is the output sequence of the (l-1)-th layer, K l is the kernel size of the l-th layer, Window l 、Window l-1 are the period lengths of the l-th layer and the (l-1)-th layer respectively, and Conv1D(·) represents one-dimensional convolution.
[0033] Preferably, the upsample module is a key part in constructing the top-down path of the pyramidal structure. The top-down path starts from the highest layer in the pyramidal structure and gradually upsamples, transferring the large-scale features to the small scale. The formula is expressed as follows:
[0034] x l-1 ' = upsample(x l );
[0035] where x l-1 ' and x l ' are the upsampling features of the (l - 1)-th layer and the l-th layer respectively, and upsample(·) represents upsampling.
[0036] Preferably, the recurrent neural network module is used to learn the sequence dependence in the time scale at different scales, and encodes the sequences at each time scale by using the temporal modeling ability of the gated recurrent unit GRU. The encoding results at different time scales are fused by weighted fusion to generate a multi-scale time series representation of the univariate time series data. The formula is as follows:
[0037] h (j) = GRU(x j ”, D|layer_num), j = 1, 2,..., l;
[0038] h = CONTACT(β 1 ·h (1) , β 2 ·h (2) ,..., β l ·h (l) );
[0039] where h (j) is the hidden state of the gated recurrent unit GRU corresponding to the j-th layer, and this hidden state serves as the time representation of the j-th layer. x j ” is the fused feature of the j-th layer, x j ” = x j + x j ', x j is the feature of the j-th layer itself, x j ' is the upsampling feature of the j-th layer, D is the embedding dimension of the univariate, layer_num is the number of scales, and the hidden dimension of the gated recurrent unit GRU is GRU(·) represents the processing of the gated recurrent unit GRU;
[0040] h is the multi-scale time series representation of the univariate time series data, β 1 , β 2 , …, β l are the weight coefficients corresponding to each layer, j = 1, 2,..., l, β j is the weight coefficient corresponding to the j-th layer, α j is a trainable parameter initialized to 1 / l, and T is the temperature parameter used to amplify the difference between α j so that the weight coefficient βj Sharper, increasing the weight difference between time representations at different scales, and CONTACT(·) represents the string concatenation function.
[0041] Preferably, in S3, anomaly detection is performed on the prediction results of multi-dimensional time series data through data analysis, and anomaly warnings are given according to the anomaly detection results, including:
[0042] Perform data analysis on the prediction results of multi-dimensional time series data of battery parameters in the future time period, judge whether anomalies occur, and give anomaly warnings according to the anomaly detection results;
[0043] Among them, anomalies include situations where battery parameters exceed preset thresholds and the change curves of battery parameters do not conform to preset rules.
[0044] (III) Beneficial Effects
[0045] Compared with the prior art, the battery warning method based on the PRformer time series prediction model provided by the present invention introduces a pyramidal recurrent neural network embedding PRE to extract multi-scale time series representations of each univariate time series data sequence, uses a Transformer encoder to learn the relationships between multivariate variables, models the relationships between variables, and finally obtains the prediction results of multi-dimensional time series data of battery parameters in the future time period. Then, anomaly detection is performed on the prediction results of multi-dimensional time series data through data analysis, so as to accurately warn of battery anomalies. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings based on these drawings without creative efforts.
[0047] Figure 1 It is a flow chart of the present invention;
[0048] Figure 2 It is an architecture diagram of the PRformer time series prediction model in the present invention;
[0049] Figure 3 It is a structural diagram of the pyramidal recurrent neural network embedding PRE in the PRformer time series prediction model of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0051] A battery warning method based on the PRformer time series prediction model, as Figure 1 shown, S1. Collect multi-dimensional time series data on battery parameters during the operation of the battery, specifically including:
[0052] S11. Collect multi-dimensional time series data on battery parameters during the operation of the battery at a preset collection frequency;
[0053] S12. Fill in missing values, smooth noise data, delete outliers, resolve inconsistencies, handle redundancy caused by data integration, and filter data with charging duration not meeting a specific time for the multi-dimensional time series data;
[0054] S13. Perform normalization or standardization processing on the multi-dimensional time series data;
[0055] Among them, the battery parameters include parameters such as voltage, current, temperature, and capacity.
[0056] Specifically, the normalization or standardization processing of the multi-dimensional time series data in S13 includes:
[0057] Use the following formula to perform normalization processing on the data x in the multi-dimensional time series data X i :
[0058]
[0059] Among them, x i ' is the normalization result of the data x i , and min(X), max(X) are the minimum value and maximum value in the multi-dimensional time series data X respectively;
[0060] Use the following formula to perform standardization processing on the data x in the multi-dimensional time series data X i :
[0061]
[0062] Among them, x i ” is the standardization result of the data x i , mean(X) is the average value of the multi-dimensional time series data X, and std(X) is the standard deviation of the multi-dimensional time series data X.
[0063] S2. Construct a PRformer time series prediction model, and use the PRformer time series prediction model to perform time series prediction on multi-dimensional time series data to obtain the prediction results of the multi-dimensional time series data of battery parameters in the future time period.
[0064] Specifically, constructing the PRformer time series prediction model in S2 includes:
[0065] As Figure 2 shown, the PRformer time series prediction model includes a Pyramid Recurrent Neural Network Embedding (PRE), a Transformer encoder, and a time series data linear projection layer;
[0066] The Pyramid Recurrent Neural Network Embedding (PRE) combines a pyramid structure with a Recurrent Neural Network (RNN) to learn multi-scale time features to preserve the time order. Its input is a univariate time series data sequence, and the output is the multi-scale time series representation of this univariate time series data sequence;
[0067] The Transformer encoder models the relationships between variables, treats the multi-scale time series representations corresponding to each univariate time series data sequence as a token, and uses the self-attention mechanism to calculate and fuse the correlations in the variable dimension;
[0068] The time series data linear projection layer performs time series prediction on multi-dimensional time series data according to the output result of the Transformer encoder to obtain the prediction results of the multi-dimensional time series data of battery parameters in the future time period.
[0069] ① As Figure 3 shown, the Pyramid Recurrent Neural Network Embedding (PRE) includes a pyramid time convolution module, an upsample module, and a recurrent neural network module;
[0070] The pyramid time convolution module is used to learn the multi-scale convolution features of the univariate time series data sequence, and is modeled in the way of a multi-layer Convolutional Neural Network (CNN). For each layer of the Convolutional Neural Network (CNN), different granularity features are extracted through different kernel sizes and stride sizes, and the length of the univariate time series data sequence is compressed through the Convolutional Neural Network (CNN) to facilitate the processing of the recurrent neural network module;
[0071] The pyramid time convolution module is the key part of constructing the bottom-up path in the pyramid structure. The bottom-up path gradually constructs the periodic features of the univariate time series data sequence through one-dimensional convolution, so as to capture the changes on different time scales. The formula is as follows:
[0072] yl = Conv1D(y l-1 , kernel = K l , stride = K l );
[0073] Among them, y l is the output sequence of the l-th layer, and y l-1 is the output sequence of the (l - 1)-th layer. K l is the convolutional kernel size of the l-th layer, Window l , Window l-1 are the period lengths of the l-th layer and the (l - 1)-th layer respectively. Conv1D(·) represents one-dimensional convolution.
[0074] ② As Figure 3 shown, the upsample module is a key part in constructing the top-down path of the pyramid structure. The top-down path starts from the highest layer in the pyramid structure and gradually upsamples, transferring large-scale features to small scales. The formula is as follows:
[0075] x l-1 ' = upsample(x l ');
[0076] Among them, x l-1 ' and x l ' are the upsampled features of the (l - 1)-th layer and the l-th layer respectively. upsample(·) represents upsampling.
[0077] ③ As Figure 3 shown, the recurrent neural network module is used to learn the sequence dependencies at different scales in the time scale. It encodes the sequences at each time scale using the temporal modeling ability of the gated recurrent unit GRU. The encoded results at different time scales are fused by weighted summation to generate a multi-scale time series representation of the univariate time series data. The formula is as follows:
[0078] h (j) = GRU(x j ”, D|layer_num), j = 1, 2,..., l;
[0079] h = CONTACT(β 1 ·h (1) , β 2 ·h (2) ,..., β l ·h (l) );
[0080] Among them, h (j)is the hidden state of the gated recurrent unit (GRU) corresponding to the j-th layer, and this hidden state serves as the temporal representation of the j-th layer, x j ” is the fused feature of the j-th layer, x j ” = x j + x j ', x j is the feature of the current layer of the j-th layer, x j ' is the upsampled feature of the j-th layer, D is the embedding dimension of a single variable, layer_num is the number of scales, and the hidden dimension of the gated recurrent unit (GRU) is GRU(·) represents the processing of the gated recurrent unit (GRU);
[0081] h is the multi-scale time series representation of the single-variable time series data sequence, β 1 、β 2 、…、β l are the weight coefficients corresponding to each layer, j = 1, 2,..., l, β j is the weight coefficient corresponding to the j-th layer, α j is a trainable parameter initialized to 1 / l, and T is the temperature parameter used to amplify the j difference between them, making the weight coefficient β j more sharp, increasing the weight difference between different scale time representations, and CONTACT(·) represents the string concatenation function.
[0082] In the technical solution of this application, the PRformer time series prediction model combines the Pyramid Recursive Network Embedding (PRE) and the Transformer encoder, utilizes the advantages of the Recurrent Neural Network (RNN) in processing time series data sequences, and at the same time uses the Transformer encoder to learn the relationships between multi-variables for variable relationship modeling, effectively improving the performance of long-time series prediction.
[0083] S3. Perform anomaly detection on the prediction results of multi-dimensional time series data through data analysis, and issue anomaly warnings according to the anomaly detection results, specifically including:
[0084] Perform data analysis on the prediction results of the multi-dimensional time series data of battery parameters in the future time period, determine whether anomalies occur, and issue anomaly warnings according to the anomaly detection results;
[0085] Among them, anomalies include situations where battery parameters exceed the preset threshold and the change curve of battery parameters does not conform to the preset rules, etc.
[0086] 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 foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.
Claims
1. A battery early warning method based on the PRformer time series prediction model, characterized by: The following steps are involved: S1. Collect multi-dimensional time series data about battery parameters during battery operation; S2. Construct a PRformer time series prediction model, use the PRformer time series prediction model to perform time series prediction on multi-dimensional time series data, and obtain the multi-dimensional time series data prediction results of battery parameters in the future time period; S3. Perform anomaly detection on the prediction results of multi-dimensional time series data through data analysis, and issue anomaly warning based on the anomaly detection results.
2. The battery early warning method based on the PRformer time series prediction model according to claim 1 is characterized in that: S1 collects multi-dimensional time series data about battery parameters during battery operation, including: S11, collecting multi-dimensional time series data about battery parameters during battery operation according to a preset collection frequency; S12, filling missing values, smoothing noise data, deleting outliers, resolving inconsistencies, processing redundancy caused by data integration, and filtering data whose charging time does not meet a specific time for multi-dimensional time series data; S13, normalizing or standardizing the multidimensional time series data; Among them, battery parameters include voltage, current, temperature and capacity.
3. The battery early warning method based on the PRformer time series prediction model according to claim 2 is characterized in that: S13 performs normalization or standardization on the multi-dimensional time series data, including: The following formula is used to analyze the data x in the multidimensional time series data X i Perform normalization: Among them, x i 'For data x i The normalized result of , min(X) and max(X) are the minimum and maximum values in the multidimensional time series data X respectively; The following formula is used to analyze the data x in the multidimensional time series data X i To standardize: Among them, x i " is the data x i The standardized result of , mean(X) is the mean value of the multidimensional time series data X, and std(X) is the standard deviation of the multidimensional time series data X.
4. The battery early warning method based on the PRformer time series prediction model according to claim 2 is characterized in that: The PRformer time series prediction model is constructed in S2, including: The PRformer time series prediction model includes a pyramidal recurrent neural network embedding PRE, a Transformer encoder, and a linear projection layer for time series data; Pyramid recurrent neural network embedding PRE combines the pyramid structure with the recurrent neural network RNN to learn multi-scale time features to preserve the time order. Its input is a univariate time series data sequence, and its output is a multi-scale time series representation of the univariate time series data sequence. Transformer encoder models the relationship between variables, treats the multi-scale time series representation corresponding to each univariate time series data sequence as a token, and uses the self-attention mechanism to calculate and fuse the correlation in the variable dimension; The linear projection layer of time series data performs time series prediction on the multi-dimensional time series data according to the output results of the Transformer encoder to obtain the multi-dimensional time series data prediction results of the battery parameters in the future time period.
5. The battery early warning method based on the PRformer time series prediction model according to claim 4 is characterized in that: The pyramidal recursive neural network embedding PRE includes a pyramidal temporal convolution module, an upsample module and a recursive neural network module; The pyramidal temporal convolution module is used to learn the multi-scale convolution features of the univariate time series data sequence, and is modeled by a multi-layer convolutional neural network (CNN). For each layer of the convolutional neural network (CNN), different convolution kernel sizes (kernelsize) and stride sizes (stride sizes) are used to extract features of different granularities, and the convolutional neural network (CNN) is used to compress the length of the univariate time series data sequence, so that the recursive neural network module can process it. The pyramidal time convolution module is the key part of the bottom-up path in building the pyramidal structure. The bottom-up path gradually builds the periodic characteristics of the univariate time series data sequence through one-dimensional convolution, thereby capturing changes on different time scales. The formula is as follows: y l =Conv1D(y l-1 ,kernel=K l ,stride=K l ); Among them, y l is the output sequence of the lth layer, y l-1 is the output sequence of the l-1th layer, K l is the convolution kernel size of the lth layer, Window l 、Window l-1 are the period lengths of the lth layer and the l-1th layer respectively, and Conv1D(·) represents one-dimensional convolution.
6. The battery early warning method based on the PRformer time series prediction model according to claim 5 is characterized in that: The upsample module is a key part of the top-down path in building a pyramid structure. The top-down path starts from the highest layer in the pyramid structure and gradually upsamples to transfer large-scale features to small-scale features. The formula is as follows: x l-1 '=upsample(x l '); Among them, x l-1 ', x l ' are the upsampled features of the l-1th layer and the lth layer respectively, and upsample(·) represents upsampling.
7. The battery early warning method based on the PRformer time series prediction model according to claim 6 is characterized in that: The recursive neural network module is used to learn the sequence dependencies at different scales on the time scale, and uses the time series modeling capability of the gated recurrent unit GRU to encode the sequences of each time scale. The encoding results of different time scales are weighted fused to generate a multi-scale time series representation of the univariate time series data sequence. The formula is as follows: h (j) =GRU(x j ”,D|layer_num),j=1,2,…,l; h=CONTACT(β1·h (1) ,β2·h (2) ,…,β l ·h (l) ); Among them, h (j) is the hidden state of the gated recurrent unit GRU corresponding to the jth layer, which is used as the time representation of the jth layer, x j ” is the fusion feature of the jth layer, x j ” = x j +x j ', x j is the j-th layer feature, x j ' is the upsampled feature of the jth layer, D is the embedding dimension of the univariate, layer_num is the number of scales, and the hidden dimension of the gated recurrent unit GRU is GRU(·) represents gated recurrent unit GRU processing; h is the multi-scale time series representation of the univariate time series data sequence, β1, β2, …, β l is the weight coefficient corresponding to each layer, j=1,2,…,l,β j is the weight coefficient corresponding to the jth layer, α j is a trainable parameter initialized to 1 / l, T is a temperature parameter used to amplify α j The difference between them makes the weight coefficient β j It is sharper, increasing the weight difference between time representations of different scales. CONTACT(·) represents the string concatenation function.
8. The battery early warning method based on the PRformer time series prediction model according to claim 4 is characterized in that: In S3, anomaly detection is performed on the prediction results of multi-dimensional time series data through data analysis, and anomaly warning is issued based on the anomaly detection results, including: Perform data analysis on the multi-dimensional time series data prediction results of battery parameters in the future time period to determine whether anomalies occur and issue abnormal warnings based on the abnormal detection results; Among them, the anomalies include battery parameters exceeding preset thresholds and battery parameter change curves not meeting preset rules.