Thermal runaway early warning method for energy storage battery based on fusion time domain characteristics

Abstract

According to the energy storage battery thermal runaway early warning method based on fusion time domain characteristics, the comprehensive operation parameters of the energy storage battery are obtained, and then the prediction model comprising the frequency domain branch network and the time domain branch network is established, so that the time domain characteristics and the frequency domain characteristics of the parameters are effectively fused, and the thermal runaway early warning accuracy of the energy storage battery is improved. Therefore, a slow but continuous temperature rise or gas production rate change rule can be accurately identified, early warning can be given out in an extremely early stage of thermal runaway chain reaction acceleration, precious emergency response time is won for operation and maintenance personnel, and the frequency domain branch network can effectively filter random interference and focus on stable mode change, so that the reliability of the system is improved. Therefore, the false alarm rate of the whole process is low.

Description

Technical Field

[0001] This invention relates to a method for early warning of thermal runaway in power systems, and more particularly to a method for early warning of thermal runaway in energy storage batteries based on fused time-domain features. Background Technology

[0002] With the rapid development of the new energy industry, large-scale electrochemical energy storage power stations have become a key support for grid peak shaving and frequency regulation, as well as renewable energy consumption. However, the inherent thermal runaway risk of lithium-ion batteries is the core bottleneck restricting the safe operation of energy storage power stations. Once thermal runaway occurs, it can easily trigger a chain reaction, leading to catastrophic accidents such as fires and explosions.

[0003] In existing technologies, thermal runaway early warning for energy storage batteries typically uses a single parameter, such as voltage or surface temperature. An alarm is triggered when this parameter exceeds a preset, fixed safety threshold. This approach suffers from severe warning lag, missing the optimal intervention window. This is because thermal runaway is a chain-reaction exothermic chemical reaction process. When macroscopic parameters such as temperature and voltage significantly exceed the threshold, the internal side reactions of the battery have often entered an accelerated or even irreversible phase, at which point the warning is too late. With technological advancements, machine learning-based thermal runaway early warning methods have been proposed. These methods extract statistical features (such as mean, variance, slope, and kurtosis) from historical data like voltage and temperature, and then use classifiers such as Support Vector Machines (SVM), Random Forests (RF), and Logistic Regression to determine normality and abnormality. However, this technology has the following drawback: most traditional machine learning methods treat data at each time point as independent samples, disrupting the temporal continuity of the data. Even considering sliding windows, it is difficult to effectively model long-range dependencies and dynamic evolution trends, which are key characteristics of the thermal runaway process. Using a single type of network cannot capture both the temporal dependencies of long sequences and uncover stable patterns or local mutations in the frequency domain. This results in insufficient and inadequate feature extraction.

[0004] Therefore, in order to solve the above-mentioned technical problems, it is urgent to propose a new technical approach. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a thermal runaway early warning method for energy storage batteries based on fused time-domain features. By acquiring relatively comprehensive operating parameters of the energy storage battery, and then establishing a prediction model including frequency domain branch networks and time domain branch networks, the time-domain and frequency-domain features of the parameters are effectively fused. This allows for accurate identification of slow but continuous temperature rise or gas production rate changes, and early warning can be issued at the very early stage of the accelerated thermal runaway chain reaction, thus gaining valuable emergency response time for maintenance personnel. Moreover, the frequency domain branch network can effectively filter random interference and focus on stable pattern changes, resulting in a low false alarm rate throughout the process.

[0006] This invention provides a method for early warning of thermal runaway in energy storage batteries based on fused time-domain features, comprising the following steps:

[0007] S1. Obtain the operating parameters of the sample energy storage battery, including normal operating parameters and abnormal operating parameters;

[0008] S2. Preprocess the operating parameters of the energy storage battery;

[0009] S3. Extract the frequency domain parameters by using Fast Fourier Transform to process the preprocessed running parameters;

[0010] S4. Construct a prediction model, which includes a frequency domain branch network, a time domain branch network, a feature fusion module, and an output module;

[0011] The time-domain parameters after preprocessing in step S2 are input into the time-domain branch network, and the frequency-domain parameters extracted in step S3 are input into the frequency-domain branch network. The output features of the frequency-domain branch network and the time-domain branch network are input into the feature fusion module for feature fusion. The fused features output by the feature fusion module are output to the output module, and the output module outputs the probability of thermal runaway risk.

[0012] S5. Obtain the operating parameters of the energy storage battery under test. After processing the operating parameters under test through steps S2 and S3, input them into the trained prediction model to obtain the thermal runaway risk probability of the energy storage battery under test.

[0013] S6. Provide early warning based on the probability of thermal runaway risk to be tested.

[0014] Furthermore, in steps S1 and S5, the operating parameters of the energy storage battery include voltage, current, temperature, current change rate, temperature change rate, and gas concentration data.

[0015] Furthermore, the preprocessing of the operating parameters of the energy storage battery in step S2 specifically includes:

[0016] A moving average filter is used to suppress noise in the operating parameters;

[0017] Linear interpolation is used to interpolate the data points in the operating parameters.

[0018] The interpolated running parameters are normalized using Z-score.

[0019] Furthermore, the time-domain branching network is a gated recurrent unit, and the gated recurrent unit outputs a time-domain feature vector.

[0020] Furthermore, the frequency domain branch network includes convolutional layers, ReLU activation layers, max pooling layers, flattening layers, and fully connected layers;

[0021] The input frequency domain parameters of the convolutional layer, the output features of the convolutional layer are input to the ReLU activation layer for processing, the output features of the ReLU activation layer are input to the max pooling layer for processing, the output features of the max pooling layer are input to the flattening layer, the output features of the flattening layer are input to the fully connected layer, and the fully connected layer outputs the frequency domain feature vector.

[0022] Furthermore, the output module includes a fully connected layer, a Dropout layer, and a sigmoid activation module;

[0023] The input features of the fully connected layer are fused with the output features of the fully connected layer. The output features of the fully connected layer are then fed into the Dropout layer. The output features of the Dropout layer are then fed into the sigmoid activation module. The sigmoid activation module outputs the thermal runaway prediction probability.

[0024] Furthermore, the early warning based on the probability of thermal runaway risk to be measured specifically includes:

[0025] When the predicted probability of thermal runaway is P≤a1, the current risk is determined to be low, and only the operating parameters of the energy storage battery are uploaded.

[0026] When the predicted probability of thermal runaway is a1<P≤a2, the current period is the early warning period. At this time, the sampling frequency of the energy storage battery operating parameters is increased and an early warning message is issued. The early warning message includes the battery ID, the current probability of thermal runaway, and abnormal parameter values.

[0027] When the predicted probability of thermal runaway is a2<P≤a3, the current period is a high-risk warning period. At this time, safety intervention is carried out. The safety intervention is to limit the charging and discharging power of the energy storage battery, control the cooling system of the current energy storage battery to operate at maximum power, and output high-risk warning information.

[0028] When the predicted probability of thermal runaway is P > a3, it is determined that thermal runaway is about to occur or is currently occurring. At this time, the electrical circuit of the current energy storage battery is cut off, fire suppression measures are triggered, and fire warning information is output.

[0029] The beneficial effects of this invention are as follows: By obtaining relatively comprehensive operating parameters of the energy storage battery, and then establishing a prediction model including frequency domain branch networks and time domain branch networks, the time domain and frequency domain characteristics of the parameters are effectively fused. This enables the accurate identification of slow but continuous temperature rise or gas production rate changes, and can issue early warnings in the very early stages of thermal runaway chain reaction acceleration, thus winning valuable emergency response time for operation and maintenance personnel. Moreover, the frequency domain branch network can effectively filter random interference and focus on stable pattern changes, resulting in a low false alarm rate throughout the process. Attached Figure Description

[0030] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0031] Figure 1 This is a schematic diagram of the process of the present invention.

[0032] Figure 2 This is a schematic diagram of the prediction model structure of the present invention.

[0033] Figure 3 This is a schematic diagram of the frequency domain branching network structure of the present invention. Detailed Implementation

[0034] The present invention will be further described in detail below:

[0035] This invention provides a method for early warning of thermal runaway in energy storage batteries based on fused time-domain features, comprising the following steps:

[0036] S1. Obtain the operating parameters of the sample energy storage battery. The operating parameters include normal operating parameters and abnormal operating parameters. Normal operating parameters are those in which all parameters of the energy storage battery are within the allowable range, i.e., normal values. Abnormal operating parameters are labeled as overcharge, over-discharge, internal short circuit, etc. The parameters are obtained by conducting overcharge, over-discharge, and internal short circuit tests in the experimental environment of the sample energy storage battery in the parameter acquisition chamber.

[0037] The operating parameters include voltage, current, temperature, current change rate, temperature change rate, and gas concentration data. The current change rate and temperature change rate are calculated based on the current and temperature data, while the current, temperature, and gas concentration are obtained using existing sensors. The gas concentration sensor is selected according to actual needs. Generally, the gases released by energy storage batteries during operation include hydrogen, carbon monoxide, methane, ethylene, ethane, hydrogen fluoride, hydrogen chloride, hydrogen sulfide, sulfur dioxide, etc. Therefore, in practice, the appropriate gas sensor is determined based on different battery types and the internal medium conditions of different batteries.

[0038] S2. Preprocess the operating parameters of the energy storage battery;

[0039] The process specifically includes:

[0040] A moving average filter is used to suppress noise in the operating parameters; the moving average filter is based on a queue structure. Linear time-invariant system By calculating data points within the sliding window Arithmetic mean To suppress the signal Random noise In addition to short-term fluctuations, this filter features low-pass characteristics and a linear phase response, and the noise suppression strength and signal fidelity can be balanced by adjusting the window size;

[0041] Linear interpolation is used to interpolate the exact data points in the operating parameters; through this process, the exact data can be supplemented, thereby smoothing the data and facilitating subsequent processing.

[0042] The interpolated running parameters are then normalized using Z-score, and the specific calculation formula is as follows:

[0043] ; where μ and σ are parameters The mean and standard deviation on the training set are used in this step to eliminate dimensional differences and accelerate the convergence of the prediction model.

[0044] S3. Extract the frequency domain parameters by using Fast Fourier Transform to process the preprocessed running parameters;

[0045] For a sequence x[n] of length N, its discrete Fourier transform is:

[0046]

[0047] L before extraction f The amplitude spectrum |X[k]| corresponding to each frequency point forms the frequency domain characteristic representation of the signal.

[0048] For time-domain parameters, i.e. the parameters processed in step S2, the sensors for each parameter output in a time sequence when acquiring the parameters, so their output signal is a time-domain data.

[0049] S4. Construct a prediction model, which includes a frequency domain branch network, a time domain branch network, a feature fusion module, and an output module;

[0050] The time-domain parameters after preprocessing in step S2 are input into the time-domain branch network, and the frequency-domain parameters extracted in step S3 are input into the frequency-domain branch network. The output features of the frequency-domain branch network and the time-domain branch network are input into the feature fusion module for feature fusion. The fused features output by the feature fusion module are output to the output module, and the output module outputs the probability of thermal runaway risk.

[0051] Specifically: The time-domain branch network is a gated recurrent unit (GRU). This branch takes a time-domain matrix as input. The GRU network is an existing technology that regulates information flow by updating zt and resetting the gate rt. Its internal computation process is as follows:

[0052] ;

[0053] The frequency domain branch network is a CNN network that captures weak anomalous harmonic or frequency component changes in voltage, current, and gas signals in the frequency domain. These changes may appear tens of seconds to minutes before a significant temperature rise. Simultaneously, the GRU performs deep analysis of time-series trends to identify slow but continuous changes in temperature rise or gas production rates. This combination of "microscopic frequency domain anomaly capture" and "macroscopic time domain trend judgment" enables the system to issue early warnings at the very early stages of accelerated thermal runaway chain reactions, buying valuable emergency response time for maintenance personnel. This shifts the focus of safety management from "accident handling" to "risk prevention," significantly improving the intrinsic safety level of the energy storage system. Specifically:

[0054] The frequency domain branch network includes convolutional layers, ReLU activation layers, max pooling layers, flattening layers, and fully connected layers;

[0055] The input frequency domain parameters of the convolutional layer, the output features of the convolutional layer are input to the ReLU activation layer for processing, the output features of the ReLU activation layer are input to the max pooling layer for processing, the output features of the max pooling layer are input to the flattening layer, the output features of the flattening layer are input to the fully connected layer, and the fully connected layer outputs the frequency domain feature vector.

[0056] The feature fusion module is a splicing module, namely the Concat module, which is a prior art technology;

[0057] The output module includes a fully connected layer, a Dropout layer, and a sigmoid activation module.

[0058] The input features of the fully connected layer are fused with the output features of the fully connected layer. The output features of the fully connected layer are then fed into the Dropout layer, and the output features of the Dropout layer are fed into the sigmoid activation module. The sigmoid activation module outputs the thermal runaway prediction probability. The formula for the output risk probability is:

[0059] ;

[0060] The entire prediction model uses the cross-entropy loss function from the Second Academy:

[0061] ;

[0062] Where yi is the true label (0 or 1) and pi is the model prediction probability.

[0063] By employing a "dual-branch deep learning model + multi-source data fusion" mechanism, the accuracy of the judgment is fundamentally improved. On one hand, CNNs have inherent noise robustness in extracting frequency domain features, effectively filtering random interference and focusing on stable pattern changes. On the other hand, the model automatically learns the complex nonlinear correlations and cross-validation relationships between multi-dimensional electrical, thermal, and gas data, without relying on any single unreliable signal. This comprehensive judgment based on multi-dimensional deep correlations makes the system more accurate in identifying real risks and more immune to various interferences, thereby reducing the false alarm rate and false negative rate to extremely low levels.

[0064] S5. Obtain the operating parameters of the energy storage battery under test. After processing the operating parameters under test through steps S2 and S3, input them into the trained prediction model to obtain the thermal runaway risk probability of the energy storage battery under test.

[0065] S6. Provide early warning based on the probability of thermal runaway risk to be tested, specifically:

[0066] When the predicted probability of thermal runaway is P≤a1, the current risk is determined to be low, and only the operating parameters of the energy storage battery are uploaded.

[0067] When the predicted probability of thermal runaway is a1<P≤a2, the current period is the early warning period. At this time, the sampling frequency of the energy storage battery operating parameters is increased and an early warning message is issued. The early warning message includes the battery ID, the current probability of thermal runaway, and abnormal parameter values.

[0068] When the predicted probability of thermal runaway is a2<P≤a3, the current period is a high-risk warning period. At this time, safety intervention is carried out. The safety intervention is to limit the charging and discharging power of the energy storage battery, control the cooling system of the current energy storage battery to operate at maximum power, and output high-risk warning information.

[0069] When the predicted probability of thermal runaway is P > a3, it is determined that thermal runaway is about to occur or is currently occurring. At this time, the electrical circuit of the current energy storage battery is cut off, fire suppression measures are triggered, and fire warning information is output.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A thermal runaway early warning method for energy storage batteries based on fusion time domain features, characterized by: The method comprises the following steps: S1. Obtain sample energy storage battery operation parameters, the operation parameters comprising normal operation parameters and abnormal operation parameters; S2. Preprocess the operation parameters of the energy storage battery; S3. Extract frequency domain parameters by using fast Fourier transform to process the preprocessed operation parameters; S4. Construct a prediction model, the prediction model comprising a frequency domain branch network, a time domain branch network, a feature fusion module, and an output module; The time domain parameters after preprocessing in step S2 are input into the time domain branch network, and the frequency domain parameters extracted in step S3 are input into the frequency domain branch network. The output features of the frequency domain branch network and the time domain branch network are input into the feature fusion module for feature fusion. The fusion features output by the feature fusion module are output to the output module, and the output module outputs a thermal runaway risk probability; S5. Obtain the operation parameters of the energy storage battery to be tested. After the operation parameters to be tested are processed through steps S2 and S3, the thermal runaway risk probability of the energy storage battery to be tested is obtained by inputting the operation parameters to be tested into the trained prediction model; S6. Perform early warning according to the thermal runaway risk probability to be tested. 2.The energy storage battery thermal runaway early warning method based on fusion time domain features according to claim 1, characterized in that: In steps S1 and S5, the operation parameters of the energy storage battery comprise voltage, current, temperature, current change rate, temperature change rate, and gas concentration data. 3.The energy storage battery thermal runaway early warning method based on fusion time domain features according to claim 1, characterized in that: In step S2, preprocessing the operation parameters of the energy storage battery comprises: using a moving average filter to perform noise suppression processing on the operation parameters; using a linear interpolation method to perform interpolation processing on the data outliers in the operation parameters; performing Z-score normalization processing on the operation parameters after interpolation processing.

4. The energy storage battery thermal runaway early warning method based on fusion time domain features according to claim 1, characterized in that: The time domain branch network is a gated recurrent unit, and the gated recurrent unit outputs a time domain feature vector.

5. The method of claim 1, wherein the method further comprises: The frequency domain branch network comprises a convolution layer, a ReLU activation layer, a max pooling layer, a flattening layer, and a fully connected layer; The input frequency domain parameters of the convolution layer, the output features of the convolution layer are input into the ReLU activation layer for processing, the output features of the ReLU activation layer are input into the max pooling layer for processing, the output features of the max pooling layer are input into the flattening layer, the output features of the flattening layer are input into the fully connected layer, and the fully connected layer outputs a frequency domain feature vector.

6. The method of claim 1, wherein the method further comprises: The output module comprises a fully connected layer, a Dropout layer, and a sigmoid activation module; The input features of the fully connected layer are the output features of the feature fusion module, the output features of the fully connected layer are input into the Dropout layer, the output features of the Dropout layer are input into the sigmoid activation module, and the sigmoid activation module outputs a thermal runaway prediction probability.

7. The method of claim 1, wherein the method further comprises: The early warning according to the thermal runaway risk probability to be tested comprises: when the thermal runaway prediction probability P≤a1, it is determined that the current is a low risk, at which time only the operation parameters of the energy storage battery are uploaded; when the thermal runaway prediction probability is a1 When the thermal runaway prediction probability is a2 When the thermal runaway prediction probability is P>a3, it is determined that the current thermal runaway is about to occur or is occurring, at which time the electrical circuit of the current energy storage battery is cut off, and a fire suppression measure is triggered, while outputting a fire warning information.

Images