Method and device for monitoring thermal runaway of energy storage system

Abstract

The invention provides a thermal runaway monitoring method and device for an energy storage system, and the method comprises the following steps: obtaining a plurality of times, obtaining the temperature and voltage of any cell at different times, and the current, gas concentration fraction, acoustic signal and structural strain of the energy storage system at different times, and generating corresponding vectors; and obtaining a trained neural network, inputting the vector into the neural network, obtaining a probability P of thermal runaway output by the neural network, and controlling the energy storage system based on the probability P. The monitoring method can be used for monitoring the thermal runaway of the energy storage system.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving, and in particular to a method and device for monitoring thermal runaway of an energy storage system. Background Art

[0002] Clean energy sources, such as wind power and photovoltaics, are intermittent and volatile, requiring energy storage systems to achieve peak load regulation and smooth output. Traditional energy storage methods (such as pumped hydro and compressed air storage) are limited by terrain and cost constraints, making lithium battery energy storage power stations the optimal choice due to their rapid operation, high energy density, and long lifespan. However, inadequate protection facilities in lithium battery energy storage power stations lead to frequent safety incidents, including fires and explosions, which seriously endanger life and property.

[0003] Therefore, how to monitor the thermal runaway of energy storage systems has become an urgent problem to be solved. Summary of the Invention

[0004] The object of the present invention is to provide a method and device for monitoring thermal runaway of an energy storage system.

[0005] In order to achieve one of the above-mentioned objects, an embodiment of the present invention provides a method for monitoring thermal runaway of an energy storage system, wherein the energy storage system comprises M battery cells Battery1, Battery2, ..., Battery M , where M is a natural number; comprising the following steps: obtaining multiple times Time1, Time2, ..., Time N Time i Earlier than Time i+1 , where i and N are both natural numbers, 1≤i≤N-1; get any battery k At Time t Temperature T t,k , voltage V t,k , the energy storage system at time Time t The current I t , gas concentration fraction G t , acoustic signal S t and structural strain ε t , generating vector X t =[T t,1 ,T t,2 ,...,T t,M ,V t,1 ,V t,2 ,...,V t,M ,I t ,G t ,St ,ε t ], where t and k are both natural numbers, 1≤t≤N, 1≤k≤M; obtain the trained neural network, vector X1, vector X2, ..., and vector X N The corresponding codes are input into the neural network to obtain the probability P of thermal runaway output by the neural network. The probability P is used to represent the probability of thermal runaway of the energy storage system in a preset time period from the current time; the energy storage system is controlled based on the probability P.

[0006] As a further improvement of an embodiment of the present invention, the "generated vector X t =[T t,1 ,T t,2 ,...,T t,M ,V t,1 ,V t,2 ,...,V t,M ,I t ,G t ,S t ,ε t ]” specifically includes: for all temperatures T t,j Perform time alignment, missing value filling, outlier suppression, and Z-score standardization; t,j Perform time alignment, missing value filling, outlier suppression, and Z-score standardization. t , all gas concentration fraction G t , all acoustic signals S t and all the structural strains ε t They are all processed by time alignment, missing value filling, outlier suppression and Z-score standardization respectively.

[0007] As a further improvement of an embodiment of the present invention, the neural network includes a graph convolutional neural network, in which X1, X2, ..., X N Form a matrix H with N rows and 2*M+4 columns (1) ; in, is the feature of the L-th layer node, is the trainable weight, σ(·) is ReLU, d L is the feature dimension, where L is a natural number, 1≤L≤M, D is the degree matrix, I is the self-loop, A is a matrix, the matrix A contains M rows and M columns of elements, when any battery m and any battery n When adjacent and sharing a heat dissipation path, element Amn = 1, otherwise, element a mn = 0; where m and n are both natural numbers, 1 ≤ m, n ≤ M; the output of the graph convolutional neural network is H k = H (k) , 1 ≤ k ≤ M.

[0008] As a further improvement of an embodiment of the present invention, the neural network includes a Transformer module, and the Transformer module expands H t-W+1 , …, H t output by the graph convolutional neural network in the time dimension, and then obtains Q, K, and V through linear projection, and then feeds them into the multi-head self-attention where Q is the query, K is the key, and V is the value; where W is a natural number, W < t; the neural network is also used for: calculating the anomaly score Calculating the lead prediction where w r is the weight vector corresponding to the anomaly branch, w p is the weight vector corresponding to the lead branch, b r and b p are both bias terms; the probability of thermal runaway output by the neural network

[0009] As a further improvement of an embodiment of the present invention, "controlling the energy storage system based on the probability P" specifically includes: when R t > the preset threshold, obtain the set U, and the set U contains several actions; C risk (u) = p fail (u)P + (1 - P)C loss , p fail (u) = exp(-kΔT u ), C total (u) = C act (u) + λC risk (u), u ∈ U; where C risk (u) represents the expected loss that still evolves into thermal runaway after taking the action u; C loss represents the fixed loss caused by a single thermal runaway accident; 0 < p fail (u) ≤ 1 represents the probability that still evolves into thermal runaway after the execution of the action u; ΔT u is the expected temperature drop after taking the action u; C act (u) represents the cost required to take the action u, 0 < λ < 1; obtain C total (u) corresponding to any action u in the set U, and obtain the minimum value of C total(u′) corresponds to the action u′, and then the energy storage system is controlled using the action u′.

[0010] The embodiment of the present invention further provides a monitoring device for thermal runaway of an energy storage system, wherein the energy storage system comprises M battery cells Battery1, Battery2, ..., Battery M , where M is a natural number; including the following modules: an information acquisition module for acquiring multiple times Time1, Time2, ..., Time N Time i Earlier than Time i+1 , where i and N are both natural numbers, 1≤i≤N-1; the pre-processing module is used to obtain the battery of any cell k At Time t Temperature T t,k , voltage V t,k , the energy storage system at time Time t The current I t , gas concentration fraction G t , acoustic signal S t and structural strain ε t , generating vector X t =[T t,1 ,T t,2 ,...,T t,M ,V t,1 ,V t,2 ,...,V t,M ,I t ,G t ,S t ,ε t ], where t and k are both natural numbers, 1≤t≤N, 1≤k≤M; an information processing module is used to obtain the trained neural network, vector X1, vector X2, ..., and vector X N The corresponding codes are input into the neural network to obtain the probability P of thermal runaway output by the neural network. The probability P is used to represent the probability of thermal runaway of the energy storage system in a preset time period from the current time; a control module is used to control the energy storage system based on the probability P.

[0011] As a further improvement of an embodiment of the present invention, the pre-processing module is further used to: t,j Perform time alignment, missing value filling, outlier suppression, and Z-score standardization; t,j Perform time alignment, missing value filling, outlier suppression, and Z-score standardization. t, all gas concentration fractions G t , all acoustic signals S t and all structural strains ε t are respectively subjected to time alignment processing, missing value filling processing, outlier suppression processing, and Z-score normalization processing.

[0012] As a further improvement of an embodiment of the present invention, the neural network includes a graph convolutional neural network. In the graph convolutional neural network, X1, X2,...., X N are combined to form a matrix H with N rows and 2*M + 4 columns (1) ; where is the feature of the node in the L-th layer, is the trainable weight, σ(·) is ReLU, d L is the feature dimension, where L is a natural number, 1 ≤ L ≤ M, D is the degree matrix, I is the self-loop, A is the matrix, and the matrix A contains elements with M rows and M columns. When any battery cell Battery m and any battery cell Battery n are adjacent and share a heat dissipation path, the element A mn = 1, otherwise, the element a mn = 0; where m and n are both natural numbers, 1 ≤ m, n ≤ M; the output of the graph convolutional neural network is H k = H (k) , 1 ≤ k ≤ M.

[0013] As a further improvement of an embodiment of the present invention, the neural network includes a Transformer module. The Transformer module expands the H t-W+1 ,…,H t output by the graph convolutional neural network in the time dimension, and then obtains Q, K, and V through linear projection, and then feeds them into the multi-head self-attention where Q is the query, K is the key, and V is the value; where W is a natural number, W < t; the neural network is also used for: calculating the anomaly score calculating the lead prediction where w r is the weight vector corresponding to the anomaly branch, w p is the weight vector corresponding to the lead branch, b r and b p are both bias terms; the probability of thermal runaway output by the neural network

[0014] As a further improvement of an embodiment of the present invention, the control module is further used for: when R t>When a threshold is preset, a set U is obtained, wherein the set U includes several actions; C risk (u) = p fail (u)P+(1-P)C loss , p fail (u) = exp(-kΔT u ), C total (u)=C act (u)+λC risk (u), u∈U; where C risk (u) represents the expected loss if the system still evolves into thermal runaway after taking action u; C loss Represents the fixed loss caused by a thermal runaway accident; 0 <p fail (u)≤1 indicates the probability of evolving into thermal runaway after action u is completed; ΔT u is the expected temperature drop after taking action u; C act (u) represents the cost of taking action u, 0<λ<1; get the C corresponding to any action u in the set U total (u), get the minimum value C total (u′) corresponds to the action u′, and then the energy storage system is controlled by the action u′

[0015] Compared to the prior art, the technical effect of the present invention is that the embodiments of the present invention provide a method and apparatus for monitoring thermal runaway of an energy storage system. The monitoring method includes the following steps: obtaining multiple times, obtaining the temperature and voltage of any battery cell at different times, the current, gas concentration fraction, acoustic signal, and structural strain of the energy storage system at different times, and generating corresponding vectors; obtaining a trained neural network, inputting the vectors into the neural network, obtaining the probability P of thermal runaway output by the neural network, and controlling the energy storage system based on the probability P. This monitoring method can monitor thermal runaway of an energy storage system. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flowchart of the monitoring method in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The present invention will be described in detail below with reference to the various embodiments shown in the accompanying drawings. However, these embodiments do not limit the present invention, and any structural, methodological, or functional modifications made by those skilled in the art based on these embodiments are all within the scope of protection of the present invention.

[0018] As used herein, terms indicating spatial relative positions such as "upper," "above," "lower," and "below" are used for ease of explanation to describe the relationship of one unit or feature relative to another unit or feature as shown in the accompanying drawings. Terms of spatial relative position may be intended to encompass different orientations of the device in use or operation in addition to the orientation shown in the drawings. For example, if the device in the drawings were turned over, units described as being "below" or "beneath" other units or features would be "above" the other units or features. Thus, the exemplary term "below" may encompass both above and below orientations. The device may be oriented in other ways (rotated 90 degrees or in other orientations), and the spatially relative descriptors used herein interpreted accordingly.

[0019] The first embodiment of the present invention provides a method for monitoring thermal runaway of an energy storage system, wherein the energy storage system comprises M battery cells Battery1, Battery2, ..., Battery M , where M is a natural number; Figure 1 As shown, the following steps are included:

[0020] Step 101: Get multiple times Time1, Time2, ..., Time N Time i Earlier than Time i+1 , where i and N are both natural numbers, 1≤i≤N-1; optional, Time i and Time i+1 The time interval between them is 1S.

[0021] Step 102: Get any battery k At Time t Temperature T t,k , voltage V t,k , the energy storage system at time Time t The current I t , gas concentration fraction G t , acoustic signal S t and structural strain ε t , generating vector X t =[T t,1 ,T t,2 ,...,T t,M ,V t,1 ,V t,2 ,...,V t,M ,I t ,G t ,S t ,ε t ], where t and k are both natural numbers, 1≤t≤N, 1≤k≤M;

[0022] Step 103: Obtain the trained neural network, vector X1, vector X2, ..., and vector X N The corresponding codes are input into the neural network to obtain the probability P of thermal runaway output by the neural network, where the probability P is used to represent the probability of thermal runaway of the energy storage system occurring in a preset time period from the current time.

[0023] Step 104: Control the energy storage system based on the probability P.

[0024] The detection method in the embodiment of the present invention realizes the early warning + immediate suppression + rapid decision-making closed loop of thermal runaway of the energy storage system through multimodal sensor fusion, data-driven prediction and active suppression, providing an innovative solution for large-scale energy storage safety.

[0025] In this embodiment, the "generated vector X t =[T t,1 ,T t,2 ,...,T t,M ,V t,1 ,V t,2 ,...,V t,M ,I t ,G t ,S t ,ε t ]” specifically include:

[0026] For all temperatures T t,j Perform time alignment, missing value filling, outlier suppression, and Z-score standardization;

[0027] Voltage V t,j Perform time alignment, missing value filling, outlier suppression, and Z-score standardization;

[0028] For all currents I t , all gas concentration fraction G t , all acoustic signals S t and all the structural strains ε t They are all processed by time alignment, missing value filling, outlier suppression and Z-score standardization respectively.

[0029] Here, for current I1, current I2, ..., current I N Perform time alignment, missing value filling, outlier suppression, and Z-score standardization; perform gas concentration scores G1, G2, ..., and G NPerform time alignment, missing value filling, outlier suppression, and Z-score normalization processing; perform time alignment, missing value filling, outlier suppression, and Z-score normalization on acoustic signals S1, S2, ..., and S N Perform time alignment, missing value filling, outlier suppression, and Z-score standardization; perform structural strain ε1, structural strain ε2, ..., structural strain ε N Perform time alignment, missing value filling, outlier suppression, and Z-score standardization.

[0030] Temperature T t , voltage V t , current I t , gas concentration fraction G t , acoustic signal S t and structural strain ε t All data is acquired by sensors, with the sampling time being the sensor's detection time. Time alignment involves synchronizing the data at the hour based on this sampling time. Sub-millisecond jitter is compensated using linear interpolation to ensure one-to-one correspondence between modalities at the same timestamp. Missing value filling uses a "forward hold" strategy for occasional packet drops (lasting <3 seconds). Continuous missing values ​​of 3 seconds or more are marked as unusable windows and discarded during training. Outlier suppression uses a sliding window mean ±3σ rule for each channel to remove transient spikes; removed data are then smoothed using double-ended interpolation.

[0031] The Z-score standardization process is called a standard score. It uses the standard deviation as a ruler to measure the distance of a raw score from the mean. The number of standard deviations in this distance is the z-score. This determines the position of this data in the overall data. This process is called standardization. The conversion formula is: Among them, Y is the original data, is the mean, and s is the standard deviation. It is understandable that after these processes, each mode satisfies zero mean and unit variance, which prevents temperature (hundreds of orders of magnitude) from "overwhelming" acoustics (double digits) and facilitates gradient stability and model convergence.

[0032] In this embodiment, the neural network includes a graph convolutional neural network, in which X1, X2, ..., X N Form a matrix H with N rows and 2*M+4 columns (1) ; in, is the feature of the L-th layer node, is the trainable weight, σ(·) is ReLU, d L is the feature dimension, where L is a natural number, 1≤L≤M, D is the degree matrix, I is the self-loop, A is a matrix, the matrix A contains M rows and M columns of elements, when any battery m and any battery n When adjacent and sharing a heat dissipation path, element A mn =1, otherwise, element a mn =0; where m and n are both natural numbers, 1≤m, n≤M; the output of the graph convolutional neural network is H k =H (k) , 1≤k≤M.

[0033] Here, the degree matrix is ​​a diagonal matrix, and the elements on the diagonal are the degrees of each vertex. i The degree of a vertex represents the number of variables associated with that vertex. k The degree d(v k )=N(k) (i.e. the number of edges connected to the vertex). In a directed graph, vertex v k The degree of vertex v k The out-degree and in-degree of vertex v i The number of directed edges outgoing and incoming to vertex v k The number of directed edges.

[0034] In an embodiment of the present invention, the method innovatively encodes the basic physical law of "conduction from adjacent battery cells" directly into the adjacency matrix structure of the graph neural network. Through this design, the model can implicitly capture the core mechanism of heat diffusion within the battery pack without relying on tedious manual feature engineering. Specifically, the connection weights in the adjacency matrix can reflect the thermal conductivity coefficient between battery cells, so that the graph convolution operation naturally simulates the heat transfer process between adjacent battery cells. This inductive bias based on physical principles not only improves the model's learning efficiency for local cluster high-temperature propagation patterns, but also better generalizes to thermal runaway prediction tasks under different battery pack configurations. Compared with traditional methods that require manual design of heat diffusion features, this end-to-end learning framework can not only maintain physical rationality, but also automatically discover complex temperature propagation laws in the data.

[0035] Optionally, stack two layers (L=0,1) to produce H t =H (2) ∈R N×d .

[0036] In this embodiment, the neural network includes a Transformer module, which converts the H output of the graph convolutional neural network into t-W+1 ,…,H t Expand by time dimension, then linearly project to get Q, K, V, and then feed into multi-head self-attention Where Q is the query, K is the key, and V is the value; where W is a natural number, W <t;

[0037] The neural network is also used to calculate anomaly scores Calculating lead time forecasts Among them, w r is the weight vector corresponding to the abnormal branch, w p is the weight vector corresponding to the advance branch, b r and b p are all bias terms; the probability of thermal runaway output by the neural network

[0038] Here, σ(·)——LogisticSigmoid, σ(z)=1 / 1+e -z , compressing real numbers to the interval (0,1).

[0039] here, d h For single head dimension, optional, d h = 6464, number of heads = 8; softmax row normalization makes the sum of coefficients equal to 1, reflecting the "attention". The multi-head results are concatenated and fed-forward network is output as Z t ∈R W×d , and then perform global mean pooling to obtain the final spatiotemporal feature vector h t ∈R d , for the fusion prediction layer to calculate R t and

[0040] For example, the GNN part = importing “who is next to whom” into the model to identify the spatial coupling of heat spreading horizontally within the PACK; the Transformer part = importing “temperature rise in the past 60 seconds → out of control” into the model to capture the temporal precursor; and finally h t It also contains two-dimensional information: "which battery cell is heating up and how long it is heating up", which is particularly critical for predicting thermal runaway.

[0041] H t ∈R d is the spatiotemporal feature vector, dimension d = 256; here, σ(·) is the Logistic Sigmoid, which can compress real numbers into the (0,1) interval.

[0042] Here, R t Reflects the deviation between the current working condition and the historical safety baseline; the closer the value is to 1, the more likely it is in an abnormal state. The system uses the threshold R t >0.6 triggers protection action.

[0043] The probability of thermal runaway occurring within the next 30 minutes is given, which is used by the operation and maintenance platform (EMS) for advance scheduling, load reduction, or dispatching maintenance orders. Although it does not directly participate in the millisecond-level closed loop, its prediction accuracy is a key indicator reflecting the "advance amount" innovation.

[0044] For example, two branches share feature H t , their respective w,b parameters are obtained through end-to-end supervised learning: t :Use contrast loss to enhance the model’s sensitivity to slight anomalies; Using binary cross entropy (BCE), the goal is to maximize the detection AUC 30 minutes before the accident; the overall loss is the weighted sum of the two (the weights are obtained by grid search on the validation set).

[0045] Intuitive explanation, the form of single-layer linear + Sigmoid is sufficient to transform high-dimensional features h t The output is compressed into a probability scale, which is easy to compare with the threshold and also easy to connect with the traditional risk control strategy (0-1 risk score). t ) and advance prediction Multi-task parallelism not only reduces the model size, but also improves the generalization ability of the two tasks by sharing features.

[0046] In this embodiment, “controlling the energy storage system based on the probability P” specifically includes:

[0047] When R t >When a threshold is preset, a set U is obtained, wherein the set U includes several actions; C risk (u) = p fail (u)P+(1-P)C loss , p fail (u) = exp(-kΔT u ), C total (u)=C act (u)+λC risk (u), u∈U; where C risk (u) represents the expected loss if the system still evolves into thermal runaway after taking action u; C loss Represents the fixed loss caused by a thermal runaway accident; 0 <p fail (u)≤1 indicates the probability of evolving into thermal runaway after action u is completed; ΔT u is the expected temperature drop after taking action u; C act (u) represents the cost of taking action u, 0<λ<1; get the C corresponding to any action u in the set U total (u), get the minimum value C total(u′) corresponds to the action u′, and then the energy storage system is controlled using the action u′.

[0048] Here, optional, in, Indicates that no active measures will be taken for the time being, only monitoring; C risk The unit of (u) may be one currency unit; C loss It can include equipment depreciation, power consumption during downtime, and safety risks; ΔT u The specific value of can be derived from the CFD-experimental hybrid calibration model; k is an empirical coefficient, for example, 0.15 / ℃ is the experimental calibration value, which means that the risk of loss of control decreases by about 15% for every 1℃ drop in temperature; C act (u) represents the cost of the action itself, including refrigerant consumables, power loss during downtime, and component wear. Its value has been explained in detail in the "Inhibition Decision Layer" and will not be repeated here.

[0049] Here, the preset threshold can be 0.6, U = {local cooling injection, reduce charging current, disconnect switch off, exhaust valve open, ...}. Each action u∈U is parameterized by a two-tuple (p, d) - p represents the intensity (such as injection power or current rate), and d represents the duration, which is convenient for enumeration on FPGA / ARM. The execution cost C act (u) = c1p + c2d + c3, where c1 includes the cost of refrigerant or atomizing gas consumables, c2 is converted into electricity loss due to downtime, and c3 is fixed depreciation such as relay wear; all three coefficients are derived from operation and maintenance records and experimental calibration and can be updated annually.

[0050] Risk Cost C risk (u)C risk (u)=P fail (u)L, where L is the expected loss of a loss of control accident - including equipment, downtime and personnel risks. Multi-objective compromise and optimal action u′, let λ∈[0,1] represent the risk aversion coefficient of the operation and maintenance party (usually 0.8 on site). Overall goal The U scale is typically less than 50; FPGAs can compute each action individually, with a worst-case time of less than 50 μs. If the number of actions increases in the future, heuristic sequential search or genetic algorithms can be used to ensure millisecond-level loop closure.

[0051] The second embodiment of the present invention provides a monitoring device for thermal runaway of an energy storage system, wherein the energy storage system includes M battery cells Battery1, Battery2, ..., Battery M , where M is a natural number; including the following modules: an information acquisition module for acquiring multiple times Time1, Time2, ..., Time N Time iEarlier than Time i+1 , where i and N are both natural numbers, 1≤i≤N-1; the pre-processing module is used to obtain the battery of any cell k At Time t Temperature T t,k , voltage V t,k , the energy storage system at time Time t The current I t , gas concentration fraction G t , acoustic signal S t and structural strain ε t , generating vector X t =[T t,1 ,T t,2 ,...,T t,M ,V t,1 ,V t,2 ,...,V t,M ,I t ,G t ,S t ,ε t ], where t and k are both natural numbers, 1≤t≤N, 1≤k≤M; an information processing module is used to obtain the trained neural network, vector X1, vector X2, ..., and vector X N The corresponding codes are input into the neural network to obtain the probability P of thermal runaway output by the neural network. The probability P is used to represent the probability of thermal runaway of the energy storage system in a preset time period from the current time; a control module is used to control the energy storage system based on the probability P.

[0052] In this embodiment, the pre-processing module is further used to: t,j Perform time alignment, missing value filling, outlier suppression, and Z-score standardization; t,j Perform time alignment, missing value filling, outlier suppression, and Z-score standardization. t , all gas concentration fraction G t , all acoustic signals S t and all the structural strains ε t They are all processed by time alignment, missing value filling, outlier suppression and Z-score standardization respectively.

[0053] In this embodiment, the neural network includes a graph convolutional neural network, in which X1, X2, ..., X N Form a matrix H with N rows and 2*M+4 columns (1) ; Among them, is the feature of the node in the L-th layer, is the trainable weight, σ(·) is ReLU, and d L is the feature dimension. Among them, L is a natural number, 1 ≤ L ≤ M, D is the degree matrix, I is the self-loop, A is a matrix. The matrix A contains elements with M rows and M columns. When any battery cell Battery m and any battery cell Battery n are adjacent and share a heat dissipation path, the element A mn = 1; otherwise, the element a mn = 0. Among them, both m and n are natural numbers, 1 ≤ m, n ≤ M; the output of the graph convolutional neural network is H k = H (k) , 1 ≤ k ≤ M.

[0054] In this embodiment, the neural network includes a Transformer module. The Transformer module expands the H t-W+1 , …, H t output by the graph convolutional neural network along the time dimension, and then obtains Q, K, and V through linear projection, and then feeds them into the multi-head self-attention Among them, Q is the query, K is the key, and V is the value; among them, W is a natural number, W < t; the neural network is also used for: calculating the anomaly score Calculating the lead prediction Among them, w r [[ID=Type:35]]is the weight vector corresponding to the anomaly branch, w p is the weight vector corresponding to the lead branch, and b r and b p are both bias terms; the probability of thermal runaway output by the neural network

[0055] In this embodiment, the control module is also used for: when R t > the preset threshold, obtaining the set U, and the set U contains several actions; C risk (u) = p fail (u)P + (1 - P)C loss , p fail (u) = exp(-kΔT u ), C total (u) = C act (u) + λC risk (u), u ∈ U; where, C risk (u) represents the expected loss that still evolves into thermal runaway after taking the action u; C loss represents the fixed loss caused by a thermal runaway accident; 0 < p fail(u)≤1 indicates the probability of evolving into thermal runaway after action u is completed; ΔT u is the expected temperature drop after taking action u; C act (u) represents the cost of taking action u, 0<λ<1; get the C corresponding to any action u in the set U total (u), get the minimum value C total (u′) corresponds to the action u′, and then the energy storage system is controlled using the action u′.

[0056] It should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each implementation method can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0057] The series of detailed descriptions listed above are only specific descriptions of feasible implementation methods of the present invention. They are not intended to limit the scope of protection of the present invention. Any equivalent implementation methods or changes that do not deviate from the technical spirit of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for monitoring thermal runaway of an energy storage system, wherein the energy storage system comprises M battery cells Battery1, Battery2, ..., Battery M ,in, M is a natural number; it is characterized by comprising the following steps: Get multiple times Time1, Time2, ..., Time N Time i Earlier than Time i+1 , where i and N are both natural numbers, 1≤i≤N-1; Get any battery k At Time t Temperature T t,k , voltage V t,k , the energy storage system at time Time t The current I t , gas concentration fraction G t , acoustic signal S t and structural strain ε t , generating vector X t =[T t,1 ,T t,2 ,...,T t,M ,V t,1 ,V t,2 ,...,V t,M ,I t ,G t ,S t ,ε t ], where t and k are both natural numbers, 1≤t≤N, 1≤k≤M; Get the trained neural network, vector X1, vector X2, ..., and vector X N The corresponding codes are input into the neural network to obtain the probability P of thermal runaway output by the neural network, where the probability P is used to represent the probability of thermal runaway of the energy storage system occurring in a preset time period from the current time. The energy storage system is controlled based on the probability P.

2. The monitoring method according to claim 1, characterized in that: The "generated vector X t =[T t,1 ,T t,2 ,...,T t,M ,V t,1 ,V t,2 ,...,V t,M ,I t ,G t ,S t ,ε t ]” specifically include: For all temperatures T t,j Perform time alignment, missing value filling, outlier suppression, and Z-score standardization; Voltage V t,j Perform time alignment, missing value filling, outlier suppression, and Z-score standardization; For all currents I t , all gas concentration fraction G t , all acoustic signals S t and all the structural strains ε t They are all processed by time alignment, missing value filling, outlier suppression and Z-score standardization respectively.

3. The monitoring method according to claim 1, wherein: The neural network includes a graph convolutional neural network, in which X1, X2, ..., X N Form a matrix H with N rows and 2*M+4 columns (1) ; in, is the feature of the L-th layer node, is the trainable weight, σ(·) is ReLU, d L is the feature dimension, where L is a natural number, 1≤L≤M, D is the degree matrix, I is the self-loop, A is a matrix, the matrix A contains M rows and M columns of elements, when any battery m and any battery n When adjacent and sharing a heat dissipation path, element A mn =1, otherwise, element a mn =0; where m and n are both natural numbers, 1≤m, n≤M; the output of the graph convolutional neural network is H k =H (k) , 1≤k≤M.

4. The monitoring method according to claim 1, wherein: The neural network includes a Transformer module, which converts the H output of the graph convolutional neural network into t-W+1 ,…,H t Expand by time dimension, then linearly project to get Q, K, V, and then feed into multi-head self-attention Where Q is the query, K is the key, and V is the value; where W is a natural number, W <t; The neural network is also used to calculate anomaly scores Calculating lead time forecasts Among them, w r is the weight vector corresponding to the abnormal branch, w p is the weight vector corresponding to the advance branch, b r and b p are all bias terms; the probability of thermal runaway output by the neural network 5. The monitoring method according to claim 1, characterized in that: The “controlling the energy storage system based on the probability P” specifically includes: When R t >When a threshold is preset, a set U is obtained, wherein the set U includes several actions; C risk (u) = p fail (u)P+(1-P)C loss , p fail (u) = exp(-kΔT u ), C total (u)=C act (u)+λC risk (u), u∈U; where C risk (u) represents the expected loss if the system still evolves into thermal runaway after taking action u; C loss Represents the fixed loss caused by a thermal runaway accident; 0 <p fail (u)≤1 indicates the probability of evolving into thermal runaway after action u is completed; ΔT u is the expected temperature drop after taking action u; C act (u) represents the cost of taking action u, 0<λ<1; get the C corresponding to any action u in the set U total (u), get the minimum value C total (u′) corresponds to the action u′, and then the energy storage system is controlled using the action u′.

6. A monitoring device for thermal runaway of an energy storage system, wherein the energy storage system comprises M battery cells Battery1, Battery2, ..., Battery M ,in, M is a natural number; it is characterized by including the following modules: Information acquisition module, used to obtain multiple time Time1, Time2, ..., Time N Time i Earlier than Time i+1 , where i and N are both natural numbers, 1≤i≤N-1; Pre-processing module, used to obtain any battery k At Time t Temperature T t,k , voltage V t,k , the energy storage system at time Time t The current I t , gas concentration fraction G t , acoustic signal S t and structural strain ε t , generating vector X t =[T t,1 ,T t,2 ,...,T t,M ,V t,1 ,V t,2 ,...,V t,M ,I t ,G t ,S t ,ε t ], where t and k are both natural numbers, 1≤t≤N, 1≤k≤M; Information processing module, used to obtain the trained neural network, vector X1, vector X2, ..., and vector X N The corresponding codes are input into the neural network to obtain the probability P of thermal runaway output by the neural network, where the probability P is used to represent the probability of thermal runaway of the energy storage system occurring in a preset time period from the current time. A control module is used to control the energy storage system based on the probability P.

7. The monitoring device according to claim 6, characterized in that The pre-processing module is further configured to: For all temperatures T t,j Perform time alignment, missing value filling, outlier suppression, and Z-score standardization; Voltage V t,j Perform time alignment, missing value filling, outlier suppression, and Z-score standardization; For all currents I t , all gas concentration fraction G t , all acoustic signals S t and all the structural strains ε t They are all processed by time alignment, missing value filling, outlier suppression and Z-score standardization respectively.

8. The monitoring device according to claim 6, characterized in that The neural network includes a graph convolutional neural network, in which X1, X2, ..., X N Form a matrix H with N rows and 2*M+4 columns (1) ; in, is the feature of the L-th layer node, is the trainable weight, σ(·) is ReLU, d L is the feature dimension, where L is a natural number, 1≤L≤M, D is the degree matrix, I is the self-loop, A is a matrix, the matrix A contains M rows and M columns of elements, when any battery m and any battery n When adjacent and sharing a heat dissipation path, element A mn =1, otherwise, element a mn =0; where m and n are both natural numbers, 1≤m, n≤M; the output of the graph convolutional neural network is H k =H (k) , 1≤k≤M.

9. The monitoring device according to claim 6, characterized in that: The neural network includes a Transformer module, which converts the H output of the graph convolutional neural network into t-W+1 ,…,H t Expand by time dimension, then linearly project to get Q, K, V, and then feed into multi-head self-attention Where Q is the query, K is the key, and V is the value; where W is a natural number, W <t; The neural network is also used to calculate anomaly scores Calculating lead time forecasts Among them, w r is the weight vector corresponding to the abnormal branch, w p is the weight vector corresponding to the advance branch, b r and b p are all bias terms; the probability of thermal runaway output by the neural network 10. The monitoring device according to claim 6, characterized in that The control module is further configured to: When R t >When a threshold is preset, a set U is obtained, wherein the set U includes several actions; C risk (u) = p fail (u)P+(1-P)C loss , p fail (u) = exp(-kΔT u ), C total (u)=C act (u)+λC risk (u), u∈U; where C risk (u) represents the expected loss if the system still evolves into thermal runaway after taking action u; C loss Represents the fixed loss caused by a thermal runaway accident; 0 <p fail (u)≤1 indicates the probability of evolving into thermal runaway after action u is completed; ΔT u is the expected temperature drop after taking action u; C act (u) represents the cost of taking action u, 0<λ<1; get the C corresponding to any action u in the set U total (u), get the minimum value C total (u′) corresponds to the action u′, and then the energy storage system is controlled using the action u′.

Images