Electrochemical energy storage power station thermal runaway hierarchical early warning system and method

Abstract

The application discloses an electrochemical energy storage power station thermal runaway grading early warning system and method, and the system comprises: a multi-parameter composite detector array, which synchronously collects multi-dimensional time sequence data such as hydrogen, carbon monoxide, VOC, temperature change rate and smoke concentration; a data preprocessing module, which carries out standardization and sliding window sample construction; a branch prediction module, which is provided with three deep learning branch networks of attention CNN-LSTM, Informer and TabNet in parallel, respectively extracts local time sequence features, global long-period trends and feature importance, and outputs a classification probability vector; a fusion decision module determines an early warning grade through first-level weighted voting fusion based on historical performance dynamic weighting and second-level Stacking fusion with a multi-classification logistic regression as a meta-classifier; and an anti-interference rule engine carries out review and verification by using space redundancy, environmental interference discrimination and trend consistency rules. The application can issue an early warning 15-30 minutes before thermal runaway, has high accuracy, and significantly reduces the false alarm and missed alarm rates.

Description

Technical Field

[0001] This invention relates to the field of safety monitoring technology for electrochemical energy storage power stations, and in particular to a graded early warning system and method for thermal runaway in electrochemical energy storage power stations based on multi-source temporal deep learning and ensemble learning. Background Technology

[0002] With the accelerated transformation of the global energy structure, electrochemical energy storage power stations, as a key supporting component of the new power system, have experienced explosive growth in installed capacity. By 2025, my country's installed capacity of new energy storage had exceeded 80GW, and the number of new energy electric vehicles exceeded 30 million. However, lithium-ion batteries are prone to thermal runaway under conditions of overcharging, overheating, and mechanical abuse, which can lead to fires and explosions, seriously threatening the safety of people and property and the stable operation of the power grid.

[0003] Currently, fire monitoring in electrochemical energy storage power stations mainly relies on the following technical means:

[0004] The cabin uses single-point temperature and smoke detectors, which only respond when the temperature or smoke concentration exceeds a preset fixed threshold. However, early signs of battery thermal runaway include trace amounts of characteristic gases being released, and temperature changes are not yet significant. The fixed threshold method is insufficient to capture these early signs, resulting in a severely inadequate warning window.

[0005] Battery Management System (BMS) voltage and temperature monitoring identifies abnormal conditions by monitoring the voltage and surface temperature of individual cells. However, voltage anomalies usually only become apparent after significant side reactions have occurred inside the battery, and surface temperature measurements suffer from a lag effect, making it impossible to provide truly early warnings.

[0006] Single-model early warning methods: Some studies have attempted to use a single machine learning model for thermal runaway prediction. However, the operating conditions of electrochemical energy storage power stations are complex and variable, and the thermal runaway characteristics of different electrochemical systems (lithium iron phosphate, ternary lithium, etc.) are significantly different. Single models have inherent defects in terms of generalization ability and robustness.

[0007] Therefore, there is an urgent need to develop an electrochemical energy storage power station thermal runaway graded early warning system and method that can integrate multi-dimensional characteristic parameters, has early identification capabilities, and adapts to complex operating conditions and disturbances. Summary of the Invention

[0008] The purpose of this invention is to provide a deep learning-based electrochemical energy storage power station thermal runaway graded early warning system and method. By using a multi-parameter detector array and a heterogeneous deep learning network for parallel prediction, combined with a two-level fusion decision and an anti-interference rule engine, early, accurate, and graded early warning of thermal runaway risk can be achieved.

[0009] To address the existing problems and achieve the above objectives, this invention provides a graded early warning system and method for thermal runaway in electrochemical energy storage power stations, the specific solution of which is as follows:

[0010] A graded early warning system for thermal runaway in an electrochemical energy storage power station includes:

[0011] The data acquisition module includes a multi-parameter composite detector array deployed in the energy storage battery compartment. The detector array is distributed in a spatial grid pattern in the front, middle, and rear areas of the battery compartment and at the upper, middle, and lower height levels. It is used to synchronously acquire multi-dimensional time-series data, including hydrogen concentration, carbon monoxide concentration, volatile organic compound concentration, temperature change rate, smoke concentration, flexible pressure deformation value, and oxygen concentration.

[0012] The data preprocessing module is used to standardize the collected raw multidimensional time series data and construct multidimensional time series sample sequences according to a preset sliding window method.

[0013] The branch prediction module includes multiple deep learning branch networks set in parallel, namely: a first branch network, a second branch network, and a third branch network. The first branch network is used to extract local temporal features, the second branch network is used to capture global long-term evolution trends, and the third branch network is used to output feature importance ranking and classification probability. Each branch network outputs a corresponding classification probability vector.

[0014] The fusion decision module is used to perform a two-level fusion decision on the classification probability vectors output by each branch network. The first-level fusion layer dynamically allocates fusion weights based on the historical performance of each branch network on the validation set and performs weighted fusion on the classification probability vectors output by each branch to obtain the first-level fusion probability vector. The second-level fusion layer uses the first-level fusion probability vector as input features and determines the preliminary thermal runaway risk level through a preset meta-classifier.

[0015] The anti-interference rule engine is designed with triple exclusivity logic for the specific scenarios of energy storage power stations:

[0016] 1. "Gas-Pressure" Coupling Logic: In conventional fire alarms, a simultaneous increase in smoke and temperature triggers an alarm. However, in energy storage scenarios, battery bulging (increased flexible pressure) often precedes the generation of smoke. This engine is configured as follows: if only smoke increases without a change in flexible pressure, it is determined to be external dust interference; if flexible pressure increases without smoke, it is determined to be a potential thermal runaway hazard.

[0017] 2. "Oxygen-Hydrogen" Antagonistic Logic: To address the confusion between hydrogen evolution during charging and hydrogen evolution during thermal runaway in the energy storage chamber, the following logic is established: if the hydrogen concentration increases while the oxygen concentration decreases simultaneously, it is determined to be thermal runaway; if the hydrogen concentration increases but the oxygen concentration remains unchanged, it is determined to be a byproduct of charging equalization.

[0018] 3. Dynamic Threshold Mechanism: The primary warning threshold for hydrogen is set at 500 ppm, and for carbon monoxide at 300 ppm. However, these thresholds are dynamically adjusted based on the battery's SOC (State of Charge). At high SOC (>90%), the threshold automatically increases by 20%; at low SOC (<20%), the threshold decreases by 15%.

[0019] The early warning output module is used to output the final thermal runaway risk level and the corresponding confidence score.

[0020] Preferably, the data preprocessing module includes a standardization processing unit and a sliding window construction unit. The standardization processing transforms the features of each dimension into a standard normal distribution with a mean of zero and a standard deviation of one. The sliding window construction unit is used to construct a multidimensional time series sample sequence. Specifically, the window length is set to a preset number of time steps, the sliding step length is set to a preset number of time steps, and continuous time segments are extracted sequentially to form a sample sequence. Each sample contains the multidimensional time series data.

[0021] Preferably, the data preprocessing module further includes a data balancing processing unit, used to perform class balancing processing on the training data using a synthetic minority oversampling method based on K-Means clustering; the data balancing processing unit first performs cluster analysis on the minority class samples to obtain multiple clusters; then allocates the number of synthetic samples proportionally according to the sample density of each cluster; finally, randomly selects samples within each cluster and performs random linear interpolation on the lines connecting the samples to their nearest neighbors in the same cluster to generate synthetic samples, so that the number of samples in each level of thermal runaway warning category is balanced.

[0022] Preferably, the first branch network is a convolutional-long short-term memory fusion network based on an attention mechanism, including a one-dimensional convolutional layer, a bidirectional long short-term memory network, and a self-attention mechanism; the one-dimensional convolutional layer extracts local features from the input, the bidirectional long short-term memory network performs temporal modeling of the convolutional features and outputs a hidden state sequence, and the self-attention mechanism calculates the correlation between the hidden state at each time step and the trainable query vector, generates attention weights after normalization, and performs weighted summation to obtain a context vector, which is then output as a first classification probability vector through a fully connected layer and a Softmax function;

[0023] The second branch network Informer adopts the ProbSparse self-attention mechanism. By evaluating the discriminative power of each query vector and key vector set, it retains only a preset number of active queries with the highest discriminative power to form a sparse query matrix to participate in attention calculation. At the same time, it reduces the sequence length layer by layer through self-attention distillation operation, and finally outputs the second classification probability vector.

[0024] The third branch network TabNet adopts a sequential attention feature selection mechanism. In each decision step, a feature selection mask is generated through the Sparsemax function to perform sparse selection of input features. After multiple decision steps, the masks from each step are aggregated to obtain the global feature importance ranking and output the third classification probability vector.

[0025] Preferably, the fusion weights of the first-level fusion layer are dynamically determined based on the F1 scores of each branch network on the validation set; the meta-classifier of the second-level fusion layer is a multi-class logistic regression model, whose input is the feature vector obtained by concatenating and flattening the probability vectors output by the three branch networks, and whose output is the probability distribution corresponding to the normal state and the multi-level warning state, and the category corresponding to the highest probability is taken as the preliminary thermal runaway risk level.

[0026] Preferably, in the anti-interference rule engine:

[0027] The spatial redundancy fusion rule is as follows: within the same decision time window, the highest thermal runaway risk level output by all valid detectors is taken as the system output level.

[0028] The environmental interference identification rule is as follows: when the maximum temperature rise rate inside the cabin is lower than the first preset threshold, the average concentration of volatile organic compounds in each detector is higher than the second preset threshold, and the average concentration of hydrogen in each detector is lower than the third preset threshold, it is determined to be external interference and the alarm output is suppressed.

[0029] The trend consistency verification rule is as follows: calculate the Pearson correlation coefficient between each pair of temperature change rate, carbon monoxide concentration, and hydrogen concentration within a preset time window. If the average correlation coefficient is lower than the fourth preset threshold, the current alarm level will be automatically downgraded.

[0030] Preferably, the system also includes a model training and hyperparameter optimization module, which uses Bayesian optimization to automatically search for hyperparameters for each branch network, with the goal of minimizing the validation set loss. The optimal hyperparameter configuration is determined iteratively through a Gaussian process surrogate model and an expected improved acquisition function. All three branch networks use a K-fold cross-validation strategy to train multiple model instances, and the outputs of multiple model instances of the same network architecture are integrated through Bagging.

[0031] This invention also discloses a method for graded early warning of thermal runaway in an electrochemical energy storage power station, comprising the following steps:

[0032] S1. Multi-dimensional time-series data are simultaneously collected by a multi-parameter composite detector array deployed in different spatial locations within the energy storage battery compartment. The multi-dimensional time-series data includes hydrogen concentration, carbon monoxide concentration, volatile organic compound concentration, temperature change rate and smoke concentration, flexible pressure deformation value and oxygen concentration.

[0033] S2. Standardize the collected raw multidimensional time series data and construct a multidimensional time series sample sequence according to a preset sliding window method.

[0034] S3. Input the preprocessed multi-dimensional time series sample sequence into three parallel deep learning branch networks respectively. The first branch network is used to extract local time series features and output the first classification probability vector. The second branch network is used to capture the global long-term evolution trend and output the second classification probability vector. The third branch network is used to output the feature importance ranking and output the third classification probability vector.

[0035] S4. In the first-level weighted voting fusion layer, the fusion weights are dynamically allocated according to the historical performance of each branch network on the validation set, and the classification probability vectors output by each branch are weighted and fused to obtain the first-level fusion probability vector.

[0036] S5. In the second-level Stacking fusion layer, the first-level fusion probability vector is used as the input feature, and the risk level is determined by a preset meta-classifier to output the preliminary warning level.

[0037] S6. The preliminary warning level is verified using an anti-interference rule engine based on spatial redundancy fusion rules, environmental interference identification rules and trend consistency verification rules to obtain the verified risk level.

[0038] S7. Output the final thermal runaway risk level and corresponding confidence score after verification.

[0039] Preferably, the training process of the three deep learning branch networks in step S3 includes: performing class balancing on the training data using a synthetic minority oversampling method based on K-Means clustering; automatically searching for the hyperparameters of each branch network using a Bayesian optimization method; training multiple model instances using a K-fold cross-validation strategy; and integrating multiple model instances of the same network architecture using a Bagging method.

[0040] Preferably, the thermal runaway risk level is divided into five levels: Level 0 represents the normal state, Level 1 represents the first-level warning (attention state), Level 2 represents the second-level warning (concern state), Level 3 represents the third-level warning (severe state), and Level 4 represents the fourth-level warning (emergency state). When the risk level reaches Level 2, an alarm message is automatically pushed to the operation and maintenance personnel; when the risk level reaches Level 3, the audible and visual alarm device is automatically triggered and the ventilation system is started; when the risk level reaches Level 4, the emergency response plan is automatically executed, including cutting off the high-voltage circuit of the battery compartment and activating the fire extinguishing system to standby status.

[0041] The beneficial effects of this invention are as follows:

[0042] 1. Early warning time: By integrating early release information of characteristic gases such as hydrogen, carbon monoxide, and volatile organic compounds, and combining the multidimensional characteristics of temperature change rate and smoke concentration, potential risk trends can be identified 15-30 minutes before thermal runaway occurs, providing sufficient time window for emergency response.

[0043] 2. Improved alarm accuracy: Three heterogeneous deep learning networks are used in parallel to extract differentiated features. Combined with a two-level ensemble learning and decision-making mechanism, the shortcomings of insufficient generalization ability of a single model are overcome. The accuracy on the test set reaches 97.46%, and the precision, recall and F1 score all exceed 97%.

[0044] 3. Strong anti-interference capability: Through a multi-layer verification mechanism of spatial redundancy fusion rules, environmental interference identification rules and trend consistency verification rules, the probability of false alarms and missed alarms caused by factors such as local sensor failures and external environmental interference is effectively reduced.

[0045] 4. High interpretability: The feature importance ranking output by the TabNet network makes the early warning decision-making process transparent and traceable, which facilitates the operation and maintenance personnel to trace the source of faults and identify responsibilities, and enhances the credibility of the system in engineering applications.

[0046] 5. High adaptability: The system adopts a modular design, which can be fine-tuned and transferred to learn the model according to the thermal runaway characteristics of batteries in different electrochemical systems. It is widely applicable to various energy storage battery systems such as lithium iron phosphate, ternary lithium, and lithium titanate.

[0047] 6. Unique Monitoring Dimensions: By introducing flexible pressure deformation monitoring and in-cabin oxygen concentration monitoring, the physical limitations of traditional gas / temperature monitoring are overcome. The flexible pressure sensor can capture the micron-level deformation caused by gas generation inside the battery in the very early stage of thermal runaway (T-30min), while the sudden drop in oxygen concentration serves as a specific indicator of electrolyte decomposition. The combination of the two effectively solves the hysteresis problem of single gas sensors in low-temperature environments. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart illustrating the thermal runaway risk classification and early warning process of the present invention.

[0050] Figure 2 This is the convergence curve for hyperparameter optimization of the deep learning network in this invention;

[0051] Figure 3 This is a comparison chart of the confusion matrices of the various models in this invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] like Figure 1-3 This invention discloses a deep learning-based electrochemical energy storage power station thermal runaway graded early warning system, including:

[0054] 1. Data acquisition layer, including a multi-parameter composite detector array deployed in the energy storage battery compartment. The detector array is distributed in a spatial grid pattern in the front, middle, and rear areas of the battery compartment and in the upper, middle, and lower height levels. Each detector integrates a hydrogen concentration sensor, a carbon monoxide concentration sensor, a volatile organic compound concentration sensor, a temperature sensor, and a smoke concentration sensor.

[0055] 2. Data preprocessing module, used to filter and denoise, fill missing values ​​and standardize the collected raw multi-source time series data, and construct multi-dimensional time series sample sequences according to the sliding window method.

[0056] Specifically, the standardization process in the data preprocessing module uses the StandardScaler standardization method to transform each dimension of features into a standard normal distribution with a mean of 0 and a standard deviation of 1. The transformation formula is as follows:

[0057]

[0058] in, Indicates the first The first time step Original values ​​of dimensional features For the first The mean of the dimensional feature across all samples. For the first Standard deviation of dimensional features.

[0059] The specific method for constructing multidimensional time-series sample sequences using the sliding window approach involves setting the window length. Each time step has a sliding step size of [number] times. Then, for a time step of 1, the 2nd time step... Each sample is represented as:

[0060]

[0061] in, for Real-time data collection 3D feature vectors The corresponding concentrations are hydrogen, carbon monoxide, volatile organic compounds, temperature change rate, and smoke concentration.

[0062] The data preprocessing module also includes a data balancing unit, which uses a synthetic minority oversampling method based on K-Means clustering to perform class balancing on the training data. The steps include:

[0063] For minority class sample sets Perform K-Means clustering to obtain Clusters For each cluster Calculate the sample distribution density within the cluster, and allocate the number of synthetic samples for the cluster according to the density ratio. :

[0064]

[0065] in, The target is the total number of synthetic samples.

[0066] Within each cluster, a sample is randomly selected. within the same cluster Random linear interpolation is performed on the lines connecting the nearest neighbor samples to generate synthetic samples. :

[0067]

[0068] in, These are randomly selected nearest neighbor samples within the same cluster. These are uniformly distributed random numbers.

[0069] 3. Branch prediction module, which includes three deep learning branch networks configured in parallel, namely:

[0070] Convolutional-Long Short-Term Memory Fusion Network (CNN-LSTM branch) based on attention mechanism is used to extract local temporal features and contextual dependencies.

[0071] The Informer long sequence prediction network is used to capture global long-term evolutionary trends.

[0072] TabNet is an interpretable deep learning network used to output feature importance rankings and preliminary classification probabilities.

[0073] Specifically, the mathematical description of the attention-based convolutional-long short-term memory fusion network is as follows: Let the input temporal features be... First, local features are extracted using a one-dimensional convolutional layer:

[0074]

[0075] in This represents a one-dimensional convolution operation. For convolution kernel weights, This is a bias term.

[0076] Convolutional features Inputting into a bidirectional long short-term memory network yields a temporal sequence of hidden states:

[0077]

[0078] We introduce a self-attention mechanism to compute the weighted representation of the hidden state at each time step:

[0079]

[0080]

[0081]

[0082] in For trainable query vectors, , , For trainable parameters, This is a weighted context vector.

[0083] Finally, the classification probability is output through a fully connected layer and a softmax function:

[0084]

[0085] The Informer long sequence prediction network employs the ProbSparse self-attention mechanism, whose core formula is:

[0086]

[0087] in, To and Same size but only includes Top- A sparse matrix of active queries, with the selection criteria for active queries as follows:

[0088]

[0089] The larger the value of M(⋅), the higher the distinguishability between the query vector and the key vector, and the more likely it is to be retained.

[0090] The Informer network reduces the sequence length layer by layer through self-attention distillation, ultimately outputting the classification probability. :

[0091]

[0092] The TabNet-based interpretable deep learning network employs a sequential attention feature selection mechanism. The decision-making process for the next step is as follows:

[0093] Feature selection mask:

[0094]

[0095] Feature processing:

[0096]

[0097] The Sparsemax function is defined as the Euclidean projection onto the simplex:

[0098]

[0099] Final output classification probability and feature importance vector .

[0100] 4. Fusion Decision Module: This module includes a first-level weighted voting fusion layer and a second-level Stacking generalization fusion layer. The first-level weighted voting fusion layer dynamically allocates fusion weights based on the historical performance of each branch network on the validation set. The second-level Stacking fusion layer uses the probability vector output by the first level as input features and performs the final risk level determination through a meta-classifier.

[0101] Specifically, the weight allocation formula for the first-level weighted voting fusion layer is as follows:

[0102]

[0103]

[0104] in, Let F1 scores be vectors for the three branch networks on the validation set. The temperature coefficient controls the smoothness of the weight distribution.

[0105] The second-level Stacking fusion layer uses the concatenated probability vectors from the outputs of the three branch networks as input features:

[0106]

[0107] Will After flattening, the data is input into the meta-classifier. This system uses multi-class logistic regression as the meta-classifier.

[0108]

[0109] The final warning level is:

[0110]

[0111] 5. Anti-interference rule engine, used to review and verify the fusion decision results based on spatial redundancy fusion rules, environmental interference identification rules, and trend consistency verification rules;

[0112] Specifically, the spatial redundancy fusion rule in the anti-interference rule engine is expressed as follows:

[0113]

[0114] in, For the set of all effective detectors, For detector exist The risk level is output in real time. The final integration level adopted by the system.

[0115] The mathematical criterion for the environmental interference identification rule is that if the following conditions are met, it is determined to be external interference and the alarm output is suppressed:

[0116]

[0117] in, This represents the maximum rate of temperature rise inside the cabin. , , This is a preset threshold.

[0118] The trend consistency verification rule uses the Pearson correlation coefficient test:

[0119]

[0120] If the mean correlation coefficient among key features is lower than the threshold If so, the alarm level will be automatically downgraded.

[0121] 6. Early warning output module: Used to output the final thermal runaway risk level and the corresponding confidence score.

[0122] 7. The system of the present invention also includes a model training and hyperparameter optimization module, which uses a Bayesian optimization method for hyperparameter search, and its objective function is:

[0123]

[0124] Bayesian optimization uses a Gaussian process as a surrogate model:

[0125]

[0126] The acquisition function is expected to be improved as follows:

[0127]

[0128] This invention also discloses a deep learning-based method for graded early warning of thermal runaway in electrochemical energy storage power stations, comprising the following steps:

[0129] Step S1: Using a multi-parameter composite detector array deployed in the energy storage battery compartment, hydrogen concentration, carbon monoxide concentration, volatile organic compound concentration, temperature change rate, and smoke concentration data are collected synchronously at a preset sampling frequency.

[0130] Step S2: Filter and denoise the collected raw multi-source time series data and fill in missing values. Standardize the data using the StandardScaler method and construct a multi-dimensional time series sample sequence using a sliding window method.

[0131] Step S3: Input the preprocessed sample sequences into three parallel deep learning sub-networks: an attention-based CNN-LSTM network extracts local temporal features, an Informer network captures global long-term trends, and a TabNet network outputs feature importance ranking and preliminary classification probabilities. Specifically, the training process of the above three parallel deep learning sub-networks includes:

[0132] (1) The K-Means SMOTE method is used to perform class balancing on the training data.

[0133] (2) The Bayesian optimization method is used to automatically search for network hyperparameters.

[0134] (3) Multiple model instances are trained using the K-fold cross-validation strategy and integrated using the Bagging method.

[0135] Step S4: The first-level weighted voting fusion layer dynamically allocates weights based on the historical performance of each branch network and performs weighted fusion of the output probabilities of each branch.

[0136] Step S5: The second-level Stacking fusion layer takes the output probability vector of the first level as input and uses the trained meta-classifier to make the final risk level determination.

[0137] Step S6: The anti-interference rule engine performs spatial redundancy verification, environmental interference identification, and trend consistency verification on the fusion judgment results.

[0138] Step S7: Output the final thermal runaway risk level and confidence score. The risk level is divided into five levels: level 0 represents normal state, level 1 represents level 1 warning - attention, level 2 represents level 2 warning - concern, level 3 represents level 3 warning - severe, and level 4 represents level 4 warning - emergency.

[0139] Example 1: System Hardware Composition and Deployment

[0140] This embodiment provides a specific hardware implementation of a deep learning-based electrochemical energy storage power station thermal runaway graded early warning system.

[0141] 1. A multi-parameter composite detector array, comprising 12 detectors deployed in a spatial grid pattern within a standard 20-foot energy storage battery compartment. The specific deployment scheme divides the battery compartment horizontally into three regions along its longitudinal direction: front, middle, and rear. Vertically, it is divided into three levels: upper (0.3m from the top), middle (1.5m from the bottom), and lower (0.3m from the bottom). Each composite detector integrates the following sensing elements:

[0142] ① Electrochemical hydrogen sensor: range 0-1000ppm, resolution 1ppm, response time T90≤30s;

[0143] ② Electrochemical carbon monoxide sensor: range 0-500ppm, resolution 0.5ppm, response time T90≤30s;

[0144] ③ Photoionization VOC sensor: range 0-50ppm, resolution 0.01ppm, response time T90≤10s;

[0145] ④PT100 platinum resistance temperature sensor: measuring range -40℃ to 150℃, accuracy ±0.1℃;

[0146] ⑤ Photoelectric smoke concentration sensor: range 0-20%obs / m, resolution 0.01%obs / m.

[0147] 2. Data Acquisition and Communication Module: The detector connects to the data acquisition gateway via a CAN bus, with a communication baud rate set to 250kbps. The gateway uses an industrial-grade CAN-Ethernet conversion module with opto-isolation design, supports a wide operating temperature range of -40℃ to 85℃, and has an IP65 protection rating. The acquisition frequency is set to 1Hz, meaning one complete set of multidimensional data is acquired per second.

[0148] 3. Edge Computing Unit: The edge computing unit uses an NVIDIA Jetson Orin NX module, equipped with 8GB of video memory and 1024 CUDA cores, supporting ONNX Runtime inference acceleration. The operating environment is Ubuntu 20.04 LTS, and the deep learning inference framework uses ONNX Runtime 1.14.

[0149] Example 2: Model Training Process

[0150] This embodiment details the training process of the three deep learning branch networks and the Stacking fusion model.

[0151] Step 1: Training Data Acquisition. On a test platform for a specific type of energy storage battery module, thermal runaway experiments were conducted on lithium iron phosphate battery modules, covering different states of charge (SOC=30%, 50%, 80%, 100%) and different triggering methods (overcharging, external heating, needle penetration). Simultaneously, typical environmental interference experiments (diesel generator exhaust, welding fumes, maintenance dust) were carried out. A total of 1256 sets of valid experimental data samples under normal operating conditions and 199 sets of data samples under thermal runaway conditions were collected.

[0152] Step 2, Data Preprocessing and Balancing: Training samples are constructed using the sliding window method, with a window length T=60 (corresponding to 60 seconds) and a sliding step size S=10. The original data is then standardized using StandardScaler.

[0153] The K-Means SMOTE method was used to oversample the minority class, with the number of clusters K=5. The target number of synthesized clusters is shown in Table 1.

[0154] Table 1: Class Distribution Before and After K-Means SMOTE Oversampling Processing

[0155] Risk level classification Original sample size Number of samples after oversampling Level 0 (Normal Status) 1256 1256 Level 1 (Level 1 Warning - Caution) 87 1250 Level 2 (Level II Warning - Attention) 53 1248 Level 3 (Level 3 Warning - Severe) 31 1255 Level 4 (Level IV Warning - Emergency) 28 1252

[0156] Step 3: Branch network hyperparameter optimization. A Bayesian optimization method is used to search for hyperparameters for each branch network, with 50 iterations. The main hyperparameters to be optimized include:

[0157] ①CNN-LSTM network: number of convolutional kernels {32, 64, 128}, number of LSTM hidden units {64, 128, 256}, learning rate [1e-4, 1e-2];

[0158] ②Informer network: number of encoder layers {2, 3, 4}, number of attention heads {4, 8}, distillation step size {2, 3};

[0159] ③ TabNet network: number of decision steps {3, 5, 8}, number of attention hidden layer units {64, 128}, relaxation factor [1.0, 2.0].

[0160] The convergence curves for hyperparameter optimization of each network are as follows: Figure 2 As shown in Table 2, the hyperparameter optimization search space and optimal configuration are as follows.

[0161] Table 2 shows the hyperparameter optimization search space and optimal configuration.

[0162] Network Name Hyperparameter name Search scope Optimal value Attention CNN-LSTM Number of convolution kernels {32, 64, 128} 64 Number of hidden units in LSTM {64, 128, 256} 128 Learning rate [1e-4, 1e-2] 8.5e-4 Dropout ratio [0.1, 0.5] 0.23 Informer Encoder layers {2, 3, 4} 3 Number of attention heads {4, 8} 8 Distillation step size {2, 3} 2 Learning rate [1e-4, 1e-2] 5.2e-4 TabNet Decision steps {3, 5, 8} 5 Attention Hidden Unit {64, 128} 128 relaxation factor [1.0, 2.0] 1.42 Learning rate [1e-3, 5e-2] 1.8e-2

[0163] Step 4: K-fold cross-validation and Bagging integration. A 5-fold cross-validation strategy is used, with 5 model instances trained for each branch of the network. Models trained on different data partitions using the same network architecture are integrated using Bagging, which involves averaging the output probabilities of the 5 model instances.

[0164] Step 5: Stacking Meta-Classifier Training. The output probability vectors of the three branch Bagging models on the validation set are concatenated as the training features of the Stacking Meta-Classifier (Multi-class Logistic Regression). The meta-classifier uses L2 regularization, and the regularization coefficient C is determined in {0.1, 1.0, 10.0} through grid search. The confusion matrix is ​​compared as follows: Figure 3 As shown.

[0165] To demonstrate that the Stacking fusion model outperforms the single-branch model and the simple voting fusion model, the performance of each model on the test set was tested. The test results are shown in Table 3. The results show that the Stacking fusion model significantly outperforms the single model and the simple weighted voting fusion model in terms of accuracy, precision, recall and F1 score.

[0166] Table 3: Comparison of performance indicators of each model

[0167] Model Name accuracy Precision (Macro Average) Recall rate (macro average) F1 score (macro average) Attention CNN-LSTM Bagging 95.62% 95.78% 95.62% 95.67% Informer Bagging 94.87% 95.03% 94.87% 94.91% TabNet Bagging 95.14% 95.29% 95.14% 95.18% Weighted voting fusion 96.83% 96.95% 96.83% 96.86% Stacking Two-Level Fusion (This Invention) 97.46% 97.58% 97.46% 97.49%

[0168] Example 3: Actual Operation Example

[0169] This embodiment uses a simulated early-stage battery thermal runaway scenario as an example to illustrate the system's workflow.

[0170] Scenario setting: The temperature of the second battery module in the No. 3 battery compartment of an energy storage power station is slowly rising due to an internal micro-short circuit.

[0171] Time T0 (normal operation): All data collected by each detector are within the normal range, the system outputs risk level 0, and confidence level 99.2%.

[0172] At time T0+5min: the upper-layer detector detected that the CO concentration increased from the background value of 2ppm to 8ppm, the VOC concentration increased from 0.1ppm to 0.4ppm, the H2 concentration did not change significantly, and the temperature change rate was 0.05℃ / min. In the probability distribution of the CNN-LSTM branch output, the probability of a Level 1 warning was 68%, the Informer branch was 57%, and the TabNet branch was 72%. After weighted voting fusion, the probability of a Level 1 warning was 66%, and the Stacking fusion determined it to be a Level 1 warning. The system main interface dashboard pointer entered the yellow attention area, displaying "Level 1 Warning - Attention" with a confidence level of 89.3%.

[0173] At time T0+12min: CO concentration rose to 25ppm, H2 concentration rose to 12ppm, VOC concentration rose to 1.2ppm, and temperature change rate rose to 0.3℃ / min. The probability of a Level 2 warning exceeded 75% in all three branch network outputs, and after fusion, it was determined to be a Level 2 warning. The system interface displayed "Level 2 Warning - Attention" with a confidence level of 94.1%, and simultaneously automatically pushed alarm information to the mobile terminals of maintenance personnel.

[0174] At time T0+18min: CO concentration rises sharply to 85ppm, H2 concentration rises to 45ppm, and temperature change rate rises to 1.5℃ / min. The system determines this to be a Level 3 warning, triggering the audible and visual alarm and automatically activating the battery compartment ventilation system.

[0175] Time T0+22min: All indicators continue to deteriorate, the system judges it to be a Level 4 warning - emergency, with a confidence level of 98.7%, and automatically executes the emergency response plan: cut off the high-voltage circuit of the battery compartment, activate the total flooding gas fire extinguishing system to standby status, and send a linkage signal to the fire department.

[0176] Example 4: Anti-interference verification

[0177] This embodiment verifies the system's performance under external interference conditions.

[0178] Scenario 1: Diesel generator exhaust interference. When a diesel generator is started near the battery compartment air inlet, exhaust fumes drift into the compartment. Multiple detectors simultaneously raise VOC concentrations to 3.5 ppm, but the temperature change rate remains below 0.02℃ / min, and the H2 concentration remains below 2 ppm. The environmental interference identification rules determine that the interference conditions are met, and the system maintains a level 0 normal output without generating false alarms.

[0179] Scenario 2: Single-point sensor failure. The CO sensor output in the upper middle layer of the simulation abnormally jumps to 80ppm, but the data from other sensors at the same location and adjacent locations are normal. The spatial redundancy fusion rule takes the highest level of all detectors for fusion. Since the anomaly is only at a single point, trend consistency verification shows that the correlation coefficient between CO and other features is only 0.12 (below the threshold of 0.6). The system automatically marks this point as a suspicious state and downgrades it, while the output level remains normal.

[0180] Table 4 shows the parameter configuration for the anti-interference rule engine:

[0181] Table 4: Anti-interference rule engine parameter configuration

[0182] Parameter symbol Parameter meaning Value unit Temperature rise rate determination threshold 0.1 ℃ / min VOC concentration determination threshold 2.0 ppm <![CDATA[H2 concentration determination threshold]]> 5.0 ppm Trend Consistency Analysis Window Length 10 Sampling points Lowest threshold for correlation coefficient 0.6 — Weighted voting temperature coefficient 0.5 —

[0183] Example 5: Performance Comparison Test

[0184] To verify the effectiveness of the technical solution of the present invention, a comparative test was conducted with existing technical solutions.

[0185] Comparison with Option 1: Fixed threshold alarm method (alarm triggered when CO > 50 ppm or temperature > 60℃);

[0186] Comparison with Option 2: Single CNN-LSTM model method;

[0187] Comparison with Option 3: Single Informer model approach.

[0188] The test dataset contains 50 thermal runaway condition samples and 200 normal operation samples. The evaluation indicators include average warning lead time, accuracy, false alarm rate, and missed alarm rate. The test results are shown in Table 5.

[0189] Table 5: Performance Comparison of Different Technical Solutions

[0190] Fixed threshold method (CO>50ppm or T>60℃) -5.2 min (lag) 82.4% 12.0% 8.5% Single CNN-LSTM model +8.7 min 92.1% 5.2% 3.1% Single Informer Model +12.3 min 91.5% 6.1% 2.9% The method of this invention (Stacking fusion + anti-interference rules) +18.6 min 97.5% 1.8% 0.7%

[0191] Note: A positive average warning lead time indicates that the warning was issued earlier than the fixed threshold method.

[0192] Test results show that the method of the present invention is significantly superior to existing technical solutions in key indicators such as early warning time, accuracy, false alarm rate, and missed alarm rate.

[0193] The present invention has been described in detail with reference to the foregoing embodiments. Those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A graded early warning system for thermal runaway in an electrochemical energy storage power station, characterized in that, include: The data acquisition module includes a multi-parameter composite detector array deployed within the energy storage battery compartment. The detector array is spatially gridded across the front, middle, and rear areas of the battery compartment, as well as at upper, middle, and lower height levels. It is used to simultaneously acquire multi-dimensional time-series data, including hydrogen concentration, carbon monoxide concentration, volatile organic compound concentration, temperature change rate, smoke concentration, flexible pressure deformation value, and oxygen concentration. The flexible pressure deformation value is acquired by a flexible piezoresistive sensor attached to the surface of the battery module, used to characterize the microscopic expansion features of the battery casing. The oxygen concentration is used to assist in determining the intensity of the oxidative environment in a thermal runaway chain reaction. The data preprocessing module is used to standardize the collected raw multidimensional time series data and construct multidimensional time series sample sequences according to a preset sliding window method. The branch prediction module includes multiple deep learning branch networks set in parallel, namely: a first branch network, a second branch network, and a third branch network. The first branch network is used to extract local temporal features, the second branch network is used to capture global long-term evolution trends, and the third branch network is used to output feature importance ranking and classification probability. Each branch network outputs a corresponding classification probability vector. The fusion decision module is used to perform a two-level fusion decision on the classification probability vectors output by each branch network. The first-level fusion layer dynamically allocates fusion weights based on the historical performance of each branch network on the validation set and performs weighted fusion on the classification probability vectors output by each branch to obtain the first-level fusion probability vector. The second-level fusion layer uses the first-level fusion probability vector as input features and determines the preliminary thermal runaway risk level through a preset meta-classifier. An anti-interference rule engine is used to verify the preliminary thermal runaway risk level based on spatial redundancy fusion rules, environmental interference identification rules, and trend consistency verification rules, so as to obtain the verified risk level. The early warning output module is used to output the final thermal runaway risk level and the corresponding confidence score.

2. The system according to claim 1, characterized in that: The data preprocessing module includes a standardization processing unit and a sliding window construction unit. The standardization processing transforms the features of each dimension into a standard normal distribution with a mean of zero and a standard deviation of one. The sliding window construction unit is used to construct a multidimensional time series sample sequence. Specifically, the window length is set to a preset number of time steps, the sliding step length is set to a preset number of time steps, and continuous time segments are extracted sequentially to form a sample sequence. Each sample contains the multidimensional time series data.

3. The system according to claim 1, characterized in that: The data preprocessing module also includes a data balancing processing unit, which is used to perform class balancing processing on the training data using a synthetic minority oversampling method based on K-Means clustering. The data balancing processing unit first performs cluster analysis on the minority class samples to obtain multiple clusters; then allocates the number of synthetic samples according to the sample density of each cluster; finally, it randomly selects samples within each cluster and performs random linear interpolation on the lines connecting the samples with the nearest neighbors in the same cluster to generate synthetic samples, so that the number of samples in each level of thermal runaway warning category is balanced.

4. The system according to claim 1, characterized in that: The first branch network is a convolutional-long short-term memory fusion network based on an attention mechanism, including a one-dimensional convolutional layer, a bidirectional long short-term memory network, and a self-attention mechanism. The one-dimensional convolutional layer extracts local features from the input, the bidirectional long short-term memory network performs temporal modeling of the convolutional features and outputs a hidden state sequence, and the self-attention mechanism calculates the correlation between the hidden state at each time step and the trainable query vector, generates attention weights after normalization, and performs weighted summation to obtain a context vector, which is then output as a first classification probability vector through a fully connected layer and a Softmax function. The second branch network Informer adopts the ProbSparse self-attention mechanism. By evaluating the discriminative power of each query vector and key vector set, it retains only a preset number of active queries with the highest discriminative power to form a sparse query matrix to participate in attention calculation. At the same time, it reduces the sequence length layer by layer through self-attention distillation operation, and finally outputs the second classification probability vector. The third branch network TabNet adopts a sequential attention feature selection mechanism. In each decision step, a feature selection mask is generated through the Sparsemax function to perform sparse selection of input features. After multiple decision steps, the masks from each step are aggregated to obtain the global feature importance ranking and output the third classification probability vector.

5. The system according to claim 1, characterized in that: The first-level fusion layer is a feature-level fusion based on dynamic weights. It assigns weights according to the historical performance of each branch network on the validation set and performs weighted aggregation of local features, long-term trends, and feature importance. The second-level fusion layer is a decision-level fusion, which takes the result of the first-level fusion as input and makes the final judgment through a preset meta-classifier. The second-level fusion layer is configured to assign dominant weights to the output of a single branch network at a specific thermal runaway evolution stage in order to overcome the decision bias caused by the homogenization of multiple models. Its input is the feature vector obtained by concatenating and flattening the probability vectors output by the three branch networks, and the output is the probability distribution corresponding to the normal state and the multi-level warning state. The category corresponding to the highest probability is taken as the preliminary thermal runaway risk level.

6. The system according to claim 1, characterized in that, In the anti-interference rule engine: The spatial redundancy fusion rule is as follows: within the same decision time window, the highest thermal runaway risk level output by all valid detectors is taken as the system output level. The environmental interference identification rules are as follows: the judgment is based on the thermal runaway gas evolution sequence unique to electrochemical energy storage power stations; when the maximum temperature rise rate inside the chamber is lower than the first preset threshold, the average concentration of volatile organic compounds in each detector is higher than the second preset threshold, and the average concentration of hydrogen in each detector is lower than the third preset threshold, it is determined to be external organic volatile interference and the alarm output is suppressed; when the oxygen concentration decrease rate is detected to exceed 0.5% / min and is accompanied by an increase in the flexible pressure value, even if the hydrogen concentration does not exceed the standard, it is also determined to be an early characteristic of thermal runaway. The trend consistency verification rule is as follows: calculate the Pearson correlation coefficient between each pair of temperature change rate, carbon monoxide concentration, and hydrogen concentration within a preset time window. If the average correlation coefficient is lower than the fourth preset threshold, the current alarm level will be automatically downgraded.

7. The system according to claim 1, characterized in that: It also includes a model training and hyperparameter optimization module, which uses Bayesian optimization to automatically search for hyperparameters for each branch network. With the goal of minimizing the validation set loss, the optimal hyperparameter configuration is determined iteratively through a Gaussian process surrogate model and the desired improved acquisition function. All three branch networks use a K-fold cross-validation strategy to train multiple model instances, and the outputs of multiple model instances of the same network architecture are integrated through Bagging.

8. A method for graded early warning of thermal runaway in an electrochemical energy storage power station, characterized in that, Includes the following steps: S1. Multi-dimensional time-series data are simultaneously collected by a multi-parameter composite detector array deployed in different spatial locations within the energy storage battery compartment. The multi-dimensional time-series data includes hydrogen concentration, carbon monoxide concentration, volatile organic compound concentration, temperature change rate and smoke concentration, flexible pressure deformation value and oxygen concentration. S2. Standardize the collected raw multidimensional time series data and construct a multidimensional time series sample sequence according to a preset sliding window method. S3. Input the preprocessed multi-dimensional time series sample sequence into three parallel deep learning branch networks respectively. The first branch network is used to extract local time series features and output the first classification probability vector. The second branch network is used to capture the global long-term evolution trend and output the second classification probability vector. The third branch network is used to output the feature importance ranking and output the third classification probability vector. S4. In the first-level weighted voting fusion layer, the fusion weights are dynamically allocated according to the historical performance of each branch network on the validation set, and the classification probability vectors output by each branch are weighted and fused to obtain the first-level fusion probability vector. S5. In the second-level Stacking fusion layer, the first-level fusion probability vector is used as the input feature, and the risk level is determined by a preset meta-classifier to output the preliminary warning level. S6. The preliminary warning level is verified using an anti-interference rule engine based on spatial redundancy fusion rules, environmental interference identification rules and trend consistency verification rules to obtain the verified risk level. S7. Output the final thermal runaway risk level and corresponding confidence score after verification.

9. The method according to claim 8, characterized in that, The training process of the three deep learning branch networks in step S3 includes: performing class balancing on the training data using a synthetic minority oversampling method based on K-Means clustering; automatically searching for hyperparameters of each branch network using a Bayesian optimization method; training multiple model instances using a K-fold cross-validation strategy; and integrating multiple model instances of the same network architecture using a Bagging method.

10. The method according to claim 8, characterized in that: The thermal runaway risk level is divided into five levels: Level 0 represents the normal state, Level 1 represents the first warning level (attention state), Level 2 represents the second warning level (concern state), Level 3 represents the third warning level (severe state), and Level 4 represents the fourth warning level (emergency state). When the risk level reaches Level 2, an alarm message is automatically pushed to the operation and maintenance personnel. When the risk level reaches Level 3, the audible and visual alarm device is automatically triggered and the ventilation system is started. When the risk level reaches Level 4, the emergency response plan is automatically executed, including cutting off the high-voltage circuit of the battery compartment and activating the fire extinguishing system to standby status.

Images