Control method for fire-fighting false alarm of energy storage system

Abstract

A control method for fire-fighting false alarm of an energy storage system comprises the following steps of: calculating a fluctuation period, a derivative and a probability density function by using fast Fourier transform and the like by acquiring data streams such as current, voltage, power and heat energy of the energy storage system, determining a system state and predicting a related trend. And predicting power by using an autoregressive moving average algorithm and a support vector machine, and setting a threshold value according to a probability density function to judge abnormity. For heat energy data, through derivative calculation and signal analysis, states are identified and early warning is carried out in combination with models such as a deep convolutional adversarial neural network and a support vector machine. The normal range of monitoring data is determined by means of the long-short-term memory neural network, the change trend and physical quantity relation of abnormal data are analyzed and predicted through a cubic exponential smoothing method and Gaussian process regression, potential risks are judged through weighted fusion and a self-adaptive Simpson algorithm, fire-fighting false alarms are effectively reduced, fire risks are accurately recognized, and the safety of an energy storage system is guaranteed.

Description

Technical Field

[0001] The present invention belongs to the technical field of fire alarm control of energy storage systems, and in particular relates to a control method for fire alarm control of energy storage systems. Background Art

[0002] During the operation of the energy storage system, due to the complex electrochemical reactions inside, the operation of electrical equipment and environmental factors, various sensors will continuously monitor the system's operating parameters, such as current, voltage, power and heat energy. These parameter data are of great significance for timely detection of potential system failures and prevention of fire accidents. However, the main technical problem currently faced is that the operating environment of the energy storage system is complex and changeable, resulting in a large amount of noise interference and uncertainty in the data collected by the sensors. These interference factors can easily cause the traditional fire alarm system to misjudge, that is, false fire alarms.

[0003] On the one hand, normal system operation fluctuations, such as current and voltage changes during charging and discharging, temperature fluctuations during heat dissipation, etc., may be misjudged as abnormal conditions and trigger alarms, which will not only interfere with the operation and maintenance personnel's judgment of the true state of the system and reduce work efficiency, but also may cause the operation and maintenance personnel to become numb to the alarm due to frequent false alarms, and fail to respond in time when real fire hazards occur. On the other hand, some minor fault hazards may show characteristics similar to normal fluctuations in the early stage. If they cannot be accurately distinguished, the fire protection system may miss the best warning opportunity, thereby increasing the risk of fire and posing a serious threat to personnel safety, equipment assets and the surrounding environment.

[0004] In addition, most of the existing energy storage system fire alarm technologies are based on single parameters or simple threshold judgments, lacking the ability to comprehensively analyze multi-source data and accurately model and predict the system's operating status. It is impossible to conduct in-depth mining and analysis based on the system's historical operating data and real-time monitoring data, making it difficult to accurately identify real fire risk signals, and it is difficult to meet the high reliability and high safety fire warning requirements of energy storage systems under complex working conditions.

[0005] Therefore, there is an urgent need to develop a control method that can effectively reduce false fire alarms and accurately identify fire risks in energy storage systems. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a control method for false fire alarms in an energy storage system, predict the development trend and severity of fire risks in advance, provide operation and maintenance personnel with sufficient time to formulate response strategies, minimize the possibility of fire accidents and the extent of losses after the fire, ensure the safe and stable operation of the energy storage system, and also effectively protect the surrounding environment and personnel safety.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is: A control method for false alarm of fire protection of an energy storage system, comprising the following steps: S1. Obtain the current value, voltage value and power value data stream transmitted by the energy storage system sensor in real time, calculate the fluctuation period of the system operation speed through fast Fourier transform, calculate the second-order derivative of the fluctuation period, and obtain the probability density function of the data change trend; S2. According to the probability density function of the data change trend, the quantiles of the normal operation range under different states are set, and combined with the pre-set empirical upper threshold of the normal operation range, if the probability of the data change trend exceeds the upper threshold, the alarm mechanism is triggered; S3, collecting the thermal energy changes of data streams at multiple time points, obtaining the first-order derivative distribution of thermal energy data with respect to time, obtaining the second-order derivative curve of temperature change data with respect to time, and constructing a thermal energy change rate index data set; S4. Train the support vector machine model based on the heat energy change rate index data set, extract the classification boundary between normal state data and abnormal state data, and build a state recognition mechanism based on the weight ratio of each state; S5. Based on the state recognition mechanism established by the support vector machine, when the monitoring data of the energy storage system triggers the upper threshold limit of the preset safe operation range, the difference of historical information records is compared to determine the safety state level, and then the corresponding warning mode is triggered; S6. By training the long short-term memory neural network model, using the existing data change trend data as features, extracting the time series features of the data change trend data and the monitoring data intensity, thereby determining the feature differences between the historical state data set and the current state data set, and outputting the normal operating range of each monitoring data of the system; S7. When the monitoring data deviates from the normal operating range after being judged by the long short-term memory neural network, the monitoring data is included in the abnormal database, and the monitoring data intensity of rare phenomena in the abnormal database is fitted by the triple exponential smoothing method to obtain the probability density distribution of the change trend of the predicted data of the future time series; S8. Calculate the standard deviation of the internal data for a period of time based on the results of rare phenomenon identification, and substitute it into the covariance kernel function of the Gaussian process for regression analysis, so as to obtain the functional relationship between the predicted data change trend and the change frequency of the physical quantity; S9. Perform weighted fusion based on the probability density distribution of the predicted data change trend and the relationship between the frequency of change of the physical quantity, perform surface integral of the function graph through the adaptive Simpson algorithm, and judge whether there is a potential safety risk based on the size of the area of ​​the integral result.

[0008] Preferably, the sub-steps of step S1 are: S1.1, collect the current value, voltage value and power value data stream output by the energy storage system sensor; S1.2, calculate the fast Fourier transform of the three parameter data streams of current value, voltage value and power value, obtain the frequency domain component signal, and determine the system fluctuation period according to the corresponding relationship between the specific frequency component and the fluctuation period in the frequency domain component signal; S1.3, calculating the second-order derivative value based on the acquired fluctuation period information to obtain the fluctuation state change rate index; S1.4. Obtain the probability density function based on the numerical calculation of the fluctuation state change rate to determine the statistical distribution probability of the system operating in each fluctuation range; S1.5. Determine the current energy storage state of the energy storage system through the existing relationship set between the probability density function and the energy storage state; S1.6. Use the known voltage and current data and the autoregressive moving average algorithm to predict the voltage and current trends over a period of time. S1.7. The obtained voltage and current trends are input into the support vector machine to determine the output power value under the preset period after the corresponding energy storage system changes.

[0009] Preferably, the sub-steps of step S2 are: S2.1. Collect state data to build a data set. The data set includes historical data under multiple different states. The state data is input into a probability density function model to obtain a probability density function corresponding to the data change trend and obtain a probability value. S2.2. Using the obtained probability density function, according to the preset quantile value, reversely solve the quantile value to the function, obtain the normal operation range corresponding to each state, obtain the corresponding upper and lower thresholds, and construct the range interval of different quantile points corresponding to each state; S2.3, obtain the state data and calculate, calculate the statistical result of the current data by histogram statistics, and use the statistical result to determine the probability of obtaining the state data; S2.4. Compare the probability of the current state data with the probability corresponding to different quantiles of the state to make a threshold judgment; S2.5. If the current state probability value is greater than the occurrence probability corresponding to the preset quantile threshold, an alarm signal is triggered, and the current state data and related time information are recorded for subsequent analysis.

[0010] Preferably, the sub-steps of step S3 are: S3.1. The time domain sensor records and obtains the data stream of the thermal energy value of the object at multiple times, calculates the difference of the thermal energy values ​​at adjacent times, and obtains the first-order derivative of the thermal energy with respect to time; S3.2, using the obtained thermal energy difference data of adjacent moments divided by the time difference of the moments, to obtain temperature change data; S3.3, using the differential method to obtain the temperature change data, and obtain the second-order derivative curve of temperature with respect to time; S3.4. Count the frequency of the second-order derivative of temperature exceeding the preset range within the preset time, and construct a thermal energy change rate index reflecting the dynamic fluctuation amplitude of thermal energy; S3.5. Perform time domain signal analysis on the acquired thermal energy values ​​and calculate the Fourier transform of its frequency components; for discrete thermal energy signal time series , is the sequence length, and its discrete Fourier transform The expression is: ; in, Corresponding to different discrete frequency points, and = 0, 1, 2, ..., N - 1; is an imaginary unit, satisfying , is the ratio of pi; the spectrum components of the thermal energy signal at different discrete frequency points can be obtained through the discrete Fourier transform; these spectrum components are then analyzed to determine the proportion of high-frequency components. If the high-frequency component exceeds the predetermined proportion threshold, the next step is executed: In the set constructed by the second-order derivative values ​​corresponding to the temperature changes at different times, the outlier detection LOF algorithm is used to detect and identify the moments that deviate from other data, where is the time series of thermal energy signal, is the interval between moments; If it is a continuous thermal energy signal time series , whose Fourier transform The expression is:

[0011] Similarly, through The analysis determines the proportion of high-frequency components, and the subsequent processing is the same as the discrete case; is the continuous frequency; S3.6. If the moment is identified as abnormal, execute the next step: The thermal energy change data is imported into the trained deep convolutional adversarial neural network model. The model uses an encoder to map the thermal energy features into latent space vectors, and the discriminator distinguishes the probability that the latent space vector is real data.

[0012] Preferably, the sub-steps of step S4 are: S4.1. Obtain a data set of thermal energy change rate according to the time series of the collected thermal energy data; S4.2, normalizing all data of the heat energy change rate data set to obtain a normalized data set, and then setting a normal state data mark and an abnormal state data mark according to preset conditions; S4.3, obtaining a labeled normal state and abnormal state data set from the normalized data set, training a support vector machine model, and obtaining an optimal support vector machine classification model; S4.4. Using all the marked normal state data and abnormal state data, obtain the classification boundary hyperplane equation of the support vector machine classification model , represents the rate of change of thermal energy, is the weight vector, is the bias term; S4.5. Obtain the number of occurrences of the normal state and the abnormal state of the marked data by statistics, and obtain the ratio of the number of occurrences of the two, and determine the state weight ratio value; S4.6. According to the support vector machine classification model and the state weight ratio value, the state of the real-time collected heat energy change rate data is judged. If the heat energy change rate is greater than the classification boundary value and the number of abnormal heat energy states is greater than the state weight ratio value, a warning state is obtained; S4.7. When the thermal energy status information corresponding to the thermal energy change rate data is collected in real time, the associated equipment type, environmental information and process parameter data are extracted from the database for data fusion, and the equipment operation status and maintenance information are obtained based on the above data.

[0013] Preferably, the sub-steps of step S5 are: S5.1, the support vector machine obtains the real-time energy storage system operation data based on the acquisition device, and generates monitoring data after performing preprocessing operations; S5.2, the monitoring data reaches the upper limit of the preset safe operation range threshold, triggering the safety module, and the operation data is transmitted to the historical database to form an information record; S5.3. Obtain all information records from the historical database as comparison objects, calculate the comparison objects and current operation data, and obtain the degree of difference in the operation data; S5.4, according to the pre-set level rules corresponding to the degree of difference, determine which security level the current monitoring data belongs to, and generate a judgment result; S5.5, obtaining one of a plurality of different pre-configured warning modes according to the judgment result, and using a pre-established support vector machine to perform state recognition; S5.6, state identification outputs a state signal set, which includes multiple state signals; S5.7. If any status signal exceeds the preset safety range, the corresponding warning mode is activated to issue an alarm prompt; S5.8. The system continuously obtains status signals. If it finds that the safety range is always out of limit, it continuously issues warnings and cyclically executes the support vector machine to perform status recognition operations and update the comparison object.

[0014] Preferably, the sub-steps of step S6 are: S6.1. Obtain historical data according to each type of monitoring point and split the data. The first historical data set is used as a training set, the second historical data set is used as a test set, and the third historical data set is used as a validation set. S6.2. Perform data cleaning, data denoising, data transformation and feature extraction on the training set data. The features include trend, intensity and fluctuation range, and construct feature vectors to obtain the preliminary training set. S6.3. Based on the preliminary training set data, a feature difference extraction model was constructed using the long short-term memory neural network algorithm. The LSTM model defined the number of hidden layers as 2, the number of neurons as 64, and the learning rate as 0.01. The objective function of the model was set to the mean square error function, and the number of iterations was 500 to obtain the first training model. S6.4. Use the validation set data to input the first training model to obtain a first prediction value, and obtain the accuracy rate through the historical normal range of the validation set historical data; S6.5. If the model accuracy is lower than the threshold, modify the network structure parameters, reset the learning rate to 0.005 and the number of hidden layers to 3, etc., update the first training model to the second training model; return to step S6.4; If the model accuracy reaches the threshold, the first final model is obtained; Input the real-time monitoring data, obtain the data N minutes before the current time, and perform data preprocessing to obtain real-time trend data and intensity data; Input real-time trend data and intensity data, calculate the characteristic difference through the first final model, compare the normal range value, and determine whether the value of each monitoring point is normal or not.

[0015] Preferably, the sub-steps of step S7 are: S7.1. Obtain the abnormal monitoring data stream after the long short-term memory network has determined that the current abnormal data stream has exceeded the preset normal data range; S7.2, judging the time point of occurrence of the current abnormal data according to the monitoring duration of the data point in the abnormal monitoring data stream; S7.3, obtaining the intensity value of the data point in the current abnormal data stream, using the intensity corresponding to all historical abnormal monitoring data points, and obtaining the optimal values ​​of the three smoothing coefficients in the cubic exponential smoothing method through the least squares method; S7.4. According to the cubic exponential smoothing method and the smoothing coefficient, the intensity corresponding to the abnormal monitoring data point that has occurred at the current moment is used to fit the intensity corresponding to all abnormal monitoring data points in the time period after the current moment; S7.5. Based on the intensity values ​​corresponding to all the fitted abnormal monitoring data points, the probability density distribution function of the intensity corresponding to all the abnormal monitoring data points is calculated. The calculation formula is: P(x)=(1 / (σ√(2π)))e^(-((x - μ)^2) / (2σ^2)), Where P(x) represents the probability density of data point x, μ represents the mean of the data point set, σ represents the standard deviation of the data point set, e represents the base of the natural logarithm, and π represents pi; S7.6. Obtain the probability density of the intensity corresponding to all abnormal monitoring data points obtained by calculation, and determine the extreme value interval of the abnormal data intensity by using a preset threshold method; S7.7. Based on the extreme value interval of abnormal data intensity and the current abnormal data monitoring time point, obtain the changing trend of this section of abnormal data at future moments.

[0016] Preferably, the sub-steps of step S8 are: S8.1. After obtaining the time series of historical data, identify rare events in the time series and obtain the standard deviation of the time series based on this result, and record the standard deviation as the first standard deviation; S8.2. After obtaining the time series of the first standard deviation and the corresponding time interval, perform data preprocessing, obtain the second time series by median filtering, obtain the moment of each data point in the second time series and mark it as the second moment; S8.3. Using the second time series and the corresponding second moment, with the data point as the center, take the three data points before and after to calculate their standard deviation as the standard deviation of the neighborhood near each second moment data point, and substitute the standard deviation of all neighborhoods together with the first standard deviation into the covariance kernel function of the Gaussian process Among them and Represents data at different time points in a time series. As an independent variable, As the dependent variable, , is the standard deviation, It is the characteristic length of time length; S8.4, constructing a Gaussian process regression model based on the covariance kernel function, and performing regression analysis on the processed second time series, thereby obtaining an initial function model between the time series change and the frequency of physical quantity change, the model being constructed based on the regression analysis and the second time series; S8.5. According to the initial function model, the processed time series is segmented, a time length is set to obtain a third moment and a plurality of third time series, and a standard deviation and a change frequency of each third time series are calculated to generate a frequency feature vector group; S8.6. Perform prediction analysis based on the initial function model, input the elements in the frequency feature vector group into the initial function model respectively, obtain the fourth time series and the corresponding fourth moment, and calculate the difference between the fourth moment series and the third moment series to obtain the frequency error.

[0017] Preferably, the sub-steps of step S9 are: S9.1. A matrix of signals in the acquisition device is established based on multiple related attribute data in the device operation timing data stream. A row or a column in the matrix is ​​called a vector of the signal. The signal amplitude value at each acquisition moment constitutes a vector component. A discrete Fourier transform is performed on a vector to obtain information data representing the frequency of change of the physical quantity. S9.2. Obtain the modulus data of the vector after discrete Fourier transformation and determine the information entropy value of each data according to the Shannon information entropy calculation method, use the maximum entropy of all data at multiple collection times as the measurement value of the risk weight, and then obtain the fused value of each vector in each column or row of the matrix; S9.3, perform principal component analysis and eigenvalue decomposition on the matrix to obtain the importance weights of the main vector groups of the signal represented by eigenvalues ​​of different sizes arranged in descending order, and obtain the normalization factor of the weighted eigenvector group after accumulating the weights according to all eigenvalues ​​in proportion; S9.4. Collect and obtain the density function that represents the probability of the change trend of the predicted data. The density function input uses the normalization factor of the weight feature vector group obtained in the previous step as the input. After the function output of the probability density function is weighted and calculated with the weight, if there is a distribution offset of the predicted density function, the gradient descent method is used to iteratively and adaptively correct the parameters in the probability density function for the offset phenomenon. S9.5. By obtaining the data parameters of the prediction probability density function in the previous step, the upper limit and lower limit of the integral interval in the probability density function are determined, and the integral function body based on the adaptive Simpson algorithm is obtained; S9.6. Determine the relationship between the set precision value and the preset threshold value for terminating the integral when the function area is obtained by integration. If the precision value is greater than the threshold, obtain the value of the lower half of the current precision, use the value as the adaptive value of the Simpson algorithm, and calculate the function in this step to obtain the sub-interval area. The multiple sub-interval areas obtained in this step are combined to obtain the integral result of this step. S9.7. Based on the multiple result values ​​obtained from all the above calculations, determine the specific location and device attribute information corresponding to the different device location data represented by each data result. The signal matrix of all device locations at different times constitutes a tensor. The risk value of different components and positions of the device in a certain time series is calculated using the tensor to make a judgment. Judge the value and the risk threshold. If the risk value is greater than or equal to the risk threshold, it is determined that the corresponding component has potential risk.

[0018] The present invention can achieve the following beneficial effects: 1. The present invention conducts comprehensive and in-depth analysis and processing of multi-source data collected by energy storage system sensors, and uses a series of complex algorithms such as fast Fourier transform to accurately calculate the fluctuation period of the system operation speed and its related indicators, which can effectively distinguish between normal operation fluctuations and abnormal danger signals. Compared with the traditional method based on simple threshold judgment, it greatly reduces the probability of false alarms caused by normal system fluctuations and improves the accuracy of fire alarms.

[0019] 2. The present invention sets the quantile and empirical upper limit threshold judgment mechanism of the normal operating range under different states based on the probability density function of the data change trend, as well as the detailed analysis and processing of thermal energy data. This can more accurately capture possible abnormal conditions in the system, avoid triggering unnecessary alarms due to data interference or misjudgment, and ensure that the fire alarm system only sounds an alarm when it is actually facing a fire risk, reducing the workload of operation and maintenance personnel in troubleshooting and handling false alarms, and improving the efficiency of the operation and maintenance management of the entire energy storage system.

[0020] 3. The present invention uses a support vector machine model trained with a heat energy change rate index data set to accurately extract the classification boundary between normal state data and abnormal state data, and based on the perfect state recognition mechanism constructed according to the state weight ratio, it can determine the operating state of the energy storage system in real time and accurately. Not only can it timely discover the current anomalies, but it can also predict the possible failure risks of the system in the future based on historical data and current state trends, providing strong support for preventive maintenance, effectively extending the service life of the energy storage system, reducing equipment failures and downtime, and improving the overall reliability and availability of the energy storage system.

[0021] 4. The present invention extracts and analyzes the time series characteristics of data change trends and monitoring data intensity changes through the long short-term memory neural network model, which can deeply explore the characteristic differences between the historical state data set and the current state data set, so as to accurately output the normal operating range of each monitoring data of the system. This helps to discover the trend of minor abnormal changes in monitoring data in advance, and take timely measures to adjust or repair before a fault occurs, further improving the stability and safety of the energy storage system operation, and also provides data basis for the optimized operation of the system, which can optimize and improve the performance of the energy storage system.

[0022] 5. The present invention uses the cubic exponential smoothing method to fit the monitoring data intensity of rare phenomena in the abnormal database, obtains the probability density distribution of the predicted data change trend of the future time series, and substitutes the data standard deviation into the covariance kernel function of the Gaussian process for regression analysis to obtain the functional relationship between the predicted data change trend and the change frequency of the physical quantity, and then performs weighted fusion and adaptive Simpson algorithm integration based on these results to judge the potential safety risks. This multi-dimensional and multi-level risk assessment method can comprehensively consider the various operating parameters and historical data characteristics of the energy storage system, and conduct a comprehensive and in-depth assessment of the fire risks that the system may face. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 It is a logic flow chart of the present invention. DETAILED DESCRIPTION

[0024] The preferred solution is Figure 1 As shown, a method for controlling false alarms of fire in an energy storage system comprises the following steps: S1. Obtain the current value, voltage value and power value data stream transmitted by the energy storage system sensor in real time, calculate the fluctuation period of the system operation speed through fast Fourier transform, calculate the second-order derivative of the fluctuation period, and obtain the probability density function of the data change trend; specifically: S1.1, collect the current value, voltage value and power value data stream output by the energy storage system sensor; In actual operation, the accuracy and stability of the sensors need to be ensured, and the sampling frequency should be reasonably set according to the characteristics of the energy storage system and the requirements for real-time data. For example, for some fast-response energy storage systems, the sampling frequency may need to reach thousands of times per second to accurately capture small changes in current, voltage and power. These sensors should be evenly distributed in key parts of the energy storage system, such as the connection points of the battery modules, the power conversion circuit, etc., so as to fully reflect the operating status of the system.

[0025] S1.2, calculate the fast Fourier transform of the three parameter data streams of current value, voltage value and power value, obtain the frequency domain component signal, and determine the system fluctuation period according to the correspondence between the specific frequency component and the fluctuation period in the frequency domain component signal (for example, the frequency and period conversion rule determined by the empirical formula or industry standard); The implementation of the fast Fourier transform algorithm requires the use of efficient computing resources, such as a dedicated digital signal processor (DSP) or a field programmable gate array (FPGA) with powerful computing power. For different types of energy storage systems, the correspondence between their specific frequency components and fluctuation periods may vary, and a large number of experiments and data analysis are required to determine the accurate conversion rules. For example, for lithium-ion battery energy storage systems, their internal chemical reactions and electrical characteristics determine that certain specific frequency ranges are closely related to the battery's charge and discharge status and health status. By analyzing these frequency components, the system's fluctuation period can be accurately determined, and the system's operating stability can be understood.

[0026] S1.3, calculating the second-order derivative value based on the acquired fluctuation period information to obtain the fluctuation state change rate index; This step can not only reflect the current fluctuation state of the system, but also provide an important basis for the subsequent prediction of the changing trend of the system. The calculation of the second-order derivative value requires the use of accurate numerical calculation methods to avoid misjudgment of the system state due to calculation errors. At the same time, in order to improve the calculation efficiency, the data can be properly preprocessed, such as data smoothing to remove noise interference, but care should be taken not to over-smooth and lose important fluctuation information.

[0027] S1.4. Obtain the probability density function based on the numerical calculation of the fluctuation state change rate to determine the statistical distribution probability of the system operating in each fluctuation range; When calculating the probability density function, methods such as kernel density estimation can be used to construct a probability distribution model of the system's operating state based on historical data and the current fluctuation state change rate value. By analyzing this model, we can clearly understand the probability of the system appearing in different fluctuation ranges, thereby providing a quantitative basis for judging whether the system is in a normal operating state. For example, if the probability of the system in a certain abnormal fluctuation range exceeds the set threshold, it may indicate that the system has a potential failure risk.

[0028] S1.5. Determine the current energy storage state of the energy storage system through the existing relationship set between the probability density function and the energy storage state; This relationship set is established through the analysis and research of a large amount of historical data, covering the probability density function characteristics of the energy storage system under different energy storage states. For example, in different states such as charging state, discharging state, and static state, the fluctuation law of the system's current, voltage, and power and the corresponding probability density function are different. By comparing the currently calculated probability density function with this relationship set, the specific energy storage state of the energy storage system can be accurately determined.

[0029] S1.6. Using the known voltage and current data, combined with the autoregressive moving average algorithm (where the parameters of the autoregressive moving average algorithm are set as: autoregressive order p = 3, moving average order q = 2, based on the preliminary analysis of historical data and industry experience, the parameter range is determined to be effective for processing such data), predict the voltage and current trends for a period of time in the future; S1.7. The obtained voltage and current trends are input into a support vector machine (the support vector machine uses a radial basis kernel function, and the kernel function parameter gamma is selected from the optimal value in the range of [0.1, 10] by cross-validation method, and the penalty parameter C is selected from the range of [1, 100] to balance the fitting ability and generalization ability of the model) to determine the output power value under the preset period after the corresponding energy storage system changes.

[0030] The radial basis kernel function can effectively handle nonlinear relationships. By selecting the gamma and C parameters within a given range through the cross-validation method, the model configuration that best suits the current data can be found. During the cross-validation process, the data set is divided into a training set, a validation set, and a test set. The parameters are continuously adjusted to optimize the model's performance on the validation set, and then the final evaluation is performed on the test set. The support vector machine model obtained in this way can accurately predict the output power value under the preset cycle based on the input voltage and current trends, providing a key basis for the energy management and control of the energy storage system.

[0031] S2. According to the probability density function of the data change trend, the quantiles of the normal operation range under different states are set. Combined with the pre-set empirical upper threshold of the normal operation range, if the probability of the data change trend exceeds the upper threshold, the alarm mechanism is triggered; specifically: S2.1. Collect state data (state data includes temperature, pressure and ambient humidity of key components of energy storage system) to build a data set. The data set includes historical data under multiple different states. The state data is input into the probability density function model to obtain the probability density function corresponding to the data change trend and obtain the probability value. When collecting status data, high-precision sensors should be used to ensure the accuracy of the data. For example, the accuracy of the temperature sensor should reach ±0.1℃, the accuracy of the pressure sensor should be within the range of ±0.01MPa, and the accuracy of the ambient humidity sensor should also be controlled within a certain error range. These sensors should be installed near the key components of the energy storage system, such as battery modules, heat dissipation devices, electrical connections, etc., so as to monitor the operating status of the system in a timely and accurate manner. For different types of energy storage systems, their key components and corresponding status data may be different, and targeted collection and analysis are required according to specific circumstances. By inputting a large amount of historical status data into the probability density function model, the relationship between the data change trend and the probability density function of the system in different states can be established, thereby providing a basis for the subsequent setting of the normal operating range.

[0032] S2.2. Using the obtained probability density function, according to the pre-set quantile value (for example, the 95% quantile is set as the basis for judging the normal operation range, which can be determined based on a large amount of historical data statistical analysis and the safety requirements for the energy storage system), reversely solve the quantile value to the function, obtain the normal operation range corresponding to each state, obtain the corresponding upper and lower thresholds, and construct the range interval of different quantiles corresponding to each state; When determining the percentile value, it is necessary to comprehensively consider factors such as the reliability requirements of the energy storage system, the probability of failure, and the false alarm rate. For some energy storage systems with extremely high safety requirements, a higher percentile value, such as 99%, may be selected to ensure that the system can be judged to be operating normally in most cases. For some systems that are more sensitive to cost and efficiency, the percentile value can be appropriately lowered, but the basic safety of the system must also be guaranteed. By reversely solving the probability density function, the specific values ​​of the normal operating range in each state can be obtained, and these values ​​will serve as an important criterion for judging whether the system is operating normally.

[0033] S2.3, obtain the state data and calculate, calculate the statistical result of the current data by histogram statistics, and use the statistical result to determine the probability of obtaining the state data; Histogram statistics is an intuitive and effective data statistics method. It can divide data into certain intervals and count the number or frequency of data in each interval. In this step, the current state data is classified and counted according to the preset interval width to obtain the distribution of data in each interval, thereby calculating the probability of the current data. For example, if the current temperature data has a large number of items in the middle interval of the normal operating range, then the corresponding probability is higher; conversely, if the data is close to or exceeds the boundary interval of the normal operating range, the probability will be reduced accordingly.

[0034] S2.4. Compare the probability of the current state data with the probability corresponding to different quantiles of the state to make a threshold judgment; This step is the key to determine whether the system is operating normally. Compare the currently calculated probability of the state data with the pre-set probability corresponding to different quantiles. If the current probability exceeds the probability corresponding to the upper threshold, it means that the system may be in an abnormal state and needs further analysis and processing. When making threshold judgments, the volatility and uncertainty of the data should be taken into account to avoid misjudgment due to accidental factors. Some filtering algorithms or statistical test methods can be used to further verify and confirm the data.

[0035] S2.5. If the current state probability value is greater than the occurrence probability corresponding to the preset quantile threshold, an alarm signal is triggered, and the current state data and related time information are recorded for subsequent analysis.

[0036] When an alarm signal is triggered, relevant information should be transmitted to the monitoring center or the operator's terminal device in a timely manner so that appropriate measures can be taken. The recorded status data and time information should include detailed sensor data, data acquisition time, system operation status, etc. This information will provide an important basis for subsequent troubleshooting and system optimization. For example, by analyzing historical alarm data, the laws and causes of system failures can be found, and targeted improvement measures can be taken to improve the reliability and stability of the system.

[0037] S3. Collect the thermal energy changes of data streams at multiple time points, obtain the first-order derivative distribution of thermal energy data with respect to time, obtain the second-order derivative curve of temperature change data with respect to time, and construct a thermal energy change rate index data set; specifically: S3.1. The time domain sensor records and obtains the data stream of the thermal energy value of the object at multiple times, calculates the difference of the thermal energy values ​​at adjacent times, and obtains the first-order derivative of the thermal energy with respect to time; The sampling frequency of the time domain sensor should be reasonably set according to the rate of thermal energy change and the requirements for data accuracy. Generally speaking, for energy storage systems with faster thermal energy changes, the sampling frequency should be higher, such as more than 10 times per second. When calculating the difference in thermal energy values ​​at adjacent moments, attention should be paid to the accuracy and stability of the data to avoid deviations in the calculation results due to sensor errors or data transmission interference. Methods such as averaging multiple measurements or data filtering can be used to improve the accuracy of the calculation.

[0038] S3.2, using the obtained thermal energy difference data of adjacent moments divided by the time difference of the moments, to obtain temperature change data; The time difference should be kept constant and is usually determined according to the sampling frequency. For example, if the sampling frequency is 5 times per second, the time difference is 0.2 seconds. By dividing the thermal energy difference by the time difference, the thermal energy change per unit time, i.e., the temperature change data, can be obtained. These data reflect the temperature change rate of the energy storage system at different times and are important indicators for judging the thermal stability of the system.

[0039] S3.3, using the differential method to obtain the temperature change data, and obtain the second-order derivative curve of temperature with respect to time; Differential calculation is a commonly used numerical calculation method used to approximate the derivative of a function. In this step, by performing differential operations on the temperature change data, a second-order derivative curve of temperature with respect to time can be obtained, which can reflect the acceleration of temperature change, that is, the change in the speed of temperature change. The shape and characteristics of the second-order derivative curve can provide important clues for judging whether the system has abnormal conditions such as thermal runaway. For example, if the second-order derivative curve rises sharply or fluctuates greatly, it may indicate problems such as local overheating or uneven heat transfer inside the system.

[0040] S3.4. Count the frequency of the second-order derivative of temperature exceeding a preset range within a preset time (the preset range is determined based on the statistical analysis of historical data of the temperature change rate during normal operation of the energy storage system, for example, set to plus or minus 0.5 degrees Celsius per second), and construct a thermal energy change rate indicator that reflects the dynamic fluctuation amplitude of thermal energy; When counting the frequency of occurrences where the second-order derivative of temperature exceeds the preset range, the length of the preset time should be determined based on the characteristics of the energy storage system and the timeliness requirements for fault detection. For example, for some energy storage systems with faster responses, the preset time can be set to a shorter period of time, such as 10 seconds; while for some systems with slower changes, the preset time can be appropriately extended, such as 1 minute. By counting the frequency of occurrence, an indicator that reflects the dynamic fluctuation amplitude of thermal energy can be constructed. This indicator can intuitively reflect the severity of the thermal energy changes in the system over a period of time, providing a basis for subsequent abnormality detection and risk assessment.

[0041] S3.5. Perform time domain signal analysis on the acquired thermal energy values ​​and calculate the Fourier transform of its frequency components; for discrete thermal energy signal time series , is the sequence length, and its discrete Fourier transform The expression is: ; in, Corresponding to different discrete frequency points, and = 0, 1, 2, ..., N - 1; is an imaginary unit, satisfying , is the ratio of pi. The spectrum components of the thermal energy signal at different discrete frequency points can be obtained through the discrete Fourier transform. Then these spectrum components are analyzed to determine the proportion of high-frequency components. If the high-frequency component exceeds a predetermined ratio threshold (the ratio threshold is determined by comparing the high-frequency components of the thermal energy signal under normal operation and fault conditions of the energy storage system, for example, set to 30%), the next step is executed: In the set constructed by the second-order derivative values ​​corresponding to the temperature changes at different times, the outlier detection LOF algorithm is used (the parameter minPts of the LOF algorithm is set to 5, and this parameter is determined according to the local density distribution of the data points to effectively identify abnormal data points) to detect and identify the moments that deviate from other data, where is the time series of thermal energy signal, is the interval between moments; If it is a continuous thermal energy signal time series , whose Fourier transform The expression is:

[0042] Similarly, through The analysis determines the proportion of high-frequency components, and the subsequent processing is the same as the discrete case; is a continuous frequency.

[0043] S3.6. If the moment is identified as abnormal, execute the next step: The thermal energy change data is imported into the trained deep convolutional adversarial neural network model (the structure of the deep convolutional adversarial neural network model is: it contains 3 convolutional layers, the convolution kernel sizes are 3x3, 5x5, and 3x3 respectively, the number of convolution kernels in each layer is 16, 32, and 64 respectively, and the ReLU activation function is used. There are 2 fully connected layers, and the number of neurons is 128 and 64 respectively. The learning rate during model training is set to 0.001, and the Adam optimization algorithm is used. The number of training iterations is 1000 times. The model structure and parameters are obtained based on training and optimization of a large amount of sample data). The model uses the encoder to feature map the thermal energy into a hidden space vector, and the discriminator distinguishes the probability that the hidden space vector is real data.

[0044] Deep convolutional adversarial neural networks have powerful feature extraction and classification capabilities in processing image, signal and other data. In this application, by inputting thermal energy change data into the trained model, the model's encoder can automatically extract deep features in the thermal energy data and map them into latent space vectors. These latent space vectors can more effectively represent the intrinsic characteristics of thermal energy data, which is convenient for subsequent analysis and processing. The discriminator further judges the authenticity and abnormality of the data by distinguishing the probability that the latent space vector is real data. In the process of model training, a large number of normal and abnormal thermal energy data samples are required. By continuously adjusting the parameters of the model, such as the weight and bias of the convolution kernel, the model can accurately identify different types of thermal energy data. At the same time, in order to avoid the occurrence of overfitting, technical means such as data enhancement and regularization can be used to improve the generalization ability of the model.

[0045] S4. Train the support vector machine model based on the heat energy change rate index data set, extract the classification boundary between normal state data and abnormal state data, and build a state recognition mechanism based on the weight ratio of each state; specifically: S4.1. Obtain a data set of thermal energy change rate according to the time series of the collected thermal energy data; When collecting thermal energy data, it is necessary to ensure the integrity and continuity of the data to avoid data loss or interruption. Data storage and backup technology can be used to regularly save and back up thermal energy data so that complete time series data can be obtained when needed. At the same time, the quality of the data must also be strictly checked and screened to remove outliers and noise data to improve data reliability.

[0046] S4.2, normalizing all data of the heat energy change rate data set to obtain a normalized data set, and then setting a normal state data mark and an abnormal state data mark according to preset conditions; Normalization can map data of different magnitudes to the same range, for example, normalizing data to the interval [0,1] or [-1, 1], which can improve the efficiency and accuracy of model training. When setting data labels, it is necessary to clarify which data belongs to normal status and which belongs to abnormal status based on the actual operation of the energy storage system and the fault diagnosis standards. The labeling rules can be determined by combining expert experience, historical fault data and relevant industry standards.

[0047] S4.3, obtaining a labeled normal state and abnormal state data set from the normalized data set, training a support vector machine model, and obtaining an optimal support vector machine classification model; When training the support vector machine model, it is necessary to reasonably select the kernel function and adjust the model parameters. As mentioned above, the radial basis kernel function is used, and the gamma and C parameters are selected within a given range through the cross-validation method to optimize the performance of the model. At the same time, the data set should be divided into a training set, a validation set, and a test set. During the training process, the performance of the model on the validation set is continuously evaluated, and the model parameters are adjusted according to the evaluation results until the model achieves a good performance indicator on the validation set.

[0048] S4.4. Using all the marked normal state data and abnormal state data, obtain the classification boundary hyperplane equation of the support vector machine classification model , represents the rate of change of thermal energy, is the weight vector, is the bias term; S4.5. Obtain the number of occurrences of the normal state and the abnormal state of the marked data by statistics, and obtain the ratio of the number of occurrences of the two, and determine the state weight ratio value; S4.6. According to the support vector machine classification model and the state weight ratio value, the state of the real-time collected heat energy change rate data is judged, and the number of abnormal states of the heat energy change rate (by setting a sliding time window, counting the number of abnormal states in the window, and the window size is set to 10 sampling cycles according to the time characteristics of the heat energy change and the system response requirements) is greater than the state weight ratio value, then the warning state is obtained; S4.7. When the thermal energy status information corresponding to the thermal energy change rate data is collected in real time, the associated equipment type, environmental information and process parameter data are extracted from the database for data fusion, and the equipment operation status and maintenance information are obtained based on the above data.

[0049] S5. Based on the state recognition mechanism established by the support vector machine, when the monitoring data of the energy storage system triggers the upper threshold limit of the preset safe operation range, the difference of historical information records is compared to determine the safety state level, and then the corresponding warning mode is triggered; specifically: S5.1, the support vector machine obtains the real-time energy storage system operation data based on the acquisition device, and generates monitoring data after performing preprocessing operations; The raw data collected by the acquisition equipment may have problems such as noise, missing data or inconsistent data formats. Preprocessing operations include data cleaning, removing obvious erroneous data and outliers; data completion, using interpolation and other methods to fill in missing data; data format conversion, converting different types of data into a format suitable for subsequent analysis. After these preprocessing steps, the monitoring data obtained can more accurately reflect the actual operating status of the energy storage system, providing a reliable data basis for subsequent status identification and judgment.

[0050] S5.2, the monitoring data reaches the upper limit of the preset safe operation range threshold, triggering the safety module, and the operation data is transmitted to the historical database to form an information record; The trigger mechanism of the safety module should have high sensitivity and low false alarm rate. The setting of the upper threshold should take into account the design parameters of the energy storage system, safety standards, and a large amount of experimental and actual operation data. Once triggered, the operation data should be stored in the historical database in a timely manner for subsequent analysis and comparison. The historical database should have good data storage and management functions, be able to quickly retrieve and call data, and ensure the integrity and security of the data to prevent data loss or tampering.

[0051] S5.3. Obtain all information records from the historical database as comparison objects, calculate the comparison objects and the current operation data, and obtain the degree of difference of the operation data (the degree of difference is calculated using the Euclidean distance formula, taking into account the differences in each data dimension); The Euclidean distance formula can effectively measure the distance between two data points in a multidimensional space. In this application, the current operating data and the data in the historical database are regarded as points in a multidimensional space. By calculating the Euclidean distance between them, we can intuitively understand the degree of deviation between the current data and the historical data. For example, if the current data is significantly different from the historical data in multiple dimensions such as current, voltage, and temperature, the calculated Euclidean distance will be larger, indicating that the system may be in an abnormal operating state.

[0052] S5.4. According to the pre-set grade rules corresponding to the degree of difference (the grade rules are divided into three levels according to the degree of difference: slight difference, medium difference, and severe difference, corresponding to different safety levels. The rules are determined by simulating a large amount of historical data and analyzing actual accident cases), determine which safety level the current monitoring data belongs to and generate a judgment result; S5.5. Obtain one of the multiple pre-configured warning modes according to the judgment result, and use the pre-established support vector machine for state recognition (the support vector machine is used for state recognition here mainly to further refine the analysis of the current state. The support vector machine parameters are consistent with the model parameters trained in S4. The current data is classified and judged again to improve the accuracy of state recognition); Different safety levels correspond to different warning modes, which can include sound alarms, light alarms, SMS notifications to maintenance personnel, and other forms. While issuing a warning, the support vector machine is used again to perform a detailed analysis of the current state to further confirm the abnormality of the system. Since the support vector machine has been trained in S4, its parameters have a certain degree of reliability and effectiveness. Using the model with the same parameters to classify and judge the current data again can improve the accuracy of state recognition and avoid misjudgment or missed judgment due to the limitations of a single judgment method.

[0053] S5.6, state identification outputs a state signal set, which includes multiple state signals; These status signals can reflect the operating status of the energy storage system in different aspects, such as battery health status, electrical connection status, thermal management status, etc. Through the comprehensive analysis of these status signals, we can have a more comprehensive understanding of the overall operation of the system and provide more information for further fault diagnosis and processing.

[0054] S5.7. If any status signal exceeds the preset safety range, the corresponding warning mode is activated to issue an alarm prompt; For each status signal, a corresponding safety range is set. When a status signal exceeds this range, the corresponding warning mode is immediately activated. For example, if the battery health status signal shows that the battery capacity drops sharply or the internal resistance increases abnormally, exceeding the preset safety range, it will trigger a warning mode specifically for battery failures, reminding the operation and maintenance personnel to check and maintain the battery in time.

[0055] S5.8. The system continuously obtains status signals. If it finds that the safety range is always out of limit, it continuously issues warnings and cyclically executes the support vector machine to perform status recognition operations and update the comparison object.

[0056] During the operation of the energy storage system, continuous monitoring of status signals is the key to ensuring system safety. If a status signal is always in an over-limit state, it means that there is a continuous safety hazard in the system and warnings need to be issued continuously until the problem is resolved. At the same time, in order to adapt to changes in the system's operating status, the support vector machine is periodically executed in a loop to perform state recognition operations and update the comparison object, that is, the latest historical data is used for comparative analysis with the current data, which can ensure the accuracy and timeliness of state recognition and judgment.

[0057] S6. By training the long short-term memory neural network model, using the existing data change trend data as features, extracting the time series features of the data change trend data and the monitoring data intensity, thereby determining the feature differences between the historical state data set and the current state data set, and outputting the normal operating range of each monitoring data of the system; specifically: S6.1. Obtain historical data according to each type of monitoring point and split the data. The first historical data set is used as a training set, the second historical data set is used as a test set, and the third historical data set is used as a validation set. During the data collection process, the integrity and accuracy of the data of various types of monitoring points must be ensured, covering all key parts and parameters of the energy storage system, such as the voltage, current, and temperature of the battery module, the efficiency and power factor of the power conversion device, etc. The data should be split in a reasonable proportion, such as the common 60% as a training set, 20% as a test set, and 20% as a validation set, so that historical data can be fully utilized for model training and evaluation, while ensuring that the model has good generalization ability.

[0058] S6.2. Perform data cleaning, data denoising, data transformation and feature extraction on the training set data. The features include trend, intensity and fluctuation range, and construct feature vectors to obtain the preliminary training set. Data cleaning and denoising can use filtering algorithms, outlier detection and correction methods to remove noise and erroneous data in the data and improve data quality. Data transformation can include operations such as logarithmic transformation and standardization to make the data more in line with the requirements of model training. Feature extraction is a key step. By analyzing the characteristics of data such as trend, intensity and fluctuation range, the inherent laws of the data can be captured more effectively. For example, for voltage data, its trend can reflect the change of battery charging and discharging state, the intensity can reflect the size of the voltage, and the fluctuation range can indicate the stability of the voltage. These features are combined into feature vectors to provide input data for the long short-term memory neural network model.

[0059] S6.3. Based on the preliminary training set data, a feature difference extraction model was constructed using the long short-term memory neural network algorithm. The LSTM model defined the number of hidden layers as 2 (determined based on the temporal complexity of the data and the model performance test), the number of neurons as 64 (determined through multiple experimental adjustments to effectively capture data features), and the learning rate as 0.01 (determined based on the trade-off between model convergence speed and accuracy). The objective function of the model was set to the mean square error function, and the number of iterations was 500, thus obtaining the first training model. S6.4. Use the validation set data to input the first training model to obtain a first prediction value, and obtain the accuracy rate through the historical normal range of the validation set historical data; Input the validation set data into the first training model to obtain the corresponding predicted value, and then compare it with the historical normal range of the validation set to calculate the correct prediction ratio, that is, the accuracy rate. If the accuracy rate is low, it means that the model may have problems such as overfitting or underfitting, and it is necessary to further adjust the model parameters or improve the model structure.

[0060] S6.5. If the model accuracy is lower than the threshold (the threshold is set to 90%, which is determined according to the system's requirements for data accuracy), the network structure parameters are modified, the learning rate is reset to 0.005 and the number of hidden layers is reset to 3, etc., and the first training model is updated to the second training model; return to step S6.4; If the model accuracy reaches the threshold, the first final model is obtained; Input the real-time monitoring data, obtain the data N minutes before the current time (N is set to 10 minutes based on the timeliness of the data and the system response requirements), and perform data preprocessing to obtain real-time trend data and intensity data; Input real-time trend data and intensity data, calculate the characteristic difference through the first final model, compare the normal range value, and determine whether the value of each monitoring point is normal or not.

[0061] S7. When the monitoring data deviates from the normal operating range after being judged by the long short-term memory neural network, it is included in the abnormal database, and the monitoring data intensity of rare phenomena in the abnormal database is fitted by the triple exponential smoothing method to obtain the probability density distribution of the change trend of the predicted data of the future time series; specifically: S7.1. Obtain the abnormal monitoring data stream after the long short-term memory network has determined, and determine that the current abnormal data stream has exceeded the preset normal data range; After the LSTM neural network determines that the monitoring data is abnormal, these abnormal data are collected in time to form an abnormal monitoring data stream. At the same time, it is necessary to clarify the specific circumstances in which these data exceed the preset normal data range, such as whether a certain parameter value is too high or too low, or multiple parameters are abnormal at the same time, which is of great significance for subsequent analysis and processing.

[0062] S7.2, judging the time point of occurrence of the current abnormal data according to the monitoring duration of the data point in the abnormal monitoring data stream; By recording the monitoring time of the data points, it is possible to determine when the abnormal phenomenon began to occur, which helps to analyze the development process and possible causes of the abnormality. For example, if the abnormal data suddenly appears after a certain period of time, it may be related to the change of the system's operating mode during that period of time or external environmental interference.

[0063] S7.3. Obtain the intensity value of the data point in the current abnormal data stream, use the intensity corresponding to all historical abnormal monitoring data points, and use the least squares method (determine the use of this method based on the distribution characteristics and fitting effect of historical data) to obtain the optimal values ​​of the three smoothing coefficients in the triple exponential smoothing method; S7.4. According to the cubic exponential smoothing method and the smoothing coefficient, the intensity corresponding to the abnormal monitoring data point that has occurred at the current moment is used to fit the intensity corresponding to all abnormal monitoring data points in the time period after the current moment; S7.5. Based on the intensity values ​​corresponding to all the fitted abnormal monitoring data points, the probability density distribution function of the intensity corresponding to all the abnormal monitoring data points is calculated. The calculation formula is: P(x)=(1 / (σ√(2π)))e^(-((x - μ)^2) / (2σ^2)), Where P(x) represents the probability density of data point x, μ represents the mean of the data point set, σ represents the standard deviation of the data point set, e represents the base of the natural logarithm, and π represents pi; S7.6. Obtain the probability density of the intensity corresponding to all abnormal monitoring data points obtained by calculation, and determine the extreme value interval of the abnormal data intensity by using the preset threshold method (the preset threshold is determined according to the risk assessment of the abnormal data intensity and the system safety requirements, for example, set to 0.05); S7.7. Based on the extreme value interval of abnormal data intensity and the current abnormal data monitoring time point, obtain the changing trend of this section of abnormal data at future moments.

[0064] S8. Calculate the standard deviation of the internal data for a period of time based on the results of rare phenomenon identification, and substitute it into the covariance kernel function of the Gaussian process for regression analysis, so as to obtain the functional relationship between the predicted data change trend and the change frequency of the physical quantity; specifically: S8.1. After obtaining the time series of historical data, identify rare events in the time series and obtain the standard deviation of the time series based on the results, and record the standard deviation as the first standard deviation; S8.2. After obtaining the time series of the first standard deviation and the corresponding time interval, perform data preprocessing, obtain the second time series by median filtering, obtain the moment of each data point in the second time series and mark it as the second moment; When analyzing historical data time series, use appropriate anomaly detection algorithms to identify rare events, such as anomaly detection based on statistical models, outlier detection based on machine learning, etc. Once a rare event is identified, calculate the standard deviation of the data in that time period. This standard deviation reflects the degree of dispersion of the data in this special case and record it as the first standard deviation for subsequent Gaussian process analysis.

[0065] S8.3. Using the second time series and the corresponding second moment, take the data point as the center, take the 3 data points before and after (determine the neighborhood range based on the local correlation and noise characteristics of the data), calculate its standard deviation as the standard deviation of the neighborhood near each data point at the second moment, and substitute the standard deviation of all neighborhoods together with the first standard deviation into the covariance kernel function of the Gaussian process. Among them and Represents data at different time points in a time series. As an independent variable, As the dependent variable, , is the standard deviation, It is the characteristic length of time length; S8.4. According to the covariance kernel function, a Gaussian process regression model is constructed, and a regression analysis is performed on the processed second time series to obtain an initial function model between the time series change and the frequency of the physical quantity change, and the model is constructed based on the regression analysis and the second time series; S8.4. According to the covariance kernel function, a Gaussian process regression model is constructed, and a regression analysis is performed on the processed second time series to obtain an initial function model between the time series change and the frequency of the physical quantity change, and the model is constructed based on the regression analysis and the second time series; S8.5. According to the initial function model, the processed time series is segmented, a time length is set (for example, 5 minutes) to obtain a third moment and a plurality of third time series, and a standard deviation and a change frequency of each third time series are calculated to generate a frequency feature vector group; S8.6. Perform forecast analysis according to the initial function model, input the elements in the frequency feature vector group into the initial function model respectively, obtain the fourth time series and the corresponding fourth moment, and calculate the difference between the fourth moment series and the third moment series to obtain the frequency error; S8.7. When the frequency error is greater than the set threshold (the threshold is determined based on the system's requirements for prediction accuracy and the sensitivity of the frequency of physical quantity changes, for example, it is set to 0.1 Hz), more time series data are collected to repeat the processing from the first step to the sixth step, and the initial function model is updated with the new time series. If the frequency error is less than the set threshold, the future time series is predicted based on the initial function model.

[0066] S9. Perform weighted fusion based on the probability density distribution of the predicted data change trend and the relationship between the frequency of physical quantity changes, perform surface integration of the function graph using the adaptive Simpson algorithm, and determine whether there is a potential safety risk based on the size of the area of ​​the integral result. Specifically: S9.1. A matrix of signals in the acquisition device is established based on multiple related attribute data in the device operation timing data stream. A row or a column in the matrix is ​​called a vector of the signal. The signal amplitude value at each acquisition moment constitutes a vector component. A discrete Fourier transform is performed on a vector to obtain information data representing the frequency of change of the physical quantity. S9.2. Obtain the modulus data of the vector after discrete Fourier transformation and determine the information entropy value of each data according to the Shannon information entropy calculation method, use the maximum entropy of all data at multiple collection times as the measurement value of the risk weight, and then obtain the fused value of each vector in each column or row of the matrix; S9.3, perform principal component analysis and eigenvalue decomposition on the matrix to obtain the importance weights of the main vector groups of the signal represented by eigenvalues ​​of different sizes arranged in descending order, and obtain the normalization factor of the weighted eigenvector group after accumulating the weights according to all eigenvalues ​​in proportion; S9.4. Collect and obtain the density function that represents the probability of the change trend of the predicted data. The density function input uses the normalization factor of the weight feature vector group obtained in the previous step as the input. After the function output of the probability density function and the weight are weighted, if there is a distribution offset of the predicted density function, then the gradient descent method (this method is determined based on the mathematical characteristics of the probability density function and the optimization goal) is used to iteratively and adaptively correct the parameters in the probability density function. S9.5. By obtaining the data parameters of the prediction probability density function in the previous step, the upper limit and lower limit of the integral interval in the probability density function are determined, and the integral function body based on the adaptive Simpson algorithm is obtained; S9.6. Determine the relationship between the precision value set when obtaining the function area by integration (the precision value is initially set to 0.01 and can be adjusted according to the computing resources and the accuracy requirements of risk judgment) and the threshold value for terminating the integration. If the precision value is greater than the threshold, obtain the value of the lower half of the current precision, use the value as the adaptive value of the Simpson algorithm, and calculate the function in this step to obtain the sub-interval area. The multiple sub-interval areas obtained in this step are combined to obtain the integral result of this step. S9.7. Based on the multiple result values ​​obtained from all the above calculations, determine the specific location and device attribute information corresponding to the different device location data represented by each data result. The signal matrix of all device locations at different times constitutes a tensor. The risk value of different components and positions of the device in a certain time series is calculated using the tensor to make a judgment. Judge the value and the risk threshold. If the risk value is greater than or equal to the risk threshold, it is determined that the corresponding component has potential risk.

[0067] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A method for controlling false alarms of fire in an energy storage system, characterized in that The following steps are involved: S1. Obtain the current value, voltage value and power value data stream transmitted by the energy storage system sensor in real time, calculate the fluctuation period of the system operation speed through fast Fourier transform, calculate the second-order derivative of the fluctuation period, and obtain the probability density function of the data change trend; S2. According to the probability density function of the data change trend, the quantiles of the normal operation range under different states are set, and combined with the pre-set empirical upper threshold of the normal operation range, if the probability of the data change trend exceeds the upper threshold, the alarm mechanism is triggered; S3, collecting the thermal energy changes of data streams at multiple time points, obtaining the first-order derivative distribution of thermal energy data with respect to time, obtaining the second-order derivative curve of temperature change data with respect to time, and constructing a thermal energy change rate index data set; S4. Train the support vector machine model based on the heat energy change rate index data set, extract the classification boundary between normal state data and abnormal state data, and build a state recognition mechanism based on the weight ratio of each state; S5. Based on the state recognition mechanism established by the support vector machine, when the monitoring data of the energy storage system triggers the upper threshold limit of the preset safe operation range, the difference of historical information records is compared to determine the safety state level, and then the corresponding warning mode is triggered; S6. By training the long short-term memory neural network model, using the existing data change trend data as features, extracting the time series features of the data change trend data and the monitoring data intensity, thereby determining the feature differences between the historical state data set and the current state data set, and outputting the normal operating range of each monitoring data of the system; S7. When the monitoring data deviates from the normal operating range after being judged by the long short-term memory neural network, the monitoring data is included in the abnormal database, and the monitoring data intensity of rare phenomena in the abnormal database is fitted by the triple exponential smoothing method to obtain the probability density distribution of the change trend of the predicted data of the future time series; S8. Calculate the standard deviation of the internal data for a period of time based on the results of rare phenomenon identification, and substitute it into the covariance kernel function of the Gaussian process for regression analysis, so as to obtain the functional relationship between the predicted data change trend and the change frequency of the physical quantity; S9. Perform weighted fusion based on the probability density distribution of the predicted data change trend and the relationship between the frequency of change of the physical quantity, perform surface integral of the function graph through the adaptive Simpson algorithm, and judge whether there is a potential safety risk based on the size of the area of ​​the integral result.

2. A method for controlling false alarms of fire in an energy storage system according to claim 1, characterized in that: The sub-steps of step S1 are: S1.1, collect the current value, voltage value and power value data stream output by the energy storage system sensor; S1.2, calculate the fast Fourier transform of the three parameter data streams of current value, voltage value and power value, obtain the frequency domain component signal, and determine the system fluctuation period according to the corresponding relationship between the specific frequency component in the frequency domain component signal and the fluctuation period; S1.3, calculating the second-order derivative value based on the acquired fluctuation period information to obtain the fluctuation state change rate index; S1.

4. Obtain the probability density function based on the numerical calculation of the fluctuation state change rate to determine the statistical distribution probability of the system operating in each fluctuation range; S1.

5. Determine the current energy storage state of the energy storage system through the existing relationship set between the probability density function and the energy storage state; S1.

6. Use the known voltage and current data and the autoregressive moving average algorithm to predict the voltage and current trends over a period of time. S1.

7. The obtained voltage and current trends are input into the support vector machine to determine the output power value under the preset period after the corresponding energy storage system changes.

3. A method for controlling false alarms of fire in an energy storage system according to claim 1, characterized in that: The sub-steps of step S2 are: S2.

1. Collect state data to build a data set. The data set includes historical data under multiple different states. The state data is input into a probability density function model to obtain a probability density function corresponding to the data change trend and obtain a probability value. S2.

2. Using the obtained probability density function, according to the preset quantile value, reversely solve the quantile value to the function, obtain the normal operation range corresponding to each state, obtain the corresponding upper and lower thresholds, and construct the range interval of different quantile points corresponding to each state; S2.3, obtain the state data and calculate, calculate the statistical result of the current data by histogram statistics, and use the statistical result to determine the probability of obtaining the state data; S2.

4. Compare the probability of the current state data with the probability corresponding to different quantiles of the state to make a threshold judgment; S2.

5. If the current state probability value is greater than the occurrence probability corresponding to the preset quantile threshold, an alarm signal is triggered, and the current state data and related time information are recorded for subsequent analysis.

4. A method for controlling false alarms of fire in an energy storage system according to claim 1, characterized in that: The sub-steps of step S3 are: S3.

1. The time domain sensor records and obtains the data stream of the thermal energy value of the object at multiple times, calculates the difference of the thermal energy values ​​at adjacent times, and obtains the first-order derivative of the thermal energy with respect to time; S3.2, using the obtained thermal energy difference data of adjacent moments divided by the time difference of the moments, to obtain temperature change data; S3.3, using the differential method to obtain the temperature change data, and obtain the second-order derivative curve of temperature with respect to time; S3.

4. Count the frequency of the second-order derivative of temperature exceeding the preset range within the preset time, and construct a thermal energy change rate index reflecting the dynamic fluctuation amplitude of thermal energy; S3.

5. Perform time domain signal analysis on the acquired thermal energy values ​​and calculate the Fourier transform of its frequency components; for discrete thermal energy signal time series , is the sequence length, and its discrete Fourier transform The expression is: ； in, Corresponding to different discrete frequency points, and = 0, 1, 2, ..., N - 1; is an imaginary unit, satisfying , is the ratio of pi; the spectrum components of the thermal energy signal at different discrete frequency points can be obtained through the discrete Fourier transform; then these spectrum components are analyzed to determine the proportion of high-frequency components; if the high-frequency components exceed the predetermined proportion threshold, the next step is executed: In the set constructed by the second-order derivative values ​​corresponding to the temperature changes at different times, the outlier detection LOF algorithm is used to detect and identify the moments that deviate from other data, where is the time series of thermal energy signal, is the interval between moments; If it is a continuous thermal energy signal time series , whose Fourier transform The expression is: Similarly, through The analysis determines the proportion of high-frequency components, and the subsequent processing is the same as the discrete case; is the continuous frequency; S3.

6. If the moment is identified as abnormal, execute the next step: The thermal energy change data is imported into the trained deep convolutional adversarial neural network model. The model uses an encoder to map the thermal energy features into latent space vectors, and the discriminator distinguishes the probability that the latent space vector is real data.

5. A method for controlling false alarms of fire in an energy storage system according to claim 1, characterized in that: The sub-steps of step S4 are: S4.

1. Obtain a data set of thermal energy change rate according to the time series of the collected thermal energy data; S4.2, normalizing all data of the heat energy change rate data set to obtain a normalized data set, and then setting a normal state data mark and an abnormal state data mark according to preset conditions; S4.3, obtaining a labeled normal state and abnormal state data set from the normalized data set, training a support vector machine model, and obtaining an optimal support vector machine classification model; S4.

4. Using all the marked normal state data and abnormal state data, obtain the classification boundary hyperplane equation of the support vector machine classification model , represents the rate of change of thermal energy, is the weight vector, is the bias term; S4.

5. Obtain the number of occurrences of the normal state and the abnormal state of the marked data by statistics, and obtain the ratio of the number of occurrences of the two, and determine the state weight ratio value; S4.

6. According to the support vector machine classification model and the state weight ratio value, the state of the real-time collected heat energy change rate data is judged. If the heat energy change rate is greater than the classification boundary value and the number of abnormal heat energy states is greater than the state weight ratio value, a warning state is obtained; S4.

7. When the thermal energy status information corresponding to the thermal energy change rate data is collected in real time, the associated equipment type, environmental information and process parameter data are extracted from the database for data fusion, and the equipment operation status and maintenance information are obtained based on the above data.

6. A method for controlling false alarms of fire in an energy storage system according to claim 1, characterized in that: The sub-steps of step S5 are: S5.1, the support vector machine obtains the real-time energy storage system operation data based on the acquisition device, and generates monitoring data after performing preprocessing operations; S5.2, the monitoring data reaches the upper limit of the preset safe operation range threshold, triggering the safety module, and the operation data is transmitted to the historical database to form an information record; S5.

3. Obtain all information records from the historical database as comparison objects, calculate the comparison objects and current operation data, and obtain the degree of difference in the operation data; S5.4, according to the pre-set level rules corresponding to the degree of difference, determine which security level the current monitoring data belongs to, and generate a judgment result; S5.5, obtaining one of a plurality of different pre-configured warning modes according to the judgment result, and using a pre-established support vector machine to perform state recognition; S5.6, state identification outputs a state signal set, which includes multiple state signals; S5.

7. If any status signal exceeds the preset safety range, the corresponding warning mode is activated to issue an alarm prompt; S5.

8. The system continuously obtains status signals. If it finds that the safety range is always out of limit, it continuously issues warnings and cyclically executes the support vector machine to perform status recognition operations and update the comparison object.

7. A method for controlling false alarms of fire in an energy storage system according to claim 1, characterized in that: The sub-steps of step S6 are: S6.

1. Obtain historical data according to each type of monitoring point and split the data. The first historical data set is used as a training set, the second historical data set is used as a test set, and the third historical data set is used as a validation set. S6.

2. Perform data cleaning, data denoising, data transformation and feature extraction on the training set data. The features include trend, intensity and fluctuation range, and construct feature vectors to obtain the preliminary training set. S6.

3. Based on the preliminary training set data, a feature difference extraction model was constructed using the long short-term memory neural network algorithm. The LSTM model defined the number of hidden layers as 2, the number of neurons as 64, and the learning rate as 0.

01. The objective function of the model was set to the mean square error function, and the number of iterations was 500 to obtain the first training model. S6.

4. Use the validation set data to input the first training model to obtain a first prediction value, and obtain the accuracy rate through the historical normal range of the validation set historical data; S6.

5. If the model accuracy is lower than the threshold, the network structure parameters are modified, the learning rate is reset to 0.005 and the number of hidden layers is reset to 3, and the first training model is updated to the second training model; Return to step S6.4; If the model accuracy reaches the threshold, the first final model is obtained; Input the real-time monitoring data, obtain the data N minutes before the current time, and perform data preprocessing to obtain real-time trend data and intensity data; Input real-time trend data and intensity data, calculate the characteristic difference through the first final model, compare the normal range value, and determine whether the value of each monitoring point is normal or not.

8. A method for controlling false alarms of fire in an energy storage system according to claim 1, characterized in that: The sub-steps of step S7 are: S7.

1. Obtain the abnormal monitoring data stream after the long short-term memory network has determined that the current abnormal data stream has exceeded the preset normal data range; S7.2, judging the time point of occurrence of the current abnormal data according to the monitoring duration of the data point in the abnormal monitoring data stream; S7.3, obtaining the intensity value of the data point in the current abnormal data stream, using the intensity corresponding to all historical abnormal monitoring data points, and obtaining the optimal values ​​of the three smoothing coefficients in the cubic exponential smoothing method through the least squares method; S7.

4. According to the cubic exponential smoothing method and the smoothing coefficient, the intensity corresponding to the abnormal monitoring data point that has occurred at the current moment is used to fit the intensity corresponding to all abnormal monitoring data points in the time period after the current moment; S7.

5. Based on the intensity values ​​corresponding to all the fitted abnormal monitoring data points, the probability density distribution function of the intensity corresponding to all the abnormal monitoring data points is calculated. The calculation formula is: P(x)=(1 / (σ√(2π)))e^(-((x - μ)^2) / (2σ^2)), Where P(x) represents the probability density of data point x, μ represents the mean of the data point set, σ represents the standard deviation of the data point set, e represents the base of the natural logarithm, and π represents pi; S7.

6. Obtain the probability density of the intensity corresponding to all abnormal monitoring data points obtained by calculation, and determine the extreme value interval of the abnormal data intensity by using the preset threshold method; S7.

7. Based on the extreme value interval of abnormal data intensity and the current abnormal data monitoring time point, obtain the changing trend of this section of abnormal data at future moments.

9. A method for controlling false alarms of fire in an energy storage system according to claim 1, characterized in that: The sub-steps of step S8 are: S8.

1. After obtaining the time series of historical data, identify rare events in the time series and obtain the standard deviation of the time series based on the results, and record the standard deviation as the first standard deviation; S8.

2. After obtaining the time series of the first standard deviation and the corresponding time interval, perform data preprocessing, obtain the second time series by using the median filtering method, obtain the moment of each data point in the second time series and mark it as the second moment; S8.

3. Using the second time series and the corresponding second moment, with the data point as the center, take the three data points before and after to calculate their standard deviation as the standard deviation of the neighborhood near each second moment data point, and substitute the standard deviation of all neighborhoods together with the first standard deviation into the covariance kernel function of the Gaussian process Among them and Represents data at different time points in a time series. As an independent variable, As the dependent variable, ， is the standard deviation, It is the characteristic length of time length; S8.4, constructing a Gaussian process regression model based on the covariance kernel function, and performing regression analysis on the processed second time series, thereby obtaining an initial function model between the time series change and the frequency of physical quantity change, the model being constructed based on the regression analysis and the second time series; S8.

5. According to the initial function model, the processed time series is segmented, a time length is set to obtain a third moment and a plurality of third time series, and a standard deviation and a change frequency of each third time series are calculated to generate a frequency feature vector group; S8.

6. Perform prediction analysis based on the initial function model, input the elements in the frequency feature vector group into the initial function model respectively, obtain the fourth time series and the corresponding fourth moment, and calculate the difference between the fourth moment series and the third moment series to obtain the frequency error.

10. A method for controlling false alarms of fire in an energy storage system according to claim 1, characterized in that: The sub-steps of step S9 are: S9.

1. A matrix of signals in the acquisition device is established based on multiple related attribute data in the device operation timing data stream. A row or a column in the matrix is ​​called a vector of the signal. The signal amplitude value at each acquisition moment constitutes a vector component. A discrete Fourier transform is performed on a vector to obtain information data representing the frequency of change of the physical quantity. S9.

2. Obtain the modulus data of the vector after discrete Fourier transformation and determine the information entropy value of each data according to the Shannon information entropy calculation method, use the maximum entropy of all data at multiple collection times as the measurement value of the risk weight, and then obtain the fused value of each vector in each column or row of the matrix; S9.3, perform principal component analysis and eigenvalue decomposition on the matrix to obtain the importance weights of the main vector groups of the signal represented by eigenvalues ​​of different sizes arranged in descending order, and obtain the normalization factor of the weighted eigenvector group after accumulating the weights according to all eigenvalues ​​in proportion; S9.

4. Collect and obtain the density function that represents the probability of the change trend of the predicted data. The density function input uses the normalization factor of the weight feature vector group obtained in the previous step as the input. After the function output of the probability density function is weighted and calculated with the weight, if there is a distribution offset of the predicted density function, the gradient descent method is used to iteratively and adaptively correct the parameters in the probability density function for the offset phenomenon. S9.

5. By obtaining the data parameters of the prediction probability density function in the previous step, the upper limit and lower limit of the integral interval in the probability density function are determined, and the integral function body based on the adaptive Simpson algorithm is obtained; S9.

6. Determine the relationship between the set precision value and the preset threshold value for terminating the integral when the function area is obtained by integration. If the precision value is greater than the threshold, obtain the value of the lower half of the current precision, use the value as the adaptive value of the Simpson algorithm, and calculate the function in this step to obtain the sub-interval area. The multiple sub-interval areas obtained in this step are combined to obtain the integral result of this step. S9.

7. Based on the multiple result values ​​obtained from all the above calculations, determine the specific location and device attribute information corresponding to the different device location data represented by each data result. The signal matrix of all device locations at different times constitutes a tensor. The risk value of different components and positions of the device in a certain time series is calculated using the tensor to make a judgment. Judge the value and the risk threshold. If the risk value is greater than or equal to the risk threshold, it is determined that the corresponding component has potential risk.

Images