Energy storage fire control method and system based on deep learning and machine learning fusion

Abstract

The invention discloses an energy storage fire control method based on deep learning and machine learning fusion. The method comprises the following steps: obtaining historical data of a battery management system; preprocessing the obtained data to obtain feature vectors, and labeling the feature vectors to obtain a training data set; based on the LSTM network model, constructing an energy storage fire control initial model; training the energy storage fire control initial model by using the training data set to obtain an energy storage fire control model; according to the energy storage fire-fighting control model, actual energy storage system state detection is carried out, and control over the energy storage fire-fighting system is completed. The invention also discloses a system for realizing the deep learning and machine learning fusion-based energy storage fire control method. According to the method, the strategy of dynamically adjusting the K value based on the prediction result is provided, then the result is checked based on the KNN algorithm, the accuracy of false alarm recognition is effectively improved, and the false alarm rate of the energy storage fire fighting system is reduced.

Description

Technical Field

[0001] The present invention belongs to the field of energy storage fire protection technology, and specifically relates to an energy storage fire protection control method and system based on the fusion of deep learning and machine learning. Background Art

[0002] As a key branch of the new energy sector, energy storage technology is gaining increasing attention for its safety as it is widely adopted across various industries. As a crucial component of ensuring the safety of energy storage facilities, the design, construction, and management of energy storage fire protection systems are crucial for preventing and reducing fire accidents within these facilities.

[0003] Existing energy storage fire protection systems use sensors installed on the battery pack (such as temperature and CO sensors) to monitor battery status in real time. These sensors, along with data collected by the battery management system (BMS), such as cell voltage, temperature, SOC (State of Charge), and depth of discharge, are used to identify fire hazards. However, due to factors such as environmental interference, equipment aging, and data drift, detectors often generate false alarms, increasing operational costs and potentially causing unnecessary downtime, impacting the normal operation of the energy storage system and the stability of its power supply. Summary of the Invention

[0004] One of the objectives of the present invention is to provide a storage fire control method based on the fusion of deep learning and machine learning, which can effectively solve the problem of suppressing false alarms and improve the accuracy of early warnings.

[0005] A second object of the present invention is to provide a system for implementing the energy storage fire control method based on the fusion of deep learning and machine learning.

[0006] The present invention provides a storage fire control method based on the fusion of deep learning and machine learning, comprising the following steps:

[0007] S1. Obtain historical data of the battery management system;

[0008] S2. Preprocess the data obtained in step S1 to obtain a feature vector, and label the feature vector to obtain a training data set;

[0009] S3. Build an initial energy storage fire control model based on the LSTM network model;

[0010] S4. Using the training data set obtained in step S2, the initial energy storage fire control model obtained in step S3 is trained to obtain an energy storage fire control model;

[0011] S5. Based on the energy storage fire control model obtained in step S4, actual energy storage system status monitoring is performed to complete the control of the energy storage fire protection system.

[0012] The data in step S1 include the voltage, temperature, state of charge, depth of discharge data of the single battery, and the temperature and CO concentration data of the battery pack.

[0013] Step S2 specifically includes: filtering the data obtained in step S1, extracting key features and constructing a feature vector. The elements in the feature vector include the average voltage of the single cell, the average temperature of the single cell, the state of charge, the average depth of discharge, the current temperature of the battery pack, and the CO concentration;

[0014] The feature vectors are then labeled to obtain a feature vector set with stable state and fire hazard labels as a training data set.

[0015] The energy storage fire control initial model in step S3 includes an input layer, an LSTM layer, a fully connected layer, and an output layer; the input layer receives the preprocessed feature vector of the input and outputs it to the LSTM layer; the LSTM layer includes several LSTM units, which use the input data to learn the dependency relationship in the time series data and output the learned features to the fully connected layer; the fully connected layer integrates the input features and outputs the results to the output layer; the output layer includes a stable state branch and a fire hazard branch;

[0016] The stable state branch in the output layer processes the input data to obtain the stable state probability predicted by the model, which is expressed using the following formula:

[0017] P s =σ(W s ·h T +b s )

[0018] Among them, P s is the stable state probability; σ is the sigmoid function; W s is the stable state branch weight; h T is the hidden state of the last time step of the LSTM layer; b s is the steady-state branch bias parameter;

[0019] The fire hazard branch in the output layer processes the input data to obtain the fire hazard probability predicted by the model, which is expressed using the following formula:

[0020] P f =σ(W f ·h T +b f )

[0021] Among them, P f is the probability of fire hazard; σ is the sigmoid function; W f is the fire hazard branch weight; h Tis the hidden state of the last time step of the LSTM layer; b f is the fire hazard branch bias parameter;

[0022] The stable state probability and fire hazard probability obtained by the output layer are used as the final output of the model.

[0023] Step S4 specifically includes: using the training data set obtained in step S2 to train the initial energy storage fire control model obtained in step S3, optimizing the model parameters through the back propagation algorithm, and obtaining the energy storage fire control model. The total loss function is expressed as follows:

[0024] L=[yslog(Ps)+(1-ys)log(1-Ps)]-[yflog(Pf)+(1-yf)log(1-Pf)]

[0025] Among them, L is the total loss function; y s is the stable state label; y f Fire hazard label.

[0026] Step S5 includes the following steps:

[0027] a. Obtain battery management system data, including single cell voltage, temperature, state of charge, depth of discharge data, as well as measured temperature and CO concentration data;

[0028] b. Preprocess the data obtained in step a to obtain a feature vector;

[0029] c. Input the characteristic vector obtained in step b into the energy storage fire control model obtained in step S4 to obtain the steady state probability P s ' and fire hazard probability P f ';

[0030] d. Based on the steady state probability P s Adjust the K value in the KNN algorithm and use the KNN algorithm to check for false alarms. If it is a true alarm, the alarm will be issued normally. Otherwise, it will be considered a false alarm and no alarm will be issued.

[0031] The step d is based on the steady state probability P s 'Adjust the K value in the KNN algorithm. The adjustment strategy is: when P s '>0.7, K value is set to 7; when P s When '≤0.7, the K value is set to 3.

[0032] In step d, when the steady state probability P s '>0.7 and the fire hazard probability P f '<0.5, the battery management system is judged to be in normal operation, otherwise it enters the troubleshooting step.

[0033] In step d, the KNN algorithm is used to perform false positive detection and investigation, including primary and secondary investigations.

[0034] The first-level troubleshooting is as follows: First, obtain the battery management system data at the current time t and construct a new sample feature vector. The elements of the new sample feature vector include the current temperature of the battery pack, CO concentration, average cell voltage, average cell temperature, state of charge, and average depth of discharge. The average cell voltage is the arithmetic mean of all cell voltages within a preset time window; the average cell temperature is the arithmetic mean of all cell temperatures within a preset time window.

[0035] Calculate the distance between the new sample feature vector and the feature vectors of all samples in the training dataset Use the following formula to express it:

[0036]

[0037] Among them, the new sample feature vector is S i =[x i1 ,x i2 ,...,x i6 ]; the jth sample in the training data set is x i1 ,x i2 ,...,x i6 They correspond to the current temperature, CO concentration, average voltage of single cells, average temperature of single cells, state of charge, and average depth of discharge of the battery pack in the new sample feature vector respectively;

[0038] Sort the obtained distances from smallest to largest, and select the K samples with the closest distance as the first-level neighbors N of the new sample, where the value of K is the preset value adjusted in step d;

[0039] Calculate the anomaly score Q1, count the number of normal state samples and false positive state samples in the first-level neighborhood, and use the following formula to express it:

[0040]

[0041] Where I is the indicator function, when the nearest neighbor is a false alarm, I = 1, otherwise it is 0;

[0042] If the anomaly score Q1 exceeds the set threshold, it is considered a true alarm and the alarm is issued normally; otherwise, it is marked as a suspected false alarm and enters the second level of investigation;

[0043] The second-level investigation specifically involves refining the feature vectors of samples marked as suspected false positives in the first-level investigation. The sample feature vectors are [current battery pack temperature, CO concentration, average cell voltage, average cell temperature, state of charge, average depth of discharge, temperature change rate over the past hour, CO concentration change rate over the past hour, difference between maximum and minimum cell temperatures, difference between maximum and minimum cell voltages, ambient temperature, and ambient humidity].

[0044] Calculate the Euclidean distance between the suspected false positive sample and all samples in the training dataset Use the following formula to express it:

[0045]

[0046] Among them, the characteristic vector of the suspected false positive sample is The jth sample in the training data set is The corresponding false positive sample feature vectors are the current temperature of the battery pack, CO concentration, average voltage of the single cell, average temperature of the single cell, state of charge, average depth of discharge, temperature change rate in the past hour, CO concentration change rate in the past hour, difference between the maximum and minimum temperatures of the single cell, difference between the highest and lowest voltages of the single cell, ambient temperature, and ambient humidity.

[0047] Sort the obtained distances from smallest to largest, and select the K samples with the closest distance as the secondary neighbors N' of the suspected false positive sample, where the value of K is the preset value adjusted in step d;

[0048] Calculate the anomaly score Q2, count the number of normal state samples and false positive state samples in the secondary neighborhood, and use the following formula to express it:

[0049]

[0050] Where I is the indicator function, when the nearest neighbor is a false alarm, I = 1, otherwise it is 0;

[0051] If the fire hazard probability P of the suspected false alarm sample f 'If the value is lower than the set threshold and the anomaly score Q2 is lower than the set threshold, the suspected false alarm sample is judged to be a false alarm and no alarm is issued. Otherwise, it is judged to be a true alarm and an alarm is issued normally.

[0052] The present invention also provides a system for implementing the energy storage fire control method based on the fusion of deep learning and machine learning, comprising a data acquisition module, a data processing module, a model construction module, a model training module, and an energy storage system fire control module;

[0053] The data acquisition module obtains the historical data of the battery management system and uploads the data to the data processing module;

[0054] The data processing module preprocesses the received data to obtain feature vectors, labels the feature vectors, obtains the training data set, and uploads the data to the model training module;

[0055] The model building module builds the initial model of energy storage fire control based on the LSTM network model and uploads the data to the model training module;

[0056] The model training module uses the training data set to train the initial energy storage fire control model based on the received data, obtains the energy storage fire control model, and uploads the data to the energy storage system fire control module;

[0057] The energy storage system fire control module uses the energy storage fire control model based on the received data to conduct actual energy storage system status monitoring and complete the control of the energy storage fire control system.

[0058] This invention discloses a method and system for energy storage fire protection control based on the fusion of deep learning and machine learning. The model incorporates dual output branches to predict the battery's stable state and fire hazard, respectively, increasing the model's predictive dimension. The method proposes a strategy for dynamically adjusting the K value based on the prediction results, and then uses the KNN algorithm to screen the results. This effectively improves the accuracy of false alarm identification and reduces the false alarm rate of the energy storage fire protection system. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 Schematic diagram of the process of the present invention;

[0060] Figure 2 Schematic diagram of the structure of the system of the present invention. DETAILED DESCRIPTION

[0061] The present invention provides a storage fire control method based on the fusion of deep learning and machine learning, the flow chart of which is as follows: Figure 1 As shown, the following steps are included:

[0062] S1. Obtain historical data of the battery management system;

[0063] The data in step S1 include the voltage, temperature, state of charge, depth of discharge data of the single battery, and the temperature and CO concentration data of the battery pack.

[0064] S2. Preprocess the data obtained in step S1 to obtain feature vectors, and label the feature vectors to obtain a training data set, specifically:

[0065] Filter the data obtained in step S1 to extract key features and construct a feature vector. The elements in the feature vector include the average voltage of the single cell, the average temperature of the single cell, the state of charge, the average depth of discharge, the current temperature of the battery pack, and the CO concentration.

[0066] The feature vectors are then labeled to obtain a feature vector set with stable state and fire hazard labels as a training data set.

[0067] S3. Build an initial energy storage fire control model based on the LSTM network model;

[0068] The energy storage fire control initial model includes an input layer, an LSTM layer, a fully connected layer, and an output layer. The input layer receives the pre-processed feature vector of the input and outputs it to the LSTM layer. The LSTM layer includes several LSTM units, which use the input data to learn the dependency relationship in the time series data and output the learned features to the fully connected layer. The fully connected layer integrates the input features and outputs the results to the output layer. The output layer includes a stable state branch and a fire hazard branch.

[0069] The stable state branch in the output layer processes the input data to obtain the stable state probability predicted by the model, which is expressed using the following formula:

[0070] P s =σ(W s ·h T +b s )

[0071] Among them, P s is the stable state probability; σ is the sigmoid function; W s is the stable state branch weight; h T is the hidden state of the last time step of the LSTM layer; b s is the steady-state branch bias parameter;

[0072] The fire hazard branch in the output layer processes the input data to obtain the fire hazard probability predicted by the model, which is expressed using the following formula:

[0073] P f =σ(W f ·h T +b f )

[0074] Among them, P f is the probability of fire hazard; σ is the sigmoid function; W f is the fire hazard branch weight; h T is the hidden state of the last time step of the LSTM layer; b fis the fire hazard branch bias parameter;

[0075] The stable state probability and fire hazard probability obtained by the output layer are used as the final output of the model.

[0076] S4. Using the training data set obtained in step S2, the initial energy storage fire control model obtained in step S3 is trained to obtain an energy storage fire control model, specifically:

[0077] The training data set obtained in step S2 is used to train the initial energy storage fire control model obtained in step S3. The model parameters are optimized by the back propagation algorithm to obtain the energy storage fire control model. The total loss function is expressed as follows:

[0078] L=[yslog(Ps)+(1-ys)log(1-Ps)]-[yflog(Pf)+(1-yf)log(1-Pf)]

[0079] Among them, L is the total loss function; y s is the stable state label; y f Fire hazard label.

[0080] S5. Based on the energy storage fire control model obtained in step S4, actual energy storage system status monitoring is performed to complete the control of the energy storage fire protection system, including the following steps:

[0081] a. Obtain battery management system data, including single cell voltage, temperature, state of charge, depth of discharge data, as well as measured temperature and CO concentration data;

[0082] b. Preprocess the data obtained in step a to obtain a feature vector;

[0083] c. Input the characteristic vector obtained in step b into the energy storage fire control model obtained in step S4 to obtain the steady state probability P s ' and fire hazard probability P f ';

[0084] d. Based on the steady state probability P s Adjust the K value in the KNN algorithm and use the KNN algorithm to check for false alarms. If it is a true alarm, the alarm will be issued normally. Otherwise, it will be considered a false alarm and no alarm will be issued.

[0085] The step d is based on the steady state probability P s 'Adjust the K value in the KNN algorithm. The adjustment strategy is: when P s '>0.7, K value is set to 7; when P s When '≤0.7, the K value is set to 3.

[0086] In step d, when the steady state probability P s '>0.7 and the fire hazard probability P f '<0.5, the battery management system is judged to be in normal operation, otherwise it enters the troubleshooting step.

[0087] In step d, the KNN algorithm is used to perform false positive detection and investigation, including primary and secondary investigations.

[0088] The first-level troubleshooting is as follows: First, obtain the battery management system data at the current time t and construct a new sample feature vector. The elements of the new sample feature vector include the current temperature of the battery pack, CO concentration, average cell voltage, average cell temperature, state of charge, and average depth of discharge;

[0089] Calculate the distance between the new sample feature vector and the feature vectors of all samples in the training dataset Use the following formula to express it:

[0090]

[0091] Among them, the new sample feature vector is S i =[x i1 ,x i2 ,...,x i6 ]; the jth sample in the training data set is x i1 ,x i2 ,...,x i6 They correspond to the current temperature, CO concentration, average cell voltage, average cell temperature, state of charge, and average depth of discharge of the battery pack in the new sample feature vector respectively; the average cell voltage is the arithmetic mean of all cell voltages within the preset time window; the average cell temperature is the arithmetic mean of all cell temperatures within the preset time window;

[0092] Sort the obtained distances from smallest to largest, and select the K samples with the closest distance as the first-level neighbors N of the new sample, where the value of K is the preset value adjusted in step d;

[0093] Calculate the anomaly score Q1, count the number of normal state samples and false positive state samples in the first-level neighborhood, and use the following formula to express it:

[0094]

[0095] Where I is the indicator function, when the nearest neighbor is a false alarm, I = 1, otherwise it is 0;

[0096] If the anomaly score Q1 exceeds the set threshold of 0.8, it is judged as a true alarm and the alarm is issued normally; otherwise, it is marked as a suspected false alarm and enters the second level of investigation;

[0097] The second-level investigation specifically involves refining the feature vectors of samples marked as suspected false positives in the first-level investigation. The sample feature vectors are [current battery pack temperature, CO concentration, average cell voltage, average cell temperature, state of charge, average depth of discharge, temperature change rate over the past hour, CO concentration change rate over the past hour, difference between maximum and minimum cell temperatures, difference between maximum and minimum cell voltages, ambient temperature, and ambient humidity].

[0098] Calculate the Euclidean distance between the suspected false positive sample and all samples in the training dataset Use the following formula to express it:

[0099]

[0100] Among them, the characteristic vector of the suspected false positive sample is The jth sample in the training data set is The corresponding false positive sample feature vectors are the current temperature of the battery pack, CO concentration, average voltage of the single cell, average temperature of the single cell, state of charge, average depth of discharge, temperature change rate in the past hour, CO concentration change rate in the past hour, difference between the maximum and minimum temperatures of the single cell, difference between the highest and lowest voltages of the single cell, ambient temperature, and ambient humidity.

[0101] Sort the obtained distances from smallest to largest, and select the K samples with the closest distance as the secondary neighbors N' of the suspected false positive sample, where the value of K is the preset value adjusted in step d;

[0102] Calculate the anomaly score Q2, count the number of normal state samples and false positive state samples in the secondary neighborhood, and use the following formula to express it:

[0103]

[0104] Where I is the indicator function, when the nearest neighbor is a false alarm, I = 1, otherwise it is 0;

[0105] If the fire hazard probability P of the suspected false alarm sample f If the value is lower than the set threshold of 0.5 and the anomaly score Q2 is lower than the set threshold of 0.8, the suspected false alarm sample is judged to be a false alarm and no alarm is issued. Otherwise, it is judged to be a true alarm and an alarm is issued normally.

[0106] The present invention also provides a system for implementing the energy storage fire control method based on the fusion of deep learning and machine learning, comprising a data acquisition module, a data processing module, a model construction module, a model training module, and an energy storage system fire control module;

[0107] The data acquisition module obtains the historical data of the battery management system and uploads the data to the data processing module;

[0108] The data processing module preprocesses the received data to obtain feature vectors, labels the feature vectors, obtains the training data set, and uploads the data to the model training module;

[0109] The model building module builds the initial model of energy storage fire control based on the LSTM network model and uploads the data to the model training module;

[0110] The model training module uses the training data set to train the initial energy storage fire control model based on the received data, obtains the energy storage fire control model, and uploads the data to the energy storage system fire control module;

[0111] The energy storage system fire control module uses the energy storage fire control model based on the received data to conduct actual energy storage system status monitoring and complete the control of the energy storage fire control system.

Claims

1. A storage fire control method based on the fusion of deep learning and machine learning, characterized in that: The following steps are involved: S1. Obtain historical data of the battery management system; S2. Preprocess the data obtained in step S1 to obtain a feature vector, and label the feature vector to obtain a training data set; S3. Based on the LSTM network model, build the initial model of energy storage fire control; S4. Using the training data set obtained in step S2, the energy storage fire control initial model obtained in step S3 is trained to obtain an energy storage fire control model; S5. According to the energy storage fire control model obtained in step S4, actual energy storage system status monitoring is performed to complete the control of the energy storage fire protection system.

2. The energy storage fire control method based on the fusion of deep learning and machine learning according to claim 1 is characterized in that: The data in step S1 include single cell voltage, temperature, state of charge, depth of discharge data, and battery pack temperature and CO concentration data.

3. The energy storage fire control method based on the fusion of deep learning and machine learning according to claim 2 is characterized in that: Step S2 specifically includes: filtering the data obtained in step S1, extracting key features and constructing a feature vector, wherein the elements of the feature vector include the average voltage of the single cell, the average temperature of the single cell, the state of charge, the average depth of discharge, the current temperature of the battery pack, and the CO concentration; The feature vectors are then labeled to obtain a feature vector set with stable state and fire hazard labels as a training data set.

4. The energy storage fire control method based on the fusion of deep learning and machine learning according to claim 3 is characterized in that: The energy storage fire control initial model in step S3 includes an input layer, an LSTM layer, a fully connected layer, and an output layer; the input layer receives the preprocessed feature vector of the input and outputs it to the LSTM layer; the LSTM layer includes a number of LSTM units, which use the input data to learn the dependency relationship in the time series data and output the learned features to the fully connected layer; the fully connected layer integrates the input features and outputs the results to the output layer; the output layer includes a stable state branch and a fire hazard branch; The stable state branch in the output layer processes the input data to obtain the stable state probability predicted by the model, which is expressed by the following formula: P s =σ(W s ·h T +b s ) Among them, P s is the stable state probability; σ is the sigmoid function; W s is the stable state branch weight; h T is the hidden state of the last time step of the LSTM layer; b s is the steady-state branch bias parameter; The fire hazard branch in the output layer processes the input data to obtain the fire hazard probability predicted by the model, which is expressed using the following formula: P f =σ(W f ·h T +b f ) Among them, P f is the probability of fire hazard; σ is the sigmoid function; W f is the fire hazard branch weight; h T is the hidden state of the last time step of the LSTM layer; b f is the fire hazard branch bias parameter; The stable state probability and fire hazard probability obtained by the output layer are used as the final output of the model.

5. The energy storage fire control method based on the fusion of deep learning and machine learning according to claim 4 is characterized in that: Step S4 is specifically as follows: using the training data set obtained in step S2, training the energy storage fire control initial model obtained in step S3, optimizing the model parameters by back propagation algorithm, and obtaining the energy storage fire control model. The total loss function is expressed by the following formula: L=[yslog(Ps)+(1-ys)log(1-Ps)]-[yflog(Pf)+(1-yf)log(1-Pf)] Among them, L is the total loss function; y s is the stable state label; y f Fire hazard label.

6. The energy storage fire control method based on the fusion of deep learning and machine learning according to claim 5 is characterized in that: Step S5 includes the following steps: a. Obtain battery management system data, including single cell voltage, temperature, state of charge, depth of discharge data, and measured temperature and CO concentration data; b. Preprocess the data obtained in step a to obtain a feature vector; c. Input the characteristic vector obtained in step b into the energy storage fire control model obtained in step S4 to obtain the stable state probability P s ' and fire hazard probability P f '; d. Based on the steady state probability P s 'Adjust the K value in the KNN algorithm and use the KNN algorithm to check for false alarms. If it is judged to be a real alarm, the alarm will be issued normally. Otherwise, it will be judged to be a false alarm and no alarm will be issued.

7. The energy storage fire control method based on the fusion of deep learning and machine learning according to claim 6 is characterized in that: The step d is based on the stable state probability P s 'Adjust the K value in the KNN algorithm. The adjustment strategy is: when P s '>0.7, K value is set to 7; when P s When '≤0.7, the K value is set to 3.

8. The energy storage fire control method based on the fusion of deep learning and machine learning according to claim 7 is characterized in that: In step d, when the steady state probability P s '>0.7 and the fire hazard probability P f '<0.5, the battery management system is judged to be in normal operation, otherwise it goes to the troubleshooting step.

9. The energy storage fire control method based on the fusion of deep learning and machine learning according to claim 8 is characterized in that: In step d, the KNN algorithm is used to perform false positive judgment and investigation, including primary investigation and secondary investigation; The first-level investigation is as follows: first, obtain the battery management system data at the current time t, and construct a new sample feature vector. The elements of the new sample feature vector include the current temperature of the battery PACK, CO concentration, average voltage of the single cell, average temperature of the single cell, state of charge, and average depth of discharge; the average voltage of the single cell is the arithmetic mean of the voltages of all single cells within the preset time window; the average temperature of the single cell is the arithmetic mean of the temperatures of all single cells within the preset time window; Calculate the distance between the feature vector of the new sample and the feature vector of all samples in the training data set Use the following formula to express it: Among them, the new sample feature vector is S i =[x i1 ,x i2 ,...,x i6 ]; the jth sample in the training data set is x i1 ,x i2 ,...,x i6 They correspond to the current temperature, CO concentration, average voltage of single cells, average temperature of single cells, state of charge, and average depth of discharge of the battery PACK in the new sample feature vector respectively; Sort the obtained distances from small to large, and select the K samples with the closest distances as the first-level neighbors N of the new sample, where the value of K is the preset value adjusted in step d; Calculate the anomaly score Q1, count the number of normal state samples and false positive state samples in the first-level neighbors, and use the following formula to express it: Among them, I is the indicator function, when the nearest neighbor is a false alarm, I = 1, otherwise it is 0; If the anomaly score Q1 exceeds the set threshold, it is considered a true alarm and the alarm is issued normally; otherwise, it is marked as a suspected false alarm and enters the second-level investigation; The second-level investigation is as follows: for samples marked as suspected false positives in the first-level investigation, feature vectors are refined. The sample feature vectors are [current temperature of the battery pack, CO concentration, average voltage of the single cell, average temperature of the single cell, state of charge, average depth of discharge, temperature change rate in the past hour, CO concentration change rate in the past hour, difference between the maximum and minimum temperatures of the single cell, difference between the highest and lowest voltages of the single cell, ambient temperature, ambient humidity]; Calculate the Euclidean distance between the suspected false positive sample and all samples in the training data set Use the following formula to express it: Among them, the feature vector of suspected false positive samples is The jth sample in the training data set is They correspond to the current temperature of the battery PACK in the false alarm sample feature vector, CO concentration, average voltage of the single cell, average temperature of the single cell, state of charge, average depth of discharge, temperature change rate in the past hour, CO concentration change rate in the past hour, difference between the maximum and minimum temperatures of the single cell, difference between the highest and lowest voltages of the single cell, ambient temperature, and ambient humidity; Sort the obtained distances from small to large, and select the K samples with the closest distances as the secondary neighbors N' of the suspected false positive samples, where the value of K is the preset value adjusted in step d; Calculate the anomaly score Q2, count the number of normal state samples and false positive state samples in the secondary neighbors, and use the following formula to express it: Among them, I is the indicator function, when the nearest neighbor is a false alarm, I = 1, otherwise it is 0; If the fire hazard probability P of the suspected false alarm sample f 'If the value is lower than the set threshold and the anomaly score Q2 is lower than the set threshold, the suspected false alarm sample is judged to be a false alarm and no alarm is issued; otherwise, it is judged to be a true alarm and an alarm is issued normally.

10. A system for implementing the energy storage fire control method based on the fusion of deep learning and machine learning as described in any one of claims 1 to 9, characterized in that: It includes data acquisition module, data processing module, model building module, model training module and energy storage system fire control module; The data acquisition module acquires the historical data of the battery management system and uploads the data to the data processing module; The data processing module preprocesses the received data to obtain feature vectors, labels the feature vectors, obtains training data sets, and uploads the data to the model training module; The model building module builds the initial model of energy storage fire control based on the LSTM network model and uploads the data to the model training module; The model training module uses the training data set to train the energy storage fire control initial model based on the received data to obtain the energy storage fire control model, and uploads the data to the energy storage system fire control module; The energy storage system fire control module uses the energy storage fire control model to perform actual energy storage system status monitoring based on the received data and complete the control of the energy storage fire control system.

Images