Battery cell early thermal runaway pre-warning method based on operation parameter fusion and isolation forest algorithm optimization

Abstract

The application relates to a battery monomer early thermal runaway early warning method based on operation parameter fusion and an isolated forest algorithm optimization, a double-engine drive of LSTM time sequence feature enhancement and isolated forest dynamic optimization is achieved, early perception-accurate identification-real-time early warning of battery thermal runaway is realized, the hysteresis limitation of traditional BMS threshold alarm is broken, early warning nodes are advanced to before the start of thermal runaway chain reaction, the positioning difficulty of abnormal monomers in large-scale battery clusters is solved, large quantity monomers / cabinet level efficient detection is realized, an active safety protection system of an energy storage power station is constructed, and technical support that can be landed is provided for intelligent operation and maintenance.

Description

Technical Field

[0001] This invention belongs to the field of electrochemical energy storage technology, specifically relating to an early warning method for thermal runaway of battery cells based on the fusion of operating parameters and optimization by the isolated forest algorithm, which is applicable to real-time safety monitoring of large-scale energy storage power stations. Background Technology

[0002] The rapid development of renewable energy sources, represented by photovoltaics and wind power, has placed higher demands on the stability and flexibility of power systems. Electrochemical energy storage systems, due to their advantages such as fast response and flexible configuration, play a crucial role in renewable energy consumption, grid peak shaving, and ancillary services, and have become an indispensable component of modern power systems. Energy storage power stations, as the main application form of electrochemical energy storage technology, are centered on battery systems composed of a large number of batteries. The operational safety and stability of these battery systems directly affect the safe operation of the entire energy storage power station and even the entire power system.

[0003] However, during long-term operation, batteries undergo complex physicochemical changes, leading to gradual performance degradation and the potential for serious safety incidents due to overcharging, over-discharging, short circuits, or overheating. Especially under extreme conditions, violent chain reactions can occur within the battery, releasing a large amount of heat in a short time, causing a rapid temperature rise and resulting in thermal runaway. Once a single battery cell experiences thermal runaway, it can easily trigger a chain reaction in adjacent batteries, causing the entire energy storage chamber to catch fire or even explode, resulting in significant economic losses and safety hazards.

[0004] Currently, energy storage power stations generally face challenges such as a large number of batteries, complex operating environments, and difficulties in condition monitoring. Traditional battery management systems (BMS) have limitations in anomaly detection accuracy and early warning capabilities, making it difficult to identify potential risks in the early stages of thermal runaway. Therefore, there is an urgent need to conduct research on the accurate identification and correlation analysis of early thermal runaway symptoms in individual battery cells. By establishing a scientific and effective early warning mechanism, it is possible to achieve early perception and accurate judgment of precursory thermal runaway signals, thereby effectively preventing the escalation of accidents and improving the overall safety and reliability of energy storage systems. Summary of the Invention

[0005] In order to overcome the shortcomings of the existing technology, the purpose of this invention is to provide an early warning method for thermal runaway of battery cells based on the fusion of operating parameters and optimization of the isolated forest algorithm, which can effectively solve the problems of high thermal runaway alarm rate and delayed response of batteries in energy storage power stations.

[0006] The technical solution solved by this invention is:

[0007] A method for early thermal runaway warning of battery cells based on operational parameter fusion and isolated forest algorithm optimization includes the following steps:

[0008] Step (1): Obtain the raw voltage and temperature data collected by the energy storage power station; the raw data is the real-time voltage and temperature data of the energy storage power station.

[0009] Step (2): Preprocess the data and combine the voltage and temperature data into a two-dimensional array, with each row representing a single battery cell and each column representing the voltage and temperature respectively;

[0010] Step (3): Use the Z-Score standardization method to standardize the battery cell voltage and temperature data from step (2);

[0011] Step (4): The standardized voltage U output in step (3) i and temperature sequence T i As input, an LSTM model is built, and after LSTM time-series feature extraction, the rate of change feature ΔU is output. i ,ΔT i and prediction residual features i From the two-dimensional vector [U i T i Expanded into a five-dimensional vector [U] i T i ,ΔU i ,ΔT i ,residual i ];

[0012] Step (5): Convert the [U] output from step (4) i T i ,ΔU i ,ΔT i ,residual i The five-dimensional vector is used as input. The weight values ​​of voltage and temperature combination in the LSTM dynamic feature weighting adjustment algorithm are adjusted, the proportion of outliers is adaptively adjusted based on the prediction confidence of the LSTM model, and multiple parameters of the isolated forest algorithm are adjusted.

[0013] Step (6): Build a single-cell level battery thermal runaway anomaly detection system based on the isolated forest algorithm, using clusters as units. Multiple isolated trees are constructed by randomly selecting features and split points to detect anomalies, thereby isolating data points.

[0014] Step (7): For each battery cell data point, traverse all isolated trees to calculate a path length for each data point, calculate the anomaly score based on the average path length, and combine the LSTM residual feature value to correct the anomaly score calculation method and improve the reliability of the anomaly score.

[0015] Step (8): Parameter optimization loop finds the optimal value and runs the algorithm. The serial number of the battery cell that may have thermal runaway abnormality is identified as the output and saved. Further, alarm information is displayed based on the battery thermal runaway detection results.

[0016] This invention achieves a closed-loop system for "early perception-accurate identification-real-time early warning" of battery thermal runaway through a dual-engine approach of LSTM temporal feature enhancement and isolated forest dynamic optimization. It overcomes the lag limitation of traditional BMS threshold alarms, advancing the warning node to before the thermal runaway chain reaction begins. It solves the problem of locating abnormal cells in large-scale battery clusters, achieving efficient detection at the large-scale cell / cabinet level. It constructs an active safety protection system for energy storage power stations, providing practical technical support for intelligent operation and maintenance. Compared with existing technologies, this invention has the following advantages:

[0017] I. Accurate Extraction of Multi-Dimensional Temporal Features

[0018] By integrating dynamic features to enhance detection dimensions, the temporal correlation of voltage and temperature parameters is deeply mined through LSTM network to generate rate of change features (dV / dt, dT / dt) and residual features (|actual value - predicted value|), expanding the original two-dimensional data [voltage, temperature] into a five-dimensional feature vector [voltage, temperature, dV / dt, dT / dt, residual]; this solves the problem of single features in traditional methods and significantly improves the ability to capture weak signs of early thermal runaway.

[0019] II. Dual-algorithm collaborative optimization improves early warning accuracy

[0020] 1. Dynamic weighting mechanism optimizes feature contribution

[0021] The voltage / temperature weights are dynamically adjusted based on the residual eigenvalues ​​of LSTM to enhance voltage-sensitive signals (voltage fluctuations are more significant in the early stages of thermal runaway) and optimize the feature space partitioning boundary of isolated forests.

[0022] 2. Residual Correction Anomaly Score Algorithm

[0023] An improved formula for calculating anomaly scores in isolated forests was developed by introducing a residual threshold and a time-series persistence determination (alarms triggered after 5 consecutive anomalies), which improves the reliability of anomaly scores and significantly reduces the false alarm rate.

[0024] 3. Parameter adaptive adjustment mechanism

[0025] The contamination parameter of the isolated forest is dynamically adjusted based on the LSTM prediction confidence: when the residual remains high, it is increased to 0.1 to match the changes in the proportion of abnormalities in real time and avoid false negatives / false positives caused by fixed thresholds.

[0026] III. Project feasibility and safety assurance effectiveness

[0027] With high-precision real-time early warning capabilities, proven in practical applications, it achieves 100% detection of thermal runaway cells, with early warning response delays controlled within 1 minute, reserving a critical time window for safe handling. Furthermore, it has low computational resource dependence, integrating unsupervised learning (Isolation Forest) and lightweight LSTM feature extraction, requiring no large amount of labeled data for training, making it suitable for energy storage power stations with massive numbers of battery cells. Through Z-Score standardization, temporal consistency verification (consecutive 5 time steps for judgment), and parameter optimization loops (grid search), it effectively addresses data noise and operating condition fluctuations. Attached Figure Description

[0028] Figure 1 This is a flowchart of the method of the present invention.

[0029] Figure 2 This is a flowchart illustrating the optimization of the isolated forest algorithm in this invention.

[0030] Figure 3 This is a clustering result diagram for thermal runaway detection in Rack1, an application example of the present invention.

[0031] Figure 4 The figure shows the experimental results of battery thermal runaway detection as an application example of the present invention. Detailed Implementation

[0032] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples.

[0033] like Figure 1 As shown, this invention provides an early warning method for thermal runaway of a single battery cell based on the fusion of operating parameters and optimization using the isolated forest algorithm, comprising the following steps:

[0034] Step (1): Obtain the raw voltage and temperature data collected by the energy storage power station;

[0035] The raw data consists of the voltage and temperature data of the energy storage power station in real time. To prevent the method from being prone to false alarms due to the decrease in consistency of voltage and temperature data in certain time periods, a set of historical data points can be used for input. Specifically, the historical data points are taken from the data under the same operating conditions in the past 24 hours, with a sampling frequency of 1Hz and a number of data points of no less than 1,000, to ensure the stability and representativeness of the input data. This process is executed based on the standard data management protocol of the energy storage power station.

[0036] The battery's real-time operating parameters, such as voltage and temperature, are collected by sensors and stored in the data management system. Selecting test run data is to ensure that the test results can reflect the current state of the battery, thereby promptly detecting potential thermal runaway risks.

[0037] Step (2): Preprocess the data and combine the voltage and temperature data into a two-dimensional array, with each row representing a single battery cell and each column representing the voltage and temperature respectively;

[0038] Preprocessing includes data cleaning and formatting; the purpose of data cleaning is to remove invalid values ​​and fill in missing values ​​to ensure the integrity and accuracy of the data; formatting is to convert the data into a format that the algorithm can process.

[0039] Step (3): Use the Z-Score standardization method to standardize the battery cell voltage and temperature data from step (2);

[0040] Step (4): The standardized voltage U output in step (3) i and temperature sequence T i As input, an LSTM model is built, and after LSTM time-series feature extraction, the rate of change feature ΔU is output. i ΔT i and prediction residual features i From the two-dimensional vector [U i T i Expanded into a five-dimensional vector [U] i ,T i ,ΔU i ,ΔT i ,residual i ];

[0041] Step (5): Convert the [U] output from step (4) i T i ,ΔU i ,ΔT i ,residual i The five-dimensional vector is used as input. The weight values ​​of voltage and temperature combination in the LSTM dynamic feature weighting adjustment algorithm are adjusted, the proportion of outliers is adaptively adjusted based on the prediction confidence of the LSTM model, and multiple parameters of the isolated forest algorithm are adjusted.

[0042] Specifically, the following steps are included:

[0043] Step (5.1): The weight values ​​of the voltage and temperature combination are used to balance the contribution of voltage and temperature features in the algorithm, directly affecting the anomaly separation boundary in the feature space, and the residual features output by the LSTM model are used to balance the contribution of voltage and temperature features in the algorithm. i As a weighting adjustment signal, dynamic weights are calculated. Since voltage is more sensitive to early thermal runaway, the voltage weight is increased when the residual is large. The formula is as follows:

[0044] w u =0.5 + 0.2 × tanh(residual) i )

[0045] Among them, w u ∈[0.6, 0.7], w t =1-w u ;

[0046] Step (5.2): The number of base learners n_estimators is used to control the number of trees in the forest. Increasing the number can reduce the model variance and improve stability, but the computational cost increases linearly. The number of base learners n_estimators is configured in stages according to the sample size as follows:

[0047]

[0048] That is: n_estimators when the sample size is <10 4 When the sample size is ∈ [10, the value of n_estimators is selected in the range of 80-100; when ... 4 10 6 When the sample size is greater than 10, the value of n_estimators is selected as 100-200, satisfying the variance reduction rate ≤ 0.1%; when the sample size is greater than 10 6 The value of n_estimators is chosen to be 200-300; here the sample size ∈ [10 4 10 6 The number of base learners, n_estimators, is set to 100.

[0049] Step (5.3): The number of samples per tree, max_samples, determines the number of samples randomly selected when building each tree. Increasing this value will make the training of each tree more stable, but it will also increase the training time. Decreasing the number of training samples per tree can speed up the training, but may result in insufficient fitting of each tree, thus affecting the accuracy. Here, max_samples = 'auto' (equivalent to min(256, n_samples)) is used to control the size of the subtrees, ensuring training stability without manual adjustment, which is suitable for most balanced datasets.

[0050] Step (5.4): Contamination of the proportion of outliers in the dataset is used to adjust the sensitivity of the algorithm. It controls what proportion of outliers the algorithm should detect. Contamination is adjusted based on the prediction confidence of the LSTM model. The adjustment formula is shown below:

[0051]

[0052] If the residuals at consecutive time steps are high, the contamination is temporarily increased to 0.1. θ is the residual threshold. The mean residual predicted by the LSTM model is 0.2±0.05, so θ can be initially set to 0.3. When an early failure occurs, the mean residual will rise to above 0.6, thus triggering the contamination to be increased.

[0053] Step (5.5): Control the randomness of random_state. This parameter is used to control the random process in model training, so that the experiment is reproducible. Here, 42 is selected as the value of random_state, so that the same random number sequence is generated each time it is run, thereby ensuring the consistency of the random process such as model training and data partitioning, and ensuring that the generated tree is the same each time.

[0054] Step (6): Build a single-cell level battery thermal runaway anomaly detection system based on the isolated forest algorithm, using clusters as units. Multiple isolated trees are constructed by randomly selecting features and split points to detect anomalies, thereby isolating data points.

[0055] Step (7): For each battery cell data point, traverse all isolated trees to calculate a path length for each data point, calculate the anomaly score based on the average path length, and combine the LSTM residual feature value to correct the anomaly score calculation method and improve the reliability of the anomaly score.

[0056] Step (8): Parameter optimization loop finds the optimal value and runs the algorithm. The serial number of the battery cell that may be thermal runaway is identified as the output and saved. Further, alarm information is displayed based on the battery thermal runaway detection results.

[0057] The standardization process in step (3) specifically involves:

[0058] The battery cell voltage and temperature data at the same time are standardized. The purpose of standardization is to eliminate the impact of differences in units and numerical ranges on algorithm performance, so that different features can be compared on the same scale. Standardized data has zero mean and unit variance, thereby improving the stability and accuracy of the algorithm. The formula for Z-Score standardization of voltage and temperature data is as follows:

[0059] Voltage standardization:

[0060] Temperature standardization:

[0061] Where: u i and t i Let μ be the voltage and temperature parameters of the i-th battery cell at the same time. uand μ t Let σ be the mean of the voltage and temperature distributions, respectively. u and σ t These are the standard deviations of the voltage and temperature distributions, respectively.

[0062] The input in step (4) is the standardized {U} i T i The data sequence is expressed by the formula ΔU. i =U i (t)-U i (t-1) Calculate the voltage change rate using the formula ΔT i =T i (t)-T i (t-1) Calculate the rate of temperature change, and use the LSTM model to predict U at the next time step. i and T i Calculate the residual i When an anomaly occurs, the residual increases, resulting in an enhanced feature matrix output from two dimensions [U]. i T i Expanded to five dimensions [U] i ,T i ,ΔU i ,ΔT i ,residual i ].

[0063] Specifically, the data standardization matrix {U} output in step (3) is as follows: i T i Using this as input, an LSTM model is built. After LSTM time-series feature extraction, the rate of change feature is calculated using the following formula:

[0064] Voltage change rate: ΔU i =U i (t)-U i (t-1)

[0065] Temperature change rate: ΔT i =T i (t)-T i (t-1)

[0066] LSTM model predicts U at the next time step i and T i Calculate the residual i = | Actual value - Predicted value |, residuals increase during anomalies; output enhanced feature matrix, from two dimensions [U i T i Expanded into a five-dimensional vector [U] i T i ,ΔU i ,ΔTi ,residual i ].

[0067] LSTM model construction, feature extraction, and prediction are well-known techniques. Specifically, a standard LSTM architecture is adopted: the input layer has a dimension of 2 (corresponding to Ui and Ti), the hidden layer has 64 units, and the output layer has a dimension of 2 (predicted Ui(t+1) and Ti(t+1)). During training, the Adam optimizer is used with a learning rate of 0.001, fitting time-series patterns based on historical data. The residual is calculated as `residuali = |actual value - predicted value|`, with the residual increasing in cases of anomalies. This process follows the standard implementations of PyTorch and TensorFlow.

[0068] In step (6), the five-dimensional vector [U] output in step (3) is... i T i ,ΔU i ,ΔT i ,residual i As the input space for the optimized isolated forest, starting from the data space, a feature dimension is randomly selected and random split points are generated. Isolated tree nodes are constructed through iterative partitioning. During the binary splitting process, normal points require multiple splits, resulting in longer path lengths, while outliers are quickly isolated, resulting in shorter path lengths. Anomaly scores are calculated based on the path length differences (the shorter the path, the higher the score). When the score is below a set threshold, it is marked as a normal point; when it is above the threshold, it is determined to be an outlier. This process quantifies the isolation difficulty of data points through a random splitting strategy, achieving efficient unsupervised anomaly identification. The specific steps are as follows:

[0069] Step (6.1): From the standardized feature set {U i T i Randomly select a feature from};

[0070] Step (6.2): ​​On the selected feature, randomly select a segmentation value, which is randomly selected between the minimum and maximum values ​​of the feature;

[0071] Step (6.3): Based on the selected features and segmentation values, divide the data into two parts: one part is less than the segmentation value, and the other part is greater than or equal to the segmentation value;

[0072] Step (6.4): Recursively construct subtrees for the two parts of data after splitting until each leaf node contains only one data point or reaches the preset depth limit;

[0073] Step (6.5): Calculate the anomaly score s(x) based on the path length and set the threshold to 0.85: if s(x) ≥ 0.85, it is determined to be an anomaly point; otherwise, it is a normal point.

[0074] In step (7), for each battery data point x, the path length h(x) in the n trees is recorded. The anomaly score is calculated based on the average of the path lengths. The specific steps are as follows:

[0075] Step (7.1): Let h(x) represent the path length of data point x in an isolated tree, and let H(x) represent the average path length of data point x in all isolated trees, then we have:

[0076]

[0077] Where T is the number of isolated trees, h t (x) is the path length of data point x in the t-th isolated tree;

[0078] Step (7.2): The anomaly score s(x) is defined as:

[0079]

[0080] Where E(H(x)) is the expected value of the path length, and c(n) is the expected value of the average path length;

[0081] Step (7.3): The built-in outlier score algorithm does not take into account time persistence. Therefore, the outlier score s(x) calculation method is corrected by combining LSTM residual feature values:

[0082]

[0083] Where I is the indicator function, θ is the residual threshold, and T is the time window size;

[0084] Step (7.4): c(n) is the expected value of the average path length, calculated using the following formula:

[0085]

[0086] Step (7.5): H(i) is the harmonic number, which can be approximated using Euler's constant:

[0087] H(i)≈ln(i)+γ

[0088] Where γ≈0.57721566 is Euler's constant.

[0089] In step (8), parameter optimization loops are performed, repeating steps (5) to (7) until the parameters are optimal. After the optimized isolated forest algorithm completes anomaly detection, the results are further analyzed. The entire algorithm optimization process is as follows: Figure 2As shown, firstly, a five-dimensional feature vector is input and dynamically weighted. The voltage and temperature weights are adjusted according to the residual feature values. Then, the contamination parameter is adaptively adjusted: it remains at 0.005 when the residual is normal and is increased to 0.01 when it is consistently high. Next, an isolated tree is constructed by random feature value segmentation, and the data is recursively segmented to calculate the path length and the original anomaly score. After time-series consistency correction and residual feature value optimization, the parameter optimization loop is entered to determine whether the optimal parameters have been reached. If not, grid search optimization is performed. Then, the isolated forest algorithm is used to optimize and fuse multi-dimensional features for anomaly detection until the final structured result matrix is ​​output. Battery cells with anomaly scores ≥0.85 are marked in red for warning. If the anomaly marking alarm is triggered 5 times consecutively, it is determined that the battery cell has experienced thermal runaway.

[0090] The process combines an anomaly score calculation method based on LSTM residual eigenvalues ​​with a dynamic feature weighting mechanism to optimize the isolated forest algorithm and perform anomaly detection, achieving accurate early thermal runaway identification and providing technical support for energy storage system safety protection. Specifically, for cells marked as potentially abnormal, if abnormal behavior is detected in five consecutive time steps, it can be determined that the cell may have experienced thermal runaway. At this point, the program automatically records the serial numbers of these abnormal cells and generates corresponding alarm information.

[0091] To test the accuracy of this method, the present invention was validated using operating data from a cabinet of an energy storage power station on July 29, 2024. The battery cabinet structure consists of 12 parallel battery clusters per cabinet, with each cluster containing 364 series-connected battery cells, resulting in 4368 battery cells per cabinet. The energy storage power station performs a constant power charge-discharge cycle once daily. During the validation process, faults were artificially injected into the experimental data to simulate the data distribution when a battery cell experiences thermal runaway. The voltage curves of cells 1 and 101 in each battery cluster were artificially lowered to simulate thermal runaway. The voltage and temperature data were Z-score standardized, and the rate of change feature ΔU was extracted using LSTM time-series feature extraction. i ,ΔT i and prediction residual features i The optimized isolated forest model is combined with random segmentation of the feature space to quickly isolate outliers.

[0092] Algorithm configuration:

[0093]

[0094] After the Isolation Forest algorithm optimizes the detection, the first single instance is located at the edge of the feature space (U i <-2.5 and T i>2.0), average path Normal single-unit path length Clustering results as follows Figure 3 As shown: Cell 1 (black) is significantly isolated. Through continuous verification and alarm, Cell 1 triggered s(x) ≥ 0.85 five times consecutively in time steps 20-25. For Rack1 data, the algorithm detected data anomalies in Cell 1 starting from time step 20, which is only 10 data points (50 seconds) away from when the cell began to show signs of thermal runaway in time step 11. The detection algorithm performed similarly in the other 11 clusters, successfully detecting abnormal cells. Furthermore, the time between the algorithm's first detection of an abnormal cell and the start of the battery temperature increase was only about 50 to 60 seconds, meeting the real-time requirements of energy storage power stations for timely detection of thermal runaway abnormal cells. The system output anomaly sequence is as follows: Figure 4 As shown, each potentially thermally runaway anomalous cell is identified and output.

[0095] The experiment verified the effectiveness of the method in detecting thermal runaway on 12 clusters. The experimental results showed that the accuracy of the method reached 100%, that is, all the monomers that the detection algorithm predicted as potentially abnormal exhibited thermal runaway phenomena; that is, the algorithm detected all monomers that exhibited thermal runaway phenomena in the experiment.

[0096] In summary, this invention proposes an early warning method for battery cell thermal runaway based on operational parameter fusion and the optimization of the isolated forest algorithm. By simultaneously monitoring the voltage and temperature changes of battery cells, the method can more comprehensively capture changes in the internal state of the battery, thereby improving the accuracy of thermal runaway detection. Furthermore, by utilizing an improved isolated forest algorithm for anomaly detection, it can quickly identify outliers in high-dimensional data without requiring a large amount of training data. This method not only effectively overcomes the difficulty of threshold setting in traditional threshold comparison methods but also provides early warning in the early stages of battery thermal runaway, thus providing strong protection for the safe and stable operation of energy storage systems.

Claims

1. A method for early warning of thermal runaway in a single battery cell based on fusion of operating parameters and optimization using the isolated forest algorithm, characterized in that, Includes the following steps: Step (1): Obtain the raw voltage and temperature data collected by the energy storage power station; the raw data is the real-time voltage and temperature data of the energy storage power station. Step (2): Preprocess the data and combine the voltage and temperature data into a two-dimensional array, with each row representing a single battery cell and each column representing the voltage and temperature respectively; Step (3): Use the Z-Score standardization method to standardize the battery cell voltage and temperature data from step (2); Step (4): The normalized voltage U output in step (3) i and temperature sequence T i As input, an LSTM model is built, and after LSTM time-series feature extraction, the rate of change feature ΔU is output. i ΔT i and prediction residual features i From [U i T i Expanded to [U] i T i ΔU i ΔT i ,residual i ]; Step (5): Take the [U] output from step (4) i T i ΔU i ΔT i ,residual i As input, the weight values ​​of voltage and temperature combination in the LSTM dynamic feature weighted adjustment algorithm and the proportion of outliers adaptively adjusted based on the prediction confidence of the LSTM model are used to adjust multiple parameters of the isolated forest algorithm. Step (6): Build a battery thermal runaway anomaly detection system based on the isolated forest algorithm at the individual level, using clusters as units. Multiple isolated trees are constructed by randomly selecting features and split points to detect anomalies, thereby isolating data points. Step (7): For each battery cell data point, traverse all isolated trees to calculate a path length for each data point, calculate the anomaly score based on the average path length, and combine the LSTM residual feature value to correct the anomaly score calculation method and improve the reliability of the anomaly score. Step (8): Parameter optimization loop finds the optimal value and runs the algorithm. The serial number of the battery cell that may have thermal runaway abnormality is identified and saved as output. Alarm information is displayed based on the battery thermal runaway detection results.

2. The method for early thermal runaway warning of a single battery cell based on fusion of operating parameters and optimization by the isolated forest algorithm as described in claim 1, characterized in that, The standardization process in step (3) specifically involves: The battery cell voltage and temperature data at the same time are standardized; the formula for Z-Score standardization of voltage and temperature data is as follows: Voltage standardization: Temperature standardization: in: and For the same moment The voltage and temperature parameters of each individual battery cell. and These are the mean values ​​of voltage and temperature distributions, respectively. and These are the standard deviations of the voltage and temperature distributions, respectively.

3. The method for early thermal runaway warning of a single battery cell based on operational parameter fusion and isolated forest algorithm optimization according to claim 1, characterized in that, The input in step (4) is the standardized {U} i T i The data sequence is expressed by the formula ΔU. i =U i (t)-U i (t-1) Calculate the voltage change rate using the formula ΔT i =T i (t)-T i (t-1) Calculate the rate of temperature change, and use the LSTM model to predict U at the next time step. i and T i Calculate the residual i When an anomaly occurs, the residual increases, resulting in an enhanced feature matrix output from [U i T i Expanded to [U] i T i ΔU i ΔT i ,residual i ].

4. The method for early thermal runaway warning of a single battery cell based on fusion of operating parameters and optimization by the isolated forest algorithm as described in claim 1, characterized in that, In step (6), the output of step (4) is [U i T i ΔU i ΔT i ,residual i As the input space for the optimized isolated forest, starting from the data space, a feature dimension is randomly selected and random split points are generated. Isolated tree nodes are constructed through iterative partitioning. During the binary splitting process, normal points require multiple splits, resulting in longer path lengths, while outliers are quickly isolated, resulting in shorter path lengths. Anomaly scores are calculated based on the path length differences. Points below a set threshold are marked as normal points, while those above the threshold are identified as outliers. This process quantifies the isolation difficulty of data points through a random splitting strategy, achieving efficient unsupervised anomaly identification. The specific steps are as follows: Step (6.1): From the output of step (4), [U i T i ΔU i ΔT i ,residual i Randomly select a feature from [ ]; Step (6.2): ​​On the selected feature, randomly select a segmentation value, which is randomly selected between the minimum and maximum values ​​of the feature; Step (6.3): Based on the selected features and segmentation values, divide the data into two parts: one part is less than the segmentation value, and the other part is greater than or equal to the segmentation value; Step (6.4): Recursively construct subtrees for the two parts of data after splitting until each leaf node contains only one data point or reaches the preset depth limit; Step (6.5): Calculate the anomaly score s(x) based on the path length and set the threshold to 0.85: if s(x) ≥ 0.85, it is determined to be an anomaly point; otherwise, it is a normal point.

5. The method for early thermal runaway warning of a single battery cell based on fusion of operating parameters and optimization by the isolated forest algorithm according to claim 1, characterized in that, In step (7), for each battery data point x, the path length h(x) in the n trees is recorded. The anomaly score is calculated based on the average of the path lengths. The specific steps are as follows: Step (7.1): Let h(x) represent the path length of data point x in an isolated tree, and let H(x) represent the average path length of data point x in all isolated trees, then we have: Where T is the number of isolated trees. It is the path length of data point x in the t-th isolated tree; Step (7.2): The anomaly score s(x) is defined as: in, It is the expected value of the path length. It is the expected value of the average path length; Step (7.3): The built-in outlier score algorithm does not take into account time persistence. By combining LSTM residual eigenvalues, the outlier score s(x) calculation method is corrected: in, Here, θ is the indicator function, θ is the residual threshold, and T is the time window size; Step (7.4): It is the expected value of the average path length, calculated using the following formula: Step (7.5): The harmonic number can be approximated using Euler's constant: in, It is Euler's constant.

6. The method for early thermal runaway warning of a single battery cell based on operational parameter fusion and isolated forest algorithm optimization according to claim 1, characterized in that, In step (8), a parameter optimization loop is performed, and steps (5) to (7) are repeated until the parameters are optimal. After the optimized isolated forest algorithm completes the anomaly detection, the results are further analyzed. First, the feature vector is input and dynamic feature weighting is performed. The voltage and temperature weights are adjusted according to the residual feature values. Then, the contamination parameter is adaptively adjusted: when the residual is normal, it is kept at 0.005, and when it is continuously high, it is increased to 0.

01. Next, an isolated tree is constructed by random feature value segmentation, and the data is recursively segmented to calculate the path length and the original anomaly score. After time-series consistency correction and residual feature value optimization, the parameter optimization loop is entered to determine whether the optimal parameters have been reached. If the target is not met, a grid search optimization is performed, followed by optimization and fusion of multi-dimensional features using the isolated forest algorithm for anomaly detection until the final structured result matrix is ​​output. Battery cells with anomaly scores ≥0.85 are marked in red for warning. If the anomaly alarm is triggered 5 times consecutively, the battery cell is determined to have experienced thermal runaway.

Images