Industrial and commercial energy storage thermal runaway early warning method based on multi-source cooperation
Through the multi-source collaborative thermal runaway early warning method, the multi-physical coupling model and multi-task learning technology are used to solve the problem of insufficient thermal runaway monitoring in energy storage power plants, and efficient and accurate thermal runaway early warning and data transmission are achieved, thereby improving energy storage The safety and operation efficiency of the power station.
Patent Information
- Application Number
- CN202411040911.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-31
- Publication Date
- 2024-11-22
AI Technical Summary
Inadequate thermal runaway monitoring in energy storage power plants leads to a sharp rise in battery temperature, which may cause fire accidents, and the existing technology is difficult to effectively monitor and early warning, affecting battery life and operation and maintenance costs.
Using multi-source collaborative thermal runaway early warning method, deploy multiple sensors for data acquisition and cleaning, establish a multi-physical coupling model, and use time Transformer and nuclear regression module to deal with data loss and different sampling frequency problems, and combine multi-task learning to perform thermal runaway Early warning.
It realizes efficient monitoring and early warning of thermal runaway in energy storage power stations, improves the intelligence level and monitoring efficiency of the system, enhances the accuracy of early warning and the reliability of data transmission, reduces operation and maintenance costs, and provides safety for energy storage power stations. Stable operation provides guarantees.
Smart Images

Figure CN119007397A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the application field of deep learning in power systems, and in particular to an industrial and commercial energy storage thermal runaway early warning method based on multi-source collaboration. Background Art
[0002] In the energy field, with the large-scale application of renewable energy and the rapid development of smart grids, energy storage power stations are becoming increasingly important as a key link in balancing power supply and demand and improving energy efficiency. However, the operating efficiency and safety of energy storage power stations are restricted by many factors, among which thermal management is particularly critical.
[0003] The battery pack in the energy storage power station will generate a lot of heat during operation, and thermal runaway refers to a chain reaction of continuous heat release from the battery. At this time, the temperature of the battery pack rises sharply, which in turn causes a battery combustion accident. Thermal runaway has three processes: induction, occurrence, and spread. The main causes of thermal runaway are overheating, overcharging, internal short circuit, collision and other factors. If thermal runaway monitoring and early warning cannot be effectively carried out, it will not only affect the charging and discharging efficiency and life of the battery, but may also cause thermal runaway of the energy storage power station and even cause safety accidents such as fire. Therefore, an accurate and efficient thermal runaway monitoring system is crucial to ensure the stable operation of the energy storage power station.
[0004] The multi-source collaborative thermal runaway monitoring strategy can improve the overall performance of energy storage power stations. Thermal runaway in energy storage power stations will have signs in many aspects, including ambient temperature, ambient humidity, battery status, etc. By adopting a multi-source collaborative thermal runaway monitoring strategy, various influencing factors can be comprehensively considered to achieve more accurate monitoring and control. This can not only improve the operating efficiency of energy storage power stations, but also extend the service life of batteries, reduce operation and maintenance costs, and provide strong support for the sustainable development of the energy industry. Summary of the invention
[0005] In order to overcome the deficiencies in the prior art, the present invention provides an industrial and commercial energy storage thermal runaway early warning method based on multi-source collaboration, which is used to improve the overall reliability and safety of the system and achieve faster real-time decision-making.
[0006] In order to achieve the above-mentioned invention object, the technical solution adopted to solve the technical problem is as follows:
[0007] A thermal runaway warning method for industrial and commercial energy storage based on multi-source collaboration includes the following steps:
[0008] Step 1: Deploy temperature sensors, humidity sensors, vibration sensors, current sensors, voltage sensors, and gas concentration sensors in the battery compartment of the energy storage power station and collect data;
[0009] Step 2: Check the integrity, handle outliers, duplicate values, and noise of the extracted data;
[0010] Step 3: Consider the mutual influence and coupling relationship between various physical parameters and establish a multi-physics coupling model;
[0011] Step 4: Compare the thermal status score with the threshold to determine whether an alarm is required;
[0012] Step 5: Upload system data.
[0013] Furthermore, step 1 includes the following:
[0014] Temperature sensors, humidity sensors, vibration sensors, current sensors, voltage sensors and gas concentration sensors are deployed in the battery compartment of the energy storage power station, and then the sensor data are collected through monitoring equipment; among them, the sampling frequency of current data, voltage data and vibration data is 1Hz, and the sampling frequency of temperature data, humidity data and gas concentration data is 0.2Hz.
[0015] Furthermore, step 2 includes the following:
[0016] First, we go through each data point in each dimension and check the completeness of the data to ensure that all required fields are filled in.
[0017] Then, the Z-score statistical method is used to calculate the degree of deviation of each data point from the mean, and a threshold is set to identify outliers. For outliers with a degree of deviation higher than the threshold, they are deleted or replaced with the mean. The formula of the Z-score statistical method is as follows:
[0018]
[0019] Among them, X is the observed value of the data, μ is the mean value of the group of data, and σ is the standard deviation of the group of data;
[0020] During the data collection process, the median filtering method is used for smoothing and denoising. The formula is as follows:
[0021] y[n]=Median(x[nk],...,x[n],...,x[n+k])
[0022] The value of each data point is replaced by the median value of its surrounding data points, where k is the size of the filter.
[0023] Furthermore, step 3 includes the following:
[0024] Since most industrial and commercial energy storage is monitored in the cloud, the multi-source data uploaded to the cloud usually have inconsistent sampling frequencies. At the same time, the cloud data usually uses wireless channels, and communication interruptions may cause data loss. Therefore, the low sampling frequency and data loss issues are considered when establishing a multi-physics coupling model.
[0025] The multi-physics coupling model has three modules, namely, a time transformer module, a kernel regression module and a mask module;
[0026] The temporal Transformer module is used to extract temporal dependencies in time series;
[0027] The kernel regression module is used to extract the correlation between different physical dimension information;
[0028] First, missing values are marked. For each dimensional time series, assume that the time index t is the window part of , where the start and end times are and Then define the function as follows:
[0029]
[0030] Among them, h fg represents the output vector produced by the mask module, I(A) is an indicator set containing the indices of all non-missing data points, X j is the observation at index j, and |I(A)| is the number of elements in set I(A), used as a normalization factor;
[0031] In the time transformer module, first, for each segment of the original temperature data tpt, humidity data h, vibration data v, current data I, voltage data U and gas concentration data gc obtained by the computing device, the features are extracted respectively to obtain Among them, T {tpt,h,v,I,U,gc} Represents the time step in each dimension time series, d {tpt,h,v,I,U,gc} Indicates the feature quantity corresponding to each data point in each dimension after feature extraction;
[0032] In order to ensure that each source data can be exposed to the data in its adjacent time domain and thus obtain more global context information, the input sequence is passed through a one-dimensional temporal convolution layer:
[0033]
[0034] Among them, k {tpt,h,v,I,U,gc} is the convolution kernel size corresponding to the original data of each dimension, and d is the common dimension;
[0035] At the same time, in order to make the feature sequence carry time information, it is necessary to embed the position encoding. The position encoding formula is as follows:
[0036]
[0037] Among them, e t,r represents the position encoding at time step t and dimension r, and p is the dimension of the position encoding vector;
[0038] Then, the multi-head attention module extracts information, and for each Y j , define Q(·) and K(·) as Y j-1 and Y j+1 The function of Y is defined as V(·) j Function:
[0039]
[0040] Among them, W q ,W k ∈R 2p×2p , W v ∈R p×p , e j is the positional encoding of index j, A i It is the availability mask flag;
[0041] Then calculate V(Y j ) is the attention weighted sum of the vectors, and the formula is as follows:
[0042]
[0043] The above formula calculates V(Y j′ ) according to their attention weights for the output vector h j The weighted contribution of j It is an aggregate representation of the information in the input sequence that is most relevant to the current time point j. For index j′ with missing values, K(Y, j′, A) is zero.
[0044] Calculate n head Such instances, using index l = 1...n head To express this, by head The output vectors of the attention are concatenated into a vector To obtain the output of multi-head attention, the formula is as follows:
[0045]
[0046] Finally, h j With Y jMultiply point by point to get the features after attention processing:
[0047]
[0048] in, represents point-wise multiplication;
[0049] In the kernel regression module, we first define the siblings Sib(k,i) of index k along dimension i as the set of all indexes k′ such that k and k′ differ only in the i-th dimension:
[0050]
[0051] Then use the RBF kernel to calculate the index k in the i-th dimension i and k′ i The similarity score between them is as follows:
[0052]
[0053] Among them, γ is the parameter of the kernel function, which controls the width of the Gaussian distribution, that is, the degree of smoothness in the feature space; E[k i ] and E[k′ i ] represent input k i and k′ i Representation in some feature space; is the square of the Euclidean distance, that is, E[k i ] and E[k′ i ];
[0054] Given a sequence X at index (k, t), for each dimension i, the sum of the kernel-weighted metrics is calculated as follows:
[0055]
[0056] Among them, A k′,t = 1 for non-missing indexes, and 0 for missing indexes; when the dimension i is large, computation is saved by pre-selecting the top L members with similarity to the kernel;
[0057] Compute two other metrics: the sum and variance of the kernel weights of the X values over each sibling dimension:
[0058]
[0059] The last layer of the kernel regression module is the concatenation of the U, V, and W components:
[0060] h kr =Concat[U (k,i),t ,V (k,i),t ,W (k,i),t]
[0061] in,
[0062] Finally, the three vectors are concatenated as the final feature M, which is used for missing value interpolation tasks and thermal runaway warning tasks;
[0063] The missing value interpolation task decodes M into output vectors for all positions through deconvolution. The formula is as follows:
[0064]
[0065] in, is a learnable weight matrix, ReLU(·) is the activation function, Represents the output vector of all indices in the jth window, which needs to be passed Modulo operation obtains the output vector of each time step, where t is the time step and w is the window size;
[0066] The features are then input into a lightweight fully connected neural network to obtain the probability distribution of missing values:
[0067]
[0068] The loss function uses the mean absolute error, and the formula is as follows:
[0069]
[0070] in, is the interpolated predicted value, is the true value;
[0071] The thermal runaway warning task inputs the feature M into another fully connected neural network, which contains two hidden layers, each layer contains 64 neurons, and uses the ReLU activation function to introduce nonlinear mapping. The output layer uses the Sigmoid activation function and sets 1 neuron.
[0072] The loss function uses the binary cross entropy loss function, the formula is as follows:
[0073]
[0074] Where L is the loss function, N is the number of samples, and y i is the true label (0 or 1) of the i-th sample, is the probability that the i-th sample predicted by the model belongs to the positive class;
[0075] The final loss function is the weighted addition of the two subtask loss functions, and the formula is as follows:
[0076]
[0077] Among them, MAX(·) is to find L task1 and L task2 , and λ is a hyperparameter used to adjust the weight between the two tasks.
[0078] Furthermore, step 4 includes the following:
[0079] After the training is completed, the multi-source data to be tested is input into the model for testing. After the thermal state score Y is obtained, it is compared with the threshold score. If the thermal state score Y is less than the threshold, it is judged as normal. Otherwise, it is judged as abnormal, that is, thermal runaway.
[0080]
[0081] Furthermore, step 5 includes the following:
[0082] Data upload is divided into two types: alarm upload and trigger upload:
[0083] When uploading alarms: After the end-edge calculations in steps 1 to 4, a thermal status score is finally obtained. Once this value exceeds the set threshold, the end-edge will directly issue an alarm and upload the data to the cloud server. The data uploaded at this time is all data within the accident recall period required by the technical specifications of the energy storage power station monitoring system;
[0084] When no alarm occurs, the triggered upload mode is used: the edge computing performs mutation judgment on the data collected in each dimension. When the new data is compared with the most recently stored data, if its value is less than the set threshold θ, the data will not be recorded until the difference between the new data and the most recently stored data is greater than the threshold θ.
[0085] Due to the adoption of the above technical solution, the present invention has the following advantages and positive effects compared with the prior art:
[0086] The invention discloses a thermal runaway warning method for industrial and commercial energy storage based on multi-source collaboration, which is a method for monitoring thermal runaway by combining multiple sensor data. The method integrates key functional steps such as data acquisition, data cleaning, multi-physical parameter fusion, collaborative warning strategy formulation and data upload to realize efficient monitoring and warning of thermal runaway of energy storage power stations. In the data acquisition stage, special sensors are used to collect temperature, humidity, vibration, current, voltage and gas concentration signals respectively; in the data cleaning stage, integrity check, abnormal value processing, duplicate value processing, noise processing and other operations are performed; in the data fusion stage, a multi-physical coupling model is built, and the mutual influence and coupling relationship between various physical parameters are jointly considered by using time Transformer and kernel regression function, key information is extracted, and the intelligent level of the monitoring system is improved. At the same time, the multi-task learning method is used to effectively alleviate the problems of missing data in the industrial and commercial energy storage cloud and inconsistent data acquisition frequency in different dimensions, and the generalization of the model is enhanced; finally, in the data upload link, the system adopts a trigger upload method of uploading key data first, which ensures the real-time transmission of important information, while reducing the pressure of data storage and analysis of the upper system, and provides a strong guarantee for the safe and stable operation of the energy storage power station. Overall, the system can improve the monitoring efficiency of thermal runaway in energy storage power plants, while enhancing the accuracy of early warning, providing important support for the safety management of energy storage power plants. BRIEF DESCRIPTION OF THE DRAWINGS
[0087] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:
[0088] Figure 1 It is a flow chart of a method for early warning of thermal runaway of industrial and commercial energy storage based on multi-source collaboration of the present invention;
[0089] Figure 2 is a schematic diagram of the structure of the multi-physics coupling model in this embodiment;
[0090] Figure 3 Schematic diagram of the data trigger upload mode flow in this embodiment. DETAILED DESCRIPTION
[0091] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. 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.
[0092] like Figure 1-3 As shown, this embodiment discloses a thermal runaway warning method for industrial and commercial energy storage based on multi-source collaboration, comprising the following steps:
[0093] Step 1: Deploy temperature sensors, humidity sensors, vibration sensors, current sensors, voltage sensors, and gas concentration sensors in the battery compartment of the energy storage power station and collect data;
[0094] Specifically, step 1 includes the following:
[0095] Temperature sensors, humidity sensors, vibration sensors, current sensors, voltage sensors and gas concentration sensors are deployed in the battery compartment of the energy storage power station, and then the sensor data are collected through monitoring equipment; among them, the sampling frequency of current data, voltage data and vibration data is 1Hz, and the sampling frequency of temperature data, humidity data and gas concentration data is 0.2Hz.
[0096] Step 2: Check the integrity, handle outliers, duplicate values, and noise of the extracted data;
[0097] Specifically, step 2 includes the following:
[0098] In order to ensure the accuracy and effectiveness of subsequent analysis, the collected raw data needs to be cleaned, including key processes such as integrity check, outlier processing, duplicate value processing, and noise processing.
[0099] First, we go through each data point in each dimension and check the completeness of the data to ensure that all required fields are filled in.
[0100] Then use the Z-score statistic to calculate the degree of deviation of each data point from the mean, set a threshold to identify outliers, and consider deleting or replacing them with the mean for outliers with a degree of deviation higher than the threshold. The formula for the Z-score statistic is as follows:
[0101]
[0102] Among them, X is the observed value of the data, μ is the mean value of the group of data, and σ is the standard deviation of the group of data;
[0103] In addition, duplicate data need to be identified and handled to avoid introducing bias in subsequent analyses.
[0104] Next, denoising is an indispensable step in data cleaning. During the data collection process, some noise may be mixed into the data due to equipment failure, transmission errors, etc. These noises may interfere with the accurate interpretation of the data. This embodiment uses the median filtering method for smoothing and denoising, and the formula is as follows:
[0105] y[n]=Median(x[nk],...,x[n],...,x[n+k])
[0106] This method replaces the value of each data point with the median value of its surrounding data points, where k is the size of the filter.
[0107] Step 3: Consider the mutual influence and coupling relationship between various physical parameters and establish a multi-physics coupling model;
[0108] Furthermore, step 3 includes the following:
[0109] In order to describe the state of the system more comprehensively and accurately, it is necessary to jointly consider the mutual influence and coupling relationship between various physical parameters and establish a multi-physics coupling model. Since most industrial and commercial energy storage is monitored on the cloud, the multi-source data on the cloud usually have inconsistent sampling frequencies; at the same time, the cloud data usually uses wireless channels, and there may be communication interruptions and other situations that lead to data loss. Therefore, when establishing a multi-physics coupling model, the problems of low sampling frequency and data loss must be considered.
[0110] like Figure 2 As shown, the multi-physics coupling model has three modules, namely, a time Transformer module, a kernel regression module and a mask module;
[0111] The temporal Transformer module is used to extract temporal dependencies in time series;
[0112] The kernel regression module is used to extract the correlation between different physical dimension information;
[0113] First, missing values are marked. For each dimensional time series, assume that the time index t is the window part of , where the start and end times are and Then define the function as follows:
[0114]
[0115] Among them, h fg represents the output vector produced by the mask module, I(A) is an indicator set containing the indices of all non-missing data points, X j is the observation at index j, and |I(A)| is the number of elements in set I(A), used as a normalization factor;
[0116] In the time transformer module, first, for each segment of the original temperature data tpt, humidity data h, vibration data v, current data I, voltage data U and gas concentration data gc obtained by the computing device, the features are extracted respectively to obtain Among them, T {tpt,h,v,I,U,gc} Represents the time step in each dimension time series, d {tpt,h,v,I,U,gc} Indicates the feature quantity corresponding to each data point in each dimension after feature extraction;
[0117] In order to ensure that each source data can be exposed to the data in its adjacent time domain and thus obtain more global context information, the input sequence is passed through a one-dimensional temporal convolution layer:
[0118]
[0119] Among them, k {tpt,h,v,I,U,gc} is the convolution kernel size corresponding to the original data of each dimension, and d is the common dimension;
[0120] At the same time, in order to make the feature sequence carry time information, it is necessary to embed the position encoding. The position encoding formula is as follows:
[0121]
[0122] Among them, e t,r represents the position encoding at time step t and dimension r, and p is the dimension of the position encoding vector;
[0123] Then, the multi-head attention module extracts information, and for each Y j , define Q(·) and K(·) as Y j-1 and Y j+1 The function of Y is defined as V(·) j Function:
[0124] Q(Y,j)=([Y j-1 , Y j+1 ]+e j )W q +b q
[0125]
[0126] V(Y j )=Y j W v +b v
[0127] Among them, W q ,W k ∈R 2p×2p , Wv ∈R p×p , e j is the positional encoding of index j, A i It is the availability mask flag;
[0128] Then calculate V(Y j ) is the attention weighted sum of the vectors, and the formula is as follows:
[0129]
[0130] The above formula calculates V(Y j′ ) according to their attention weights for the output vector h j The weighted contribution of j It can be viewed as an aggregate representation of the most relevant information in the input sequence for the current time point j. It should be noted that for index j′ with missing values, K(Y, j′, A) is zero; therefore, these indexes are not considered in the attention mechanism.
[0131] Calculate n head Such instances, using index l = 1...n head To express this, by head The output vectors of the attention are concatenated into a vector To obtain the output of multi-head attention, the formula is as follows:
[0132]
[0133] Finally, h j With Y j Multiply point by point to get the features after attention processing:
[0134]
[0135] in, represents point-wise multiplication;
[0136] In the kernel regression module, we first define the siblings Sib(k,i) of index k along dimension i as the set of all indexes k′ such that k and k′ differ only in the i-th dimension:
[0137]
[0138] Then use the RBF kernel to calculate the index k in the i-th dimension i and k′ i The similarity score between them is as follows:
[0139]
[0140] Among them, γ is the parameter of the kernel function, which controls the width of the Gaussian distribution, that is, the degree of smoothness in the feature space; E[k i ] and E[k′ i ] represent input k i and k′ i Representation in some feature space; is the square of the Euclidean distance, that is, E[k i ] and E[k′ i ];
[0141] Given a sequence X at index (k, t), for each dimension i, the sum of the kernel-weighted metrics is calculated as follows:
[0142]
[0143] Among them, A k′,t = 1 for non-missing indexes, and 0 for missing indexes; when the dimension i is large, computation is saved by pre-selecting the top L members with similarity to the kernel;
[0144] Compute two other metrics: the sum and variance of the kernel weights of the X values over each sibling dimension:
[0145]
[0146] The last layer of the kernel regression module is the concatenation of the U, V, and W components:
[0147] h kr =Concat[U (k,i),t ,V (k,i),t ,W (k,i),t ]
[0148] in,
[0149] Finally, the three vectors are concatenated as the final feature M, which is used for missing value interpolation tasks and thermal runaway warning tasks;
[0150] The missing value interpolation task decodes M into output vectors for all positions through deconvolution. The formula is as follows:
[0151]
[0152] in, is a learnable weight matrix, ReLU(·) is the activation function, Represents the output vector of all indices in the jth window, which needs to be passed Modulo operation obtains the output vector of each time step, where t is the time step and w is the window size;
[0153] The features are then input into a lightweight fully connected neural network to obtain the probability distribution of missing values:
[0154]
[0155] The loss function uses the mean absolute error, and the formula is as follows:
[0156]
[0157] in, is the interpolated predicted value, is the true value;
[0158] The thermal runaway warning task inputs the feature M into another fully connected neural network, which contains two hidden layers, each layer contains 64 neurons, and uses the ReLU activation function to introduce nonlinear mapping. The output layer uses the Sigmoid activation function and sets 1 neuron.
[0159] The loss function uses the binary cross entropy loss function, the formula is as follows:
[0160]
[0161] Where L is the loss function, N is the number of samples, and y i is the true label (0 or 1) of the i-th sample, is the probability that the i-th sample predicted by the model belongs to the positive class;
[0162] The final loss function is the weighted addition of the two subtask loss functions, and the formula is as follows:
[0163]
[0164] Among them, MAX(·) is to find L task1 and L task2 The maximum value of , λ is a hyperparameter used to adjust the weight between the two tasks. Based on experience, it is set to 0.9 in this embodiment.
[0165] Step 4: Compare the thermal status score with the threshold to determine whether an alarm is required;
[0166] Specifically, step 4 includes the following contents:
[0167] After the training is completed, the multi-source data to be tested is input into the model for testing. After the thermal state score Y is obtained, it is compared with the threshold score. If the thermal state score Y is less than the threshold, it is judged as normal. Otherwise, it is judged as abnormal, that is, thermal runaway.
[0168]
[0169] According to empirical values, the score is set to 0.95 in this embodiment.
[0170] Step 5: Upload system data.
[0171] Specifically, step 5 includes the following contents:
[0172] This system adopts the "cloud-edge" architecture, and data upload is divided into two situations: alarm upload and trigger upload:
[0173] When uploading alarms: After the end-edge calculations in steps 1 to 4, a thermal status score is finally obtained. Once this value exceeds the set threshold, the end-edge will directly issue an alarm and upload the data to the cloud server. The data uploaded at this time is all data within the accident recall period required by the technical specifications of the energy storage power station monitoring system;
[0174] When no alarm occurs, the triggered upload mode is used: the edge computing performs mutation judgment on the data collected in each dimension. When the new data is compared with the most recently stored data, if its value is less than the set threshold θ, the data will not be recorded until the difference between the new data and the most recently stored data is greater than the threshold θ. However, due to the existence of measurement errors, the amount of data may not decrease. Therefore, based on the above method, data recording is considered after the data changes p times. The overall flow chart is as follows Figure 3 .
[0175] This approach can greatly reduce the delay and redundancy in the data transmission process by optimizing the data transmission and processing process, ensuring that key data can still be transmitted in real time and accurately under rapidly changing circumstances. This not only improves the efficiency of data transmission, but also greatly improves the reliability of data communication. On this basis, it can also effectively reduce the pressure on upper-level system data storage, making more reasonable use of storage space, while reducing the burden of data analysis and processing.
[0176] This embodiment is a thermal runaway warning method for industrial and commercial energy storage based on multi-source collaboration. The method first constructs a multi-physics coupling model. Through the time Transformer and kernel regression function, the mutual influence and coupling relationship between the physical parameters in the energy storage system are comprehensively considered, and the complex dynamics inside the system are accurately captured. Since most industrial and commercial energy storage is monitored on the cloud, the multi-source data on the cloud usually have inconsistent sampling frequencies; at the same time, the cloud data usually uses wireless channels, and there may be communication interruptions and other situations that lead to data loss. Therefore, when establishing a multi-physics coupling model, the problems of low sampling frequency and data loss must be considered. This embodiment adopts a multi-task learning method to effectively alleviate the problems of inconsistent data collection frequencies of different dimensions of industrial and commercial energy storage and missing cloud data, and significantly improves the generalization ability and robustness of the model. In the data upload link, the system innovatively adopts a trigger mechanism for uploading key data first, ensuring the real-time transmission of important information, while reducing the pressure on the upper system in data storage and analysis, and providing a strong guarantee for the safe and stable operation of energy storage power stations. In general, the present invention not only improves the monitoring efficiency of thermal runaway of energy storage power stations, but also enhances the accuracy of early warning, providing important support for the safety management of energy storage power stations.
[0177] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.
Claims
1. A thermal runaway early warning method for industrial and commercial energy storage based on multi-source collaboration, characterized in that: The following steps are involved: Step 1: Deploy temperature sensors, humidity sensors, vibration sensors, current sensors, voltage sensors, and gas concentration sensors in the battery compartment of the energy storage power station and collect data; Step 2: Check the integrity, handle outliers, duplicate values, and noise of the extracted data; Step 3: Consider the mutual influence and coupling relationship between various physical parameters and establish a multi-physics coupling model; Step 4: Compare the thermal status score with the threshold to determine whether an alarm is required; Step 5: Upload system data.
2. The method for early warning of thermal runaway of industrial and commercial energy storage based on multi-source collaboration according to claim 1 is characterized in that: Step 1 includes the following: Temperature sensors, humidity sensors, vibration sensors, current sensors, voltage sensors and gas concentration sensors are deployed in the battery compartment of the energy storage power station, and then the sensor data are collected through monitoring equipment; among them, the sampling frequency of current data, voltage data and vibration data is 1Hz, and the sampling frequency of temperature data, humidity data and gas concentration data is 0.2Hz.
3. The method for early warning of thermal runaway of industrial and commercial energy storage based on multi-source collaboration according to claim 1 is characterized in that: Step 2 includes the following: First, we go through each data point in each dimension and check the completeness of the data to ensure that all required fields are filled in. Then, the Z-score statistical method is used to calculate the degree of deviation of each data point from the mean, and a threshold is set to identify outliers. For outliers with a degree of deviation higher than the threshold, they are deleted or replaced with the mean. The formula of the Z-score statistical method is as follows: Among them, X is the observed value of the data, μ is the mean value of the group of data, and σ is the standard deviation of the group of data; During the data collection process, the median filtering method is used for smoothing and denoising. The formula is as follows: y[n]=Median(x[nk],...,x[n],...,x[n+k]) The value of each data point is replaced by the median value of its surrounding data points, where k is the size of the filter.
4. The method for early warning of thermal runaway of industrial and commercial energy storage based on multi-source collaboration according to claim 1 is characterized in that: Step 3 Includes the following: Since most industrial and commercial energy storage is monitored in the cloud, the multi-source data uploaded to the cloud usually have inconsistent sampling frequencies. At the same time, the cloud data usually uses wireless channels, and communication interruptions may cause data loss. Therefore, the low sampling frequency and data loss issues are considered when establishing a multi-physics coupling model. The multi-physics coupling model has three modules, namely, a time transformer module, a kernel regression module and a mask module; The temporal Transformer module is used to extract temporal dependencies in time series; The kernel regression module is used to extract the correlation between different physical dimension information; First, missing values are marked. For each dimensional time series, assume that the time index t is the window part of , where the start and end times are and Then define the function as follows: Among them, h fg represents the output vector produced by the mask module, I(A) is an indicator set containing the indices of all non-missing data points, X j is the observation at index j, and |I(A)| is the number of elements in set I(A), used as a normalization factor; In the time transformer module, first, for each segment of the original temperature data tpt, humidity data h, vibration data v, current data I, voltage data U and gas concentration data gc obtained by the computing device, the features are extracted respectively to obtain Among them, T {tpt,h,v,I,U,gc} Represents the time step in each dimension time series, d {tpt,h,v,I,U,gc} Indicates the feature quantity corresponding to each data point in each dimension after feature extraction; In order to ensure that each source data can be exposed to the data in its adjacent time domain and thus obtain more global context information, the input sequence is passed through a one-dimensional temporal convolution layer: Among them, k {tpt,h,v,I,U,gc} is the convolution kernel size corresponding to the original data of each dimension, and d is the common dimension; At the same time, in order to make the feature sequence carry time information, it is necessary to embed the position encoding. The position encoding formula is as follows: Among them, e t,r represents the position encoding at time step t and dimension r, and p is the dimension of the position encoding vector; Then, the multi-head attention module extracts information, and for each Y j , define Q(·) and K(·) as Y j-1 and Y j+1 The function of Y is defined as V(·) j Function: Q(Y,j)=([Y j-1 ,Y j+1 ]+e j )W q +b q V(Y j )=Y j W v +b v Among them, W q ,W k ∈R 2p×2p , W v ∈R p×p , e j is the positional encoding of index j, A i It is the availability mask flag; Then calculate V(Y j ) is the attention weighted sum of the vectors, and the formula is as follows: The above formula calculates V(Y) at all time points j' in the context of time point j. j' ) according to their attention weights to the output vector h j The weighted contribution of j It is an aggregate representation of the information in the input sequence that is most relevant to the current time point j. For index j' with missing values, K(Y, j', A) is zero. Calculate n head Such instances, using index l = 1...n head To express this, by head The output vectors of the attention are concatenated into a vector To obtain the output of multi-head attention, the formula is as follows: Finally, h j With Y j Multiply point by point to get the features after attention processing: in, represents point-wise multiplication; In the kernel regression module, we first define the siblings Sib(k,i) of index k along dimension i as the set of all indexes k', such that k and k' differ only in the i-th dimension: Then use the RBF kernel to calculate the index k in the i-th dimension i and k' i The similarity score between them is as follows: Among them, γ is the parameter of the kernel function, which controls the width of the Gaussian distribution, that is, the degree of smoothness in the feature space; E[k i ] and E[k' i ] represent input k i and k' i Representation in some feature space; is the square of the Euclidean distance, that is, E[k i ] and E[k' i ]; Given a sequence X at index (k, t), for each dimension i, the sum of the kernel-weighted metrics is calculated as follows: Among them, A k′,t = 1 for non-missing indexes, and 0 for missing indexes; when the dimension i is large, computation is saved by pre-selecting the top L members with similarity to the kernel; Compute two other metrics: the sum and variance of the kernel weights of the X values over each sibling dimension: V (k,i),t =There is(X Sib(k,t),t ) The last layer of the kernel regression module is the concatenation of the U, V, and W components: h kr =Concat[U (k,i),t ,V (k,i),t ,W (k,i),t ] in, Finally, the three vectors are concatenated as the final feature M, which is used for missing value interpolation tasks and thermal runaway warning tasks; The missing value interpolation task decodes M into output vectors for all positions through deconvolution. The formula is as follows: in, is a learnable weight matrix, ReLU(·) is the activation function, Represents the output vector of all indices in the jth window, which needs to be passed Modulo operation obtains the output vector of each time step, where t is the time step and w is the window size; The features are then input into a lightweight fully connected neural network to obtain the probability distribution of missing values: The loss function uses the mean absolute error, and the formula is as follows: in, is the interpolated predicted value, is the true value; The thermal runaway warning task inputs the feature M into another fully connected neural network, which contains two hidden layers, each layer contains 64 neurons, and uses the ReLU activation function to introduce nonlinear mapping. The output layer uses the Sigmoid activation function and sets 1 neuron. The loss function uses the binary cross entropy loss function, the formula is as follows: Where L is the loss function, N is the number of samples, and y i is the true label (0 or 1) of the i-th sample, is the probability that the i-th sample predicted by the model belongs to the positive class; The final loss function is the weighted addition of the two subtask loss functions, and the formula is as follows: Among them, MAX(·) is to find L task1 and L task2 , and λ is a hyperparameter used to adjust the weight between the two tasks.
5. The method for early warning of thermal runaway of industrial and commercial energy storage based on multi-source collaboration according to claim 1 is characterized in that: Step 4 includes the following: After the training is completed, the multi-source data to be tested is input into the model for testing. After the thermal state score Y is obtained, it is compared with the threshold score. If the thermal state score Y is less than the threshold, it is judged as normal. Otherwise, it is judged as abnormal, that is, thermal runaway.
6. The method for early warning of thermal runaway of industrial and commercial energy storage based on multi-source collaboration according to claim 1 is characterized in that: Step 5 includes the following: Data upload is divided into two types: alarm upload and trigger upload: When uploading alarms: After the end-edge calculations in steps 1 to 4, a thermal status score is finally obtained. Once this value exceeds the set threshold, the end-edge will directly issue an alarm and upload the data to the cloud server; The data uploaded at this time is all the data within the accident recall period required by the technical specifications of the energy storage power station monitoring system; When no alarm occurs, the triggered upload mode is used: the edge computing performs mutation judgment on the data collected in each dimension. When the new data is compared with the most recently stored data, if its value is less than the set threshold θ, the data will not be recorded until the difference between the new data and the most recently stored data is greater than the threshold θ.
Citation Information
Cited By
Industrial Internet of Things equipment operation state time series data missing value filling method
CN120277344A
Method and device for monitoring thermal runaway of energy storage system
CN120595129A
Method and device for processing thermal runaway event of autonomous vehicle
CN121492664A
A method and apparatus for handling thermal runaway events in autonomous vehicles.
CN121492664B
