Energy storage battery system insulation failure early warning method based on deep learning time sequence prediction

Abstract

The invention discloses an energy storage battery system insulation failure early warning method based on deep learning time sequence prediction. The method comprises the following steps: constructing an insulation detection circuit; bimodal voltage sampling is realized by periodically switching bridge arm resistors, and a voltage time sequence data set is formed; performing feature extraction and recursive prediction on the voltage time sequence data based on a CNN-LSTM-Attention deep learning model, and outputting a voltage predicted value; establishing an evaluation index and a threshold value of a voltage prediction value, and realizing rapid diagnosis and early warning of insulation failure; performing parameter identification on the abnormal voltage data by adopting a self-adaptive forgetting factor recursive least square method; and inversion calculation is carried out on positive and negative electrode ground insulation resistance values through a capacitor charge and discharge characteristic equation, and a grading early warning mechanism is established to carry out insulation failure early warning. The method can shorten the detection period, reduce the calculation error of the insulation resistance value, realize rapid fault response, and effectively solve the problems of large detection blind area and response lag in the prior art.

Description

Technical Field

[0001] The present invention relates to the technical field of insulation failure detection, and particularly to a method for warning insulation failure of an energy storage battery system based on deep learning time series prediction. Background Art

[0002] The existing insulation detection technologies for energy storage batteries mainly have the following defects: 1) The balanced bridge method has a detection blind area and cannot identify the same proportion of insulation decline in the positive and negative poles; 2) The unbalanced bridge method is affected by the charging and discharging of Y capacitors, and the detection period is as long as several seconds, making it difficult to meet the requirements of rapid response; 3) The AC injection method requires additional hardware and has poor anti-interference ability. In addition, traditional methods rely on steady-state voltage values to calculate insulation resistance, and there is a significant lag in dynamic working conditions (such as the non-linear sudden drop of insulation resistance value at the initial stage of battery leakage). For example, when electrolyte leaks from the battery shell, the insulation resistance value may drop from 10 MΩ to the kΩ level within 6 seconds. However, due to the need to switch the sampling period multiple times, conventional detection algorithms often delay fault judgment, resulting in the protection action lagging behind the thermal runaway trigger point. In the prior art, algorithms such as Kalman filtering are tried to optimize the detection speed, but the parameter identification accuracy is insufficient in the face of complex working conditions, and the resistance value calculation error still exceeds 15%.

[0003] Therefore, there is an urgent need to develop an insulation failure warning technology with fast response, high-precision identification, and strong anti-interference ability to solve the core problems of detection blind area, response delay, and high false alarm rate in the existing technology, and to meet the protection requirements of high-safety-level energy storage systems. Summary of the Invention

[0004] In order to overcome the defects and deficiencies existing in the prior art, the present invention provides a method for warning insulation failure of an energy storage battery system based on deep learning time series prediction, which solves the problems such as detection blind area, response delay (≥6 seconds), and large resistance value calculation error (>15%) existing in the existing insulation detection technology for energy storage batteries, and realizes millisecond-level fault identification, resistance value calculation error within 5%, and hierarchical rapid protection.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] The present invention provides a method for warning insulation failure of an energy storage battery system based on deep learning time series prediction, including the following steps:

[0007] Construct an insulation detection circuit, including adjustable bridge arm resistors and a bidirectional switch matrix, and dynamically switch the bridge arm resistors through the bidirectional switch matrix;

[0008] Periodically switch the resistance value configuration of the bridge arm resistors at a preset frequency, and collect the positive pole-to-ground voltage and the negative pole-to-ground voltage in a dual-mode to form a voltage time series data set;

[0009] Input the voltage time series dataset into the pre-trained CNN-LSTM-Attention deep learning model. Extract spatial features through the convolutional neural network, capture temporal dependencies through the long short-term memory network, and dynamically weight key features through the attention mechanism to output the voltage prediction value;

[0010] Construct evaluation indicators and their thresholds for the voltage prediction value to detect whether the energy storage battery system has insulation failure;

[0011] Collect the abnormal voltage data of the sampling points within the set time after exceeding the evaluation indicator threshold. Based on the function model of the voltage of the capacitor during charging and discharging at time t, use the adaptive forgetting factor recursive least squares method to identify the parameters of the abnormal voltage data and obtain the voltage stable value;

[0012] Invert and calculate the insulation resistance values of the positive and negative electrodes to the ground according to the capacitor charging and discharging characteristic equation;

[0013] Establish a hierarchical warning mechanism, set the hierarchical threshold of the insulation resistance value, cut off the main circuit after triggering, and start the insulation failure warning.

[0014] As a preferred technical solution, in the CNN-LSTM-Attention deep learning model, the output of the convolutional neural network is expressed as:

[0015]

[0016] Among them, represents the output of the convolutional neural network, f cov (·) represents the activation function, k represents the sliding window size, w n,m represents the weight of the convolutional kernel with n rows and m columns, X i+n,j+m represents the value of the nth row and mth column of the input voltage data feature matrix, b n,m represents the convolutional kernel bias;

[0017] Input the output of the convolutional neural network into the long short-term memory network to obtain the intermediate state vector of the voltage data;

[0018] The attention mechanism calculates the similarity between the intermediate state feature vector and the hidden vector of the voltage data, calculates the attention weight of the hidden layer vector of the voltage data, and weights it with the intermediate state feature vector to obtain the output of the attention layer, and outputs the predicted value of the sampled voltage through the fully connected layer.

[0019] As a preferred technical solution, the attention mechanism calculates the similarity between the intermediate state feature vector and the hidden vector of the voltage data, specifically through the function score([h t,i ,h t ) to calculate the similarity between the intermediate state feature vector and the hidden vector of the voltage data, which is expressed as:

[0020]

[0021] Among them, h t,i represents the intermediate state feature vector, and h t represents the hidden vector, W s represents the weight matrix of the fully connected layer, and b s represents the bias vector;

[0022] Calculate the attention weights of the voltage data hidden layer vector, and weight it with the intermediate state feature vector to obtain the output of the attention layer, which is expressed as:

[0023]

[0024] Among them, represents the output of the attention layer, α i represents the attention weight, and τ represents the output node of the fully connected layer.

[0025] As a preferred technical solution, based on the function model of the voltage of the capacitor during charging and discharging at time t, the adaptive forgetting factor recursive least squares method is used to identify the parameters of the abnormal voltage data, specifically including:

[0026] The function model of the voltage of the capacitor during charging and discharging at time t is expressed as:

[0027]

[0028] Among them, V ∞ is the sampled point voltage when the capacitor is stable, V0 is the initial voltage, RC is the time constant, and t is the time;

[0029] Convert the function model into the discrete-time function of U f which is expressed as:

[0030]

[0031] Among them, Δt is the sampling time interval, and U f (k) represents the reflected wave voltage value at time k;

[0032] Let U f (k) = F(A, k), where A represents the model parameters, and A = [a1 a2 a3];

[0033] Set the initial parameter values The initial value of the error covariance The forgetting factor Construct the observation matrix H(k):

[0034]

[0035] Calculate the prediction error e(k):

[0036]

[0037] Calculate the gain matrix K(k):

[0038] K(k) = P(k - 1)·H T (k)·[λ(k - 1) + H(k)·P(k - 1)·H T (k)] -1

[0039] Update the parameter estimate

[0040]

[0041] Update the covariance matrix P(k):

[0042]

[0043] Update the forgetting factor λ(k):

[0044] When λ(k) = max[0.9, λ(k - 1) - 0.01];

[0045] When λ(k) = min[0.99, λ(k - 1) + 0.01].

[0046] As a preferred technical solution, inversely calculate the insulation resistance values of the positive and negative electrodes to the ground according to the capacitance charge-discharge characteristic equation, specifically including:

[0047] Close the main control switch S1 to connect the insulation detection circuit;

[0048] Open the on-off switch S2 and close the on-off switch S3. At this time, the Y capacitor C p discharges, and the Y capacitor C n charges, and wait for the time T1;

[0049] It is known that the voltage at the sampling point A1 is U A1 , calculate the voltages U 12 across the resistors R1 and R2. It is known that the total battery voltage is U A2 , and the following formula can be obtained:

[0050]

[0051] Close the on-off switch S2 and open the on-off switch S3. At this time, the Y capacitor C p charges, and the Y capacitor C n discharges, and wait for the time T2;

[0052] It is known that the second voltage at the voltage sampling point A1 is U A '1, calculate the voltage U1'2 between the resistor R1 and the resistor R2, knowing that the total battery voltage U A ′2, the following formula is obtained:

[0053]

[0054] The combined formulas are used to obtain the insulation resistance R of the high voltage to ground. p And the high voltage negative insulation resistance to ground R n , the result is as follows:

[0055]

[0056] Among them, the resistor R1, the resistor R2, the resistor R3, and the resistor R4 are adjustable bridge arm resistors and are in an asymmetric configuration.

[0057] The present invention also provides an insulation failure warning system for an energy storage battery system based on deep learning time series prediction, comprising: an insulation detection circuit construction module, a voltage time series data set construction module, a voltage prediction value output module, an evaluation index construction module, a parameter identification module, an insulation resistance value calculation module, and an insulation failure warning module;

[0058] The insulation detection circuit construction module is used to construct an insulation detection circuit, including an adjustable bridge arm resistor and a bidirectional switch matrix, and the bridge arm resistor is dynamically switched through the bidirectional switch matrix; the voltage time series data set construction module is used to periodically switch the resistance configuration of the bridge arm resistor at a preset frequency, collect the positive pole-to-ground voltage and the negative pole-to-ground voltage in a dual mode, and form a voltage time series data set; the voltage prediction value output module is used to input the voltage time series data set into a pre-trained CNN-LSTM-Attention deep learning model, extract spatial features through a convolutional neural network, capture temporal dependencies through a long short-term memory network, and dynamically weight key features through an attention mechanism, and output a voltage prediction value; The evaluation index construction module is used to construct the evaluation index and threshold of the voltage prediction value, and detect whether the energy storage battery system has insulation failure; the parameter identification module is used to collect the abnormal voltage data of the sampling point within the set time after exceeding the evaluation index threshold, and based on the function model of the voltage of the capacitor under time t when charging and discharging, the adaptive forgetting factor recursive least squares method is used to perform parameter identification on the abnormal voltage data to obtain the voltage stability value; the insulation resistance calculation module is used to inversely calculate the insulation resistance of the positive and negative electrodes to the ground according to the capacitor charging and discharging characteristic equation; the insulation failure warning module is used to establish a hierarchical warning mechanism, set the insulation resistance hierarchical threshold, cut off the main circuit after triggering, and start the insulation failure warning.

[0059] As a preferred technical solution, in the CNN-LSTM-Attention deep learning model, the output of the convolutional neural network is expressed as:

[0060]

[0061] Wherein, represents the output of the convolutional neural network, f cov (·) represents the activation function, k represents the sliding window size, w n,m represents the weight of the convolutional kernel with n rows and m columns, X i+n,j+m represents the value of the nth row and mth column of the input voltage data feature matrix, b n,m represents the convolutional kernel bias;

[0062] The output of the convolutional neural network is input into the long short-term memory network to obtain the intermediate state vector of the voltage data;

[0063] The attention mechanism calculates the similarity between the intermediate state feature vector and the hidden vector of the voltage data, calculates the attention weight of the hidden layer vector of the voltage data, and weights it with the intermediate state feature vector to obtain the output of the attention layer, and outputs the predicted value of the sampled voltage through the fully connected layer.

[0064] As a preferred technical solution, the attention mechanism calculates the similarity between the intermediate state feature vector and the hidden vector of the voltage data, specifically calculates the similarity between the intermediate state feature vector and the hidden vector of the voltage data through the function score([h t,i ,h t ) and is expressed as:

[0065]

[0066] Wherein, h t,i represents the intermediate state feature vector, h t represents the hidden vector, W s represents the weight matrix of the fully connected layer, b s represents the bias vector;

[0067] The attention weight of the hidden layer vector of the voltage data is calculated and weighted with the intermediate state feature vector to obtain the output of the attention layer, which is expressed as:

[0068]

[0069] Wherein, represents the output of the attention layer, α i represents the attention weight, and τ represents the output node of the fully connected layer.

[0070] As a preferred technical solution, based on the function model of the voltage of the capacitor over time t during charging and discharging, the adaptive forgetting factor recursive least squares method is used to identify the parameters of the abnormal voltage data, specifically including:

[0071] The function model of the voltage of the capacitor over time t during charging and discharging is expressed as:

[0072]

[0073] where V ∞ is the voltage at the sampling point when the capacitor is stable, V0 is the initial voltage, RC is the time constant, and t is the time;

[0074] The function model is transformed into a discrete-time function of U f and is expressed as:

[0075]

[0076] where Δt is the sampling time interval, and U f (k) represents the reflected wave voltage value at time k;

[0077] Let U f (k) = F(A, k), where A represents the model parameters and A = [a1 a2 a3];

[0078] Set the initial parameter values Initial value of the error covariance Forgetting factor Construct the observation matrix H(k):

[0079]

[0080] Calculate the prediction error e(k):

[0081]

[0082] Calculate the gain matrix K(k):

[0083] K(k) = P(k - 1) · H T (k) · [λ(k - 1) + H(k) · P(k - 1) · H T (k)] -1

[0084] Update the parameter estimation value

[0085]

[0086] Update the covariance matrix P(k):

[0087]

[0088] Update forgetting factor λ(k):

[0089] When λ(k) = max[0.9, λ(k - 1) - 0.01];

[0090] When λ(k) = min[0.99, λ(k - 1) + 0.01].

[0091] As a preferred technical solution, the insulation resistance values of the positive and negative electrodes to the ground are inversely calculated according to the capacitance charge and discharge characteristic equation, specifically including:

[0092] Close the main control switch S1 to connect the insulation detection circuit;

[0093] Disconnect the on-off switch S2 and close the on-off switch S3. At this time, the Y capacitor C p discharges, and the Y capacitor C n charges, and wait for the time T1;

[0094] Given that the voltage at the sampling point A1 is U A1 , calculate the voltages U 12 across the resistors R1 and R2. Given the total battery voltage U A2 , the following formula can be obtained:

[0095]

[0096] Close the on-off switch S2 and disconnect the on-off switch S3. At this time, the Y capacitor C p charges, and the Y capacitor C n discharges, and wait for the time T2;

[0097] Given that the second voltage at the voltage sampling point A1 is U A ′1, calculate the voltages U1′2 across the resistors R1 and R2. Given the second collected total battery voltage U A ′2, the following formula can be obtained:

[0098]

[0099] Solve the equations simultaneously to obtain the insulation resistance value R p of the high-voltage positive electrode to the ground and the insulation resistance value R n of the high-voltage negative electrode to the ground. The results are as follows:

[0100]

[0101] Among them, the resistors R1, R2, R3, and R4 are adjustable bridge arm resistors and are configured asymmetrically.

[0102] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0103] (1) The present invention completely solves the problem of missed detection in the traditional balanced bridge method when the insulation of the positive and negative poles decreases proportionally through the dynamic bridge arm switching of the unbalanced bridge (asymmetric configuration of R1, R2, R3, and R4), and the detection blind area elimination rate is 100%;

[0104] (2) Based on the recursive prediction of the CNN-LSTM-Attention model, the present invention can complete the voltage trend prediction for the next 6 seconds instantaneously (the traditional method requires 6 seconds of steady-state sampling). Combining the three-level evaluation indicators (RMSE / MAE / R 2 ), it can determine the fault in real time, and the response time from insulation failure to warning trigger is ≤0.8 seconds, with the response speed increased by 83% compared to the traditional method (≥6 seconds).

[0105] (3) The adaptive forgetting factor recursive least squares method (AFFRLS) is used to dynamically identify the capacitance charge and discharge parameters, and the resistance value inversion is completed within 0.5 seconds. The calculation error of the insulation resistance is ≤5% (the traditional method has an error of 15%-30%), meeting the high-precision detection requirements in IEC61557-8 standard;

[0106] (4) Combining the feature extraction ability of the deep learning model for noise data, in the Gaussian white noise environment with a noise power of 0.001, it still maintains a prediction accuracy of 99.9% (R 2 = 99.9%), enhancing the anti-interference ability.

[0107] (5) The present invention is based on a hierarchical response mechanism based on the resistance value threshold. After the secondary warning is triggered, the main circuit is cut off, optimizing the hierarchical protection mechanism.

[0108] (6) Through the optimized design of the bridge arm resistors (R1 = 30 kΩ, R2 = 30 kΩ, R3 = 30 MΩ, R4 = 3 MΩ), the present invention replaces the isolation power supply module of the AC injection method, reducing the hardware cost by 40% while maintaining the detection accuracy, and is suitable for distributed deployment in large-scale energy storage power stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0109] Figure 1 is a schematic structural diagram of the insulation detection circuit of the present invention;

[0110] Figure 2 is a schematic diagram of the sampling voltage waveform of the insulation detection circuit of the present invention;

[0111] Figure 3 is a schematic diagram of the architecture of the CNN-LSTM-Attention model of the present invention;

[0112] Figure 4 is a schematic diagram of the comparison between the voltage simulation value and the real value of the present invention;

[0113] Figure 5 Schematic diagram for predicting future voltage data of the present invention;

[0114] Figure 6 Schematic diagram for comparing predicted data and real data of the present invention;

[0115] Figure 7 (a) Schematic diagram for comparing predicted value and real value of CNN of the present invention;

[0116] Figure 7 (b) Schematic diagram for comparing predicted value and real value of LSTM of the present invention;

[0117] Figure 7 (c) Schematic diagram for comparing predicted value and real value of CNN-LSTM of the present invention;

[0118] Figure 7 (d) Schematic diagram for comparing predicted value and real value of CNN-LSTM-Attention of the present invention;

[0119] Figure 8 Schematic diagram for insulation failure diagnosis process of the present invention;

[0120] Figure 9 Schematic diagram for voltage change of insulation resistance decrease of the present invention;

[0121] Figure 10(a) Root mean square error R MSE of the present invention; Schematic diagram of change;

[0122] Figure 10(b) Mean absolute error M AE of the present invention; Schematic diagram of change;

[0123] Figure 10(c) Coefficient of determination R 2 of the present invention; Schematic diagram of change;

[0124] Figure 11 (a) Schematic diagram for comparing real value and future value of voltage waveform after insulation failure of the present invention;

[0125] Figure 11 (b) Schematic diagram of real value of sampled voltage at 0.5 s after insulation failure of the present invention;

[0126] Figure 12 (a) Schematic diagram of change of identification parameter a1 of the present invention;

[0127] Figure 12 (b) Schematic diagram of algorithm fitting result of the present invention;

[0128] Figure 12 (c) Root mean square error R MSE of the present invention; Schematic diagram of change;

[0129] Figure 12 (d) Schematic diagram of the change of the forgetting factor λ of the present invention;

[0130] Figure 13 Schematic diagram of the calculation results of the present invention under different failure resistance values;

[0131] Figure 14 Schematic diagram of the insulation fault level classification standard of the present invention;

[0132] Figure 15 Schematic diagram of the judgment logic for insulation failure early warning of the present invention. Detailed implementation manners

[0133] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0134] Embodiment 1

[0135] This embodiment provides a method for insulation failure early warning of an energy storage battery system based on deep learning time series prediction, including the following steps:

[0136] S1: Construct an insulation detection circuit based on the unbalanced bridge method. The circuit includes adjustable bridge arm resistors R1-R4 and a bidirectional switch matrix, and the bridge arm resistors are dynamically switched through the bidirectional switch matrix;

[0137] As Figure 1 shown, the schematic diagram of the insulation detection circuit built using the Matlab / Simulink simulation platform, R p is the equivalent insulation resistance value between the positive pole of the high-voltage system and the ground, R n is the equivalent insulation resistance value between the negative pole of the high-voltage system and the ground, C p and C n are the equivalent Y capacitors incorporated into the battery, and their values are set to 10 μF. Battery is a high-voltage battery of 1500 V. The bridge arm resistors of this unbalanced bridge circuit are R1 = 30 kΩ, R2 = 30 kΩ, R3 = 30 MΩ, and R4 = 3 MΩ, that is, the fixed resistors in the insulation detection circuit. S1, S2, and S3 are the on-off switches in the circuit. S1 is turned on during insulation detection. The switching periods of S2 and S3 are 6 s. The sampling voltage U A1 is the voltage across resistor R2, and U A2 is the total voltage of the battery.

[0138] Calculate R p and R n using the insulation detection method. The length of the detection time mainly depends on the waiting time for the charging and discharging of the Y capacitor. As Figure 2As shown, the sampled voltage U A1 is obtained. From the waveform diagram, it can be seen that due to the influence of the Y capacitor, the sampled voltage U A1 shows a situation of ramping up and slow decline. After waiting for the times T1 and T2, the voltage only stabilizes, and then the resistance values of R p and R n can be calculated by substituting into the formula.

[0139] S2: Periodically switch the resistance configuration of the bridge arm resistance module at a preset frequency, and collect the positive electrode voltage U + (t) and the negative electrode voltage U - (t) with respect to the ground in the first mode and the second mode respectively, forming a voltage time series data set;

[0140] In this embodiment, the first mode means that the S2 switch is disconnected and the S3 switch is closed, and the second mode is that the S2 switch is closed and the S3 switch is disconnected;

[0141] S3: Input the voltage time series data set into the pre-trained CNN-LSTM-Attention deep learning model. Input the voltage data corresponding to 600 sampling points in units of 0.1 s. Extract spatial features through the convolutional neural network, capture temporal dependence relationships through the long short-term memory network, and dynamically weight key features through the attention mechanism to output the voltage prediction values for the next 6 seconds

[0142] As Figure 3 shown, in the CNN-LSTM-Attention deep learning model, the pooling layer of the convolutional neural network (CNN) uses a filter of size 3×1 and performs feature sampling with a sliding window of stride 1 to reduce the spatial size of the feature data and reduce network parameters. The feature data after being processed by the pooling layer will be passed through the fully connected layer to the long short-term memory network (LSTM) layer for deeper temporal information processing. The output of the convolutional layer of the convolutional neural network (CNN) is expressed as:

[0143]

[0144] Among them, f cov (·) represents the activation function, k represents the size of the sliding window, w n,m represents the weight of the convolutional kernel with n rows and m columns, X i+n,j+m represents the value of the nth row and the mth column of the input voltage data feature matrix, and b n,m represents the convolutional kernel bias;

[0145] Then, through the calculation of the hidden layer in the LSTM, the intermediate state of the voltage data is obtained. The attention mechanism passes through the function score([h t,i ,h t) to calculate the intermediate state feature vector h t,i with the hidden vector h t similarity, and its expression is as shown in the formula:

[0146]

[0147] where, W s represents the weight matrix of the fully connected layer, and b s represents the bias vector;

[0148] Use a function to calculate the attention weight α of the voltage data hidden layer vector i , and perform weighted summation with h t,i to obtain the output of the attention layer α i , The expressions are respectively:

[0149]

[0150] where, τ represents the output node of the fully connected layer, and input to the fully connected layer to obtain the predicted value y t ' of the sampled voltage;

[0151] Based on the CNN-LSTM-Attention combined training model structure, recursively predict the voltage data of the next 60 sampling points, and then through the root mean square error R MSE (threshold 2.5V), mean absolute error M AE (threshold 2V), coefficient of determination R 2 (threshold 97%) three-level indicators to determine insulation abnormality in real time;

[0152] S4: Calculate the predicted voltage value and the root mean square error R MSE , mean absolute error M AE and coefficient of determination R 2 with the measured value. When R MSE > 2.5V, M AE > 2V and R 2 < 0.97, trigger an insulation failure warning;

[0153] Specifically, the root mean square error R MSE , mean absolute error M AE and coefficient of determination R 2 are used as evaluation indicators of the CNN-LSTM-Attention deep learning model, and are specifically expressed as:

[0154]

[0155] where, yt and y t ′ are the true value and predicted value of the sampled voltage at time t, represents the average value of the true value of the sampled voltage, and N is the number of test samples;

[0156] As Figure 4 shown, it shows the simulation results of the algorithm based on the CNN-LSTM-Attention combined model structure. It can be seen that the voltage simulation value fits well with the true value, and its evaluation indexes are: R MSE = 1.3966V, M AE = 1.0025V and R 2 = 99.9%, indicating that the model has learned the data characteristics and distribution laws of the sampled voltage.

[0157] As Figure 5 shown in Figure 6 it shows the structure of the CNN-LSTM-Attention combined training model, and the future 60 sampled voltage data predicted recursively. Similarly, the predicted value is compared with the true value, and the obtained evaluation indexes are: R MSE = 1.5855V, M AE = 1.2583V and R 2 = 99.9%, indicating that the model can achieve good prediction of short-term future values.

[0158] To further highlight the advantages of the algorithm based on the CNN-LSTM-Attention combined training model structure, four prediction algorithms: CNN, LSTM, CNN-LSTM combination, and CNN-LSTM-Attention combination are compared. By making four consecutive predictions for the next 6s based on historical data, the fitting situation between the predicted value and the true value is observed respectively to judge the advantages and disadvantages of the algorithms. The fitting situations of each algorithm are as Figure 7 (a)- Figure 7 (d) shown.

[0159] As Figure 7 (a), Figure 7 (b) shown, the future data predicted by the CNN and LSTM algorithms has a large difference from the true value, and there are obvious fluctuations, which will have a greater impact on the diagnosis of insulation failure. As Figure 7 (c) shown, although the CNN-LSTM combination algorithm can better fit the smooth trend of the waveform change, there is still a certain deviation between its predicted data and the true value. As Figure 7(d), the fitting effect between the multiple predicted values and the true values by the CNN-LSTM-Attention combined model algorithm is good, indicating the superiority of this algorithm in predicting future data based on historical time-series data, greatly improving the diagnostic accuracy of insulation failure and reducing the probability of false alarms.

[0160] As Figure 8 shown, it is the insulation failure diagnosis process, and its strategy is as follows: ① Turn on the insulation detection system, close the circuit switch S1, and S2 and S3 are periodically opened and closed to process the voltage data of the sampling points: collect a voltage data every 0.1 s, and a total of 60 s need to be collected, that is, 600 sampling point voltages, which form the training data set required by the algorithm; ② Use the algorithm to predict the training data set and recursively predict the voltage data of the next 60 sampling points, that is, the voltage values in the next 6 s; ③ During the next 6 s, compare the actual value of the sampling point voltage with the algorithm-predicted future value in real time, and based on the calculated comparison with the true value, each evaluation index R MSE , M AE and R 2 , and considering the fault tolerance of the system, set their thresholds as: R MSE ≤2.5 V, M AE ≤2 V, R 2 ≤97%, in the actual application process, its threshold can be adjusted accordingly in combination with the data of multiple failure scenarios. If each evaluation index is within the threshold, repeat steps ① and ② to predict the future value of the voltage in the next 6 s; ④ If any one of the evaluation indexes appears abnormal during the next 6 s, immediately start the early warning.

[0161] As Figure 9 shown, it is the simulation of the sampling point voltage change curve when the total negative pole of the battery side is grounded and the resistance R n has an insulation failure. By setting its resistance to instantaneously drop from 10 MΩ to 100 kΩ, 200 kΩ, 750 kΩ, and 1500 kΩ at the 73 s moment respectively. It can be seen that when an insulation failure occurs, there will be a voltage offset area between the future value predicted based on the historical sampling voltage data in the previous period and the true voltage value. The size of this voltage offset area increases with the decrease of the insulation failure resistance, indicating that it will cause corresponding changes in each evaluation index R MSe , M Ae and R 2 . Its each evaluation index changes between 72 - 75 s. As Figure 10(a)-Figure 10(c) shown, each evaluation index R MSE , M AE and R 3 changes significantly with the decrease of the insulation failure resistance. The evaluation index R MSEThe times to reach the set threshold of 2.5 V at 100 kΩ, 200 kΩ, 750 kΩ, and 1500 kΩ are 73.05 s, 73.30 s, 73.48 s, and 74.31 s respectively, and the evaluation index M AE The times to reach the set threshold of 2.0 V at 100 kΩ, 200 kΩ, 750 kΩ, and 1500 kΩ are 73.08 s, 73.35 s, 73.47 s, and 74.33 s respectively, and the evaluation index R 2 The times to reach 97% of the set threshold at 100 kΩ, 200 kΩ, and 750 kΩ are 73.23 s, 73.62 s, and 74.50 s respectively. The times for each evaluation index to reach the set threshold all indicate that as the insulation resistance value decreases, the system diagnosis time becomes faster, meeting the requirement of timely identification of more severe insulation failures.

[0162] S5: When the CNN-LSTM-Attention combined model algorithm detects an insulation failure in the energy storage battery system, continue to collect abnormal voltage data of the sampling points within 0.5 s after exceeding the evaluation index threshold. Based on the function model of the voltage of the capacitor during charge and discharge at time t:

[0163]

[0164] where V ∞ is the sampling point voltage when the capacitor is stable, V0 is the initial voltage, RC is the time constant, and t is the time;

[0165] Based on the above function model formula, as long as the stable value of the sampling voltage is obtained and it is determined whether the voltage is in the rising or falling period, the corresponding resistance value of the insulation failure can be calculated. Next, the fast calculation of the stable value of the sampling voltage will be carried out based on the adaptive forgetting factor recursive least squares method, and the above function model formula will be transformed into the discrete-time function of U f The expression is:

[0166]

[0167] where Δt is the sampling time interval, and U f (k) represents the reflected wave voltage value at time k, a1 corresponds to V in the function model formula ∞ , a2 corresponds to (V0 - V ∞ ), and a3 corresponds to RC. These three are all parameters to be identified by the forgetting factor, mainly the identification of parameter a1;

[0168] Let U f (k) = F(A, k), where A represents the model parameters and A = [a1 a2 a3];

[0169] Set the initial values of the parameters Initial value of error covariance Forgetting factor

[0170] Construct the observation matrix H(k):

[0171]

[0172] Calculate the prediction error e(k):

[0173]

[0174] Calculate the gain matrix K(k):

[0175] K(k) = P(k - 1)·H T (k)·[λ(k - 1)+H(k)·P(k - 1)·H T (k)] -1

[0176] Update the parameter estimate

[0177]

[0178] Update the covariance matrix P(k):

[0179]

[0180] Update the forgetting factor λ(k)

[0181] When , λ(k) = max[0.9, λ(k - 1)-0.01];

[0182] When , λ(k) = min[0.99, λ(k - 1)+0.01];

[0183] As Figure 11 (a)- Figure 11 (b) shows, in order to set the resistance value of R n to instantaneously drop from 10 MΩ to 200 KΩ at 73 s, the voltage of the sampling point is sampled and recorded. The sampling frequency is 100 Hz, the sampling time is 0.5 s, and the sampling voltage waveform diagram.

[0184] As Figure 12 (a) shows, for Figure 11 (b), the parameter a1 of the sampling voltage is identified. The identified parameter a1 converges, and the convergence value is 118.24 V, indicating that when the capacitor in the circuit is stable during the rising period of the sampling voltage, the voltage value U of the sampling point A1 is 118.24 V. Substitute the identified values of the remaining parameters into the formula. AsFigure 12 As shown in (b), it can be seen that the fitting values of the algorithm and the true values have a good fitting effect within 0.5 s, such as Figure 12 (c) shows that R MSE also finally tends to 0. Similarly, sample and record the voltage 0.5 s before the voltage drop period of the sampling point. As Figure 12 (d) shows, identify the parameter a1 through the recursive least squares algorithm with an adaptive forgetting factor, and obtain U A ′1 as 0.36 V. Substitute U A1 , U A ′1 into the formula respectively to calculate the negative pole-to-ground resistance R n = 193.96 KΩ, and the relative error is 3.02%. The positive pole-to-ground resistance R p = 9.56 MΩ, and its relative error is 4.40%, which reflects the accuracy of the algorithm.

[0185] As Figure 13 shown, it shows the calculation results of R n , R p based on the recursive least squares algorithm with an adaptive forgetting factor at different failure resistance values of the positive and negative pole-to-ground resistances in the insulation detection circuit. It can be seen that the smaller the failure resistance value, the higher the accuracy of the insulation resistance value calculated by the algorithm. The measurement of the failure resistance value when the energy storage battery system has an insulation failure is realized, and the accuracy requirement is met at the same time.

[0186] S6: Invert and calculate the positive and negative pole-to-ground insulation resistances R p and R n according to the capacitance charge and discharge characteristic equation. The specific insulation detection process is as follows:

[0187] (1) The BMS automatically activates the detection module;

[0188] (2) Close the main control switch S1 to connect the detection circuit;

[0189] (3) Disconnect the S2 switch and close the S3 switch. At this time, the Y capacitor C p discharges, and C n charges. Considering the charge and discharge time of the Y capacitor, it is necessary to wait for a certain time T1;

[0190] (4) Given that the voltage at the sampling point A1 is U A1 , calculate the voltages U 12 across R1 and R2. Given the total battery voltage U A2 , the following formula can be obtained:

[0191]

[0192] (5) Close the S2 switch and disconnect the S3 switch. At this time, the Y capacitor C p charges, and Cn Discharge, considering the charge and discharge time of capacitor Y, wait for a certain time T2;

[0193] (6) It is known that the second voltage at the voltage sampling point A1 is U A ′1, calculate the voltage U1′2 between R1 and R2, knowing that the total voltage U A ′2, the following formula is obtained:

[0194]

[0195] (7) Combining the above formulas, we can obtain the insulation resistance R of the high voltage to ground: p , high voltage negative insulation resistance to ground R n , the result is as follows:

[0196]

[0197] (8) Cycle insulation test and repeat steps 3 to 7.

[0198] By collecting time series data within 0.5 seconds, the capacitor charging and discharging parameters are identified and the insulation resistance R is inverted. p and R n And control the error <5%;

[0199] S7: Establish a graded warning mechanism, set the insulation resistance grade threshold (>10MΩ normal, 750kΩ-10MΩ warning, <750kΩ protection), and cut off the main circuit after triggering.

[0200] like Figure 14 As shown in the figure, the insulation fault level classification standard is set. The CNN-LSTM-Attention combined model algorithm can be used to quickly detect whether insulation failure has occurred, and the adaptive forgetting factor recursive least squares algorithm can be used to calculate the insulation failure resistance value. Combining the two constitutes the energy storage battery insulation failure warning implementation process. Its warning logic is as follows: Figure 15 As shown in the figure, when the CNN-LSTM-Attention combined model algorithm is evaluated by the real-time evaluation index R MSE 、M AE and R 2 When insulation failure of the energy storage battery is detected, the insulation failure warning is ready to start. Collect 0.5s data of the voltage corresponding to the sampling point, perform parameter identification based on the AFFRLS algorithm, and obtain the voltage stability value U A1 and U A ′1, use the formula to calculate the corresponding ground resistance R n , R p, so as to determine the warning level. If any of its resistance values is less than 750 KΩ, a red warning is initiated and the main circuit is disconnected. If any of its resistance values is greater than 750 KΩ and less than 1500 KΩ, an orange warning is initiated and the main circuit is disconnected. If any of its resistance values is greater than 1500 KΩ and less than 10 MΩ, a yellow warning is initiated, and the change of its resistance value is continuously monitored through the AFFRLS algorithm until the two resistance values return to the normal value of 10 MΩ.

[0201] Embodiment 2

[0202] The present invention provides an insulation failure warning system for an energy storage battery system based on deep learning time series prediction, which is used to implement the insulation failure warning method for the energy storage battery system based on deep learning time series prediction in the above Embodiment 1. The system includes: an insulation detection circuit construction module, a voltage time series data set construction module, a voltage predicted value output module, an evaluation index construction module, a parameter identification module, an insulation resistance calculation module, and an insulation failure warning module;

[0203] In this embodiment, the insulation detection circuit construction module is used to construct an insulation detection circuit, including adjustable bridge arm resistors and a bidirectional switch matrix, and the bridge arm resistors are dynamically switched through the bidirectional switch matrix; the voltage time series data set construction module is used to periodically switch the resistance value configuration of the bridge arm resistors at a preset frequency, and collect the positive pole-to-ground voltage and the negative pole-to-ground voltage in a dual-mode to form a voltage time series data set; the voltage predicted value output module is used to input the voltage time series data set into a pre-trained CNN-LSTM-Attention deep learning model, extract spatial features through a convolutional neural network, capture time series dependencies through a long short-term memory network, and dynamically weight key features through an attention mechanism to output a voltage predicted value; the evaluation index construction module is used to construct an evaluation index and its threshold of the voltage predicted value to detect whether an insulation failure occurs in the energy storage battery system; based on the function model of the voltage of the capacitor during charging and discharging at time t, the adaptive forgetting factor recursive least squares method is used to identify the parameters of abnormal voltage data to obtain the voltage stable value; the insulation resistance calculation module is used to inversely calculate the insulation resistance of the positive and negative poles to the ground according to the capacitor charging and discharging characteristic equation; the insulation failure warning module is used to establish a hierarchical warning mechanism, set the insulation resistance hierarchical threshold, and cut off the main circuit and initiate an insulation failure warning after being triggered.

[0204] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.

Claims

1. A method for warning of insulation failure of an energy storage battery system based on deep learning time series prediction, characterized in that, It includes the following steps: Construct an insulation detection circuit, including adjustable arm resistors and a bidirectional switch matrix, and dynamically switch the arm resistors through the bidirectional switch matrix; Periodically switch the resistance configuration of the arm resistors at a preset frequency, collect the positive-pole-to-ground voltage and the negative-pole-to-ground voltage in a dual mode, and form a voltage time-series data set; Input the voltage time-series data set into a pre-trained CNN-LSTM-Attention deep learning model. Extract spatial features through a convolutional neural network, capture temporal dependence relationships through a long short-term memory network, and dynamically weight key features through an attention mechanism to output voltage prediction values; Construct an evaluation index and its threshold for the voltage prediction value to detect whether an insulation failure occurs in the energy storage battery system; Collect abnormal voltage data of sampling points within a set time after exceeding the evaluation index threshold. Based on the function model of the voltage of the capacitor during charging and discharging at time t, use the adaptive forgetting factor recursive least squares method to identify parameters for the abnormal voltage data to obtain the voltage stability value; Invert and calculate the insulation resistances of the positive and negative poles to the ground according to the capacitor charging and discharging characteristic equation; Establish a hierarchical early warning mechanism, set hierarchical thresholds for insulation resistances, and cut off the main circuit and start an insulation failure early warning after being triggered; 2. The insulation failure warning method for an energy storage battery system based on deep learning time series prediction according to claim 1, wherein In the CNN-LSTM-Attention deep learning model, the output of the convolutional neural network is expressed as: Among them, represents the output of the convolutional neural network, and f cov (·) represents the activation function, k represents the sliding window size, and w n,m represents the weight of the convolutional kernel with n rows and m columns, and X i+n,j+m represents the value of the input voltage data feature matrix at the nth row and the mth column, and b n,m represents the convolutional kernel bias; Input the output of the convolutional neural network into the long short-term memory network to obtain an intermediate state vector of the voltage data; The attention mechanism calculates the similarity between the intermediate state feature vector and the hidden vector of the voltage data, calculates the attention weights of the hidden layer vector of the voltage data, weights them with the intermediate state feature vector to obtain the output of the attention layer, and outputs the prediction value of the sampling voltage through a fully connected layer; 3. The insulation failure warning method for an energy storage battery system based on deep learning time series prediction according to claim 2, wherein The attention mechanism calculates the similarity between the intermediate state feature vector of the voltage data and the hidden vector, specifically through the function score([h t,i ,h t ) to calculate the similarity between the intermediate state feature vector of the voltage data and the hidden vector, expressed as: Among them, h t,i represents the intermediate state feature vector, h t represents the hidden vector, W s represents the weight matrix of the fully connected layer, b s represents the bias vector; Calculate the attention weights of the hidden layer vector of the voltage data, and weight them with the intermediate state feature vector to obtain the output of the attention layer, which is expressed as: Among them, represents the output of the attention layer, and α i represents the attention weight, and τ represents the output node of the fully connected layer.

4. The insulation failure warning method for the energy storage battery system based on deep learning time series prediction according to claim 1, characterized in that Based on the function model of the voltage of the capacitor during charging and discharging at time t, use the adaptive forgetting factor recursive least squares method to identify parameters for the abnormal voltage data, specifically including: The function model of the voltage of the capacitor during charging and discharging at time t is expressed as: Among them, V ∞ is the voltage at the sampling point when the capacitance is stable, V0 is the initial voltage, RC is the time constant, and t is the time; Convert the function model to U f The discrete-time function of, expressed as: where Δt is the sampling time interval, and U f (k) represents the reflected wave voltage value at time k; Let U f (k) = F(A, k), where A represents model parameters, and A = [a1 a2 a3]; Set the initial value of the parameter Initial value of error covariance Forgetting factor Construct an observation matrix H(k): Calculate the prediction error e(k): Calculate the gain matrix K(k): K(k) = P(k - 1)·H T (k)·[λ(k - 1) + H(k)·P(k - 1)·H T (k)] -1 Update parameter estimates Update the covariance matrix P(k): Update the forgetting factor λ(k): When , λ(k) = max[0.9, λ(k - 1) - 0.01]; When is true, λ(k) = min[0.99, λ(k - 1)+0.01].

5. The insulation failure warning method for an energy storage battery system based on deep learning time series prediction according to claim 1, characterized in that Invert and calculate the insulation resistances of the positive and negative poles to the ground according to the capacitor charging and discharging characteristic equation, specifically including: Close the main control switch S1 to connect the insulation detection circuit; Disconnect the on-off switch S2 and close the on-off switch S3. At this time, the Y capacitor C p discharges, and the Y capacitor C n charges and waits for the time T1; The voltage at the known sampling point A1 is U A1 , calculate the voltages U across resistors R1 and R2 12 , given the total battery voltage U A2 , the following formula is obtained: Close the on-off switch S2 and open the on-off switch S3. At this time, the Y-capacitor C p is charged, and the Y-capacitor C n is discharged, and wait for the time T2; The second voltage at the known voltage sampling point A1 is U A ′ 1. Calculate the voltages U1 at the resistors R1 and R2 ′ 2. The total battery voltage U is known from the second acquisition A ′ 2. The following formula can be obtained: The high-voltage positive-to-ground insulation resistance value R is obtained by solving the simultaneous equations p and the high-voltage negative-to-ground insulation resistance value R n , and the results are as follows: Among them, the resistors R1, R2, R3, and R4 are adjustable arm resistors and are configured asymmetrically.

6. An insulation failure warning system for an energy storage battery system based on deep learning time series prediction, characterized in that, It includes: An insulation detection circuit construction module, a voltage time-series data set construction module, a voltage prediction value output module, an evaluation index construction module, a parameter identification module, an insulation resistance calculation module, and an insulation failure early warning module; The insulation detection circuit construction module is used to construct an insulation detection circuit, including adjustable arm resistors and a bidirectional switch matrix, and dynamically switch the arm resistors through the bidirectional switch matrix; The voltage time-series data set construction module is used to periodically switch the resistance configuration of the arm resistors at a preset frequency, collect the positive-pole-to-ground voltage and the negative-pole-to-ground voltage in a dual mode, and form a voltage time-series data set; The voltage prediction value output module is used to input the voltage time series data set into a pre-trained CNN-LSTM-Attention deep learning model, extract spatial features through a convolutional neural network, capture temporal dependence relationships through a long short-term memory network, and dynamically weight key features through an attention mechanism, and output the voltage prediction value; The evaluation index construction module is used to construct the evaluation index and its threshold of the voltage prediction value to detect whether the energy storage battery system has insulation failure; The parameter identification module is used to collect the abnormal voltage data of the sampling points within a set time after the evaluation index threshold is exceeded, and based on the function model of the voltage of the capacitor during charge and discharge at time t, use the adaptive forgetting factor recursive least squares method to identify the parameters of the abnormal voltage data to obtain the voltage stable value; The insulation resistance calculation module is used to inversely calculate the insulation resistance of the positive and negative electrodes to the ground according to the capacitor charge and discharge characteristic equation; The insulation failure warning module is used to establish a hierarchical warning mechanism, set the hierarchical threshold of the insulation resistance, cut off the main circuit after being triggered, and start the insulation failure warning.

7. The insulation failure warning system for an energy storage battery system based on deep learning time series prediction according to claim 6, wherein In the CNN-LSTM-Attention deep learning model, the output of the convolutional neural network is expressed as: Among them, represents the output of the convolutional neural network, and f cov (·) represents the activation function, k represents the sliding window size, and w n,m represents the weight of the convolutional kernel with n rows and m columns, and X i+n,j+n represents the value of the n-th row and m-th column of the input voltage data feature matrix, and b n,m represents the convolutional kernel bias; Input the output of the convolutional neural network into the long short-term memory network to obtain the intermediate state vector of the voltage data; The attention mechanism calculates the similarity between the intermediate state feature vector and the hidden vector of the voltage data, calculates the attention weight of the hidden layer vector of the voltage data, and weights it with the intermediate state feature vector to obtain the output of the attention layer, and outputs the prediction value of the sampling voltage through the fully connected layer.

8. The insulation failure warning system for an energy storage battery system based on deep learning time series prediction according to claim 7, wherein, The attention mechanism calculates the similarity between the intermediate state feature vector of the voltage data and the hidden vector, specifically through the function score([h t,i ,h t ) to calculate the similarity between the intermediate state feature vector of the voltage data and the hidden vector, expressed as: Among them, h t,i represents the intermediate state feature vector, h t represents the hidden vector, W s represents the weight matrix of the fully connected layer, b s represents the bias vector; The attention weight of the hidden layer vector of the voltage data is calculated and weighted with the intermediate state feature vector to obtain the output of the attention layer, which is expressed as: Among them, represents the output of the attention layer, and α i represents the attention weight, and τ represents the output node of the fully connected layer.

9. The insulation failure warning system for an energy storage battery system based on deep learning time series prediction according to claim 6, characterized in that Based on the function model of the voltage of the capacitor during charge and discharge at time t, the adaptive forgetting factor recursive least squares method is used to identify the parameters of the abnormal voltage data, specifically including: The function model of the voltage of the capacitor during charge and discharge at time t is expressed as: Among them, V ∞ is the voltage at the sampling point when the capacitor is stable, V0 is the initial voltage, RC is the time constant, and t is the time; Convert the function model into a discrete-time function of U f which is expressed as: where Δt is the sampling time interval, and U f (k) represents the reflected wave voltage value at time k; Let U f (k) = F(A, k), where A represents the model parameters, and A = [a1 a2 a3]; Set the initial value of the parameter Initial value of the error covariance Forgetting factor Construct the observation matrix H(k): Calculate the prediction error e(k): Calculate the gain matrix K(k): K(k) = P(k - 1)·H T (k)·[λ(k - 1)+H(k)·P(k - 1)·H T (k)] -1 Update parameter estimates Update the covariance matrix P(k): Update the forgetting factor λ(k): When is true, λ(k) = max[0.9, λ(k - 1) - 0.01]; When is true, λ(k) = min[0.99, λ(k - 1)+0.01].

10. The insulation failure warning system for an energy storage battery system based on deep learning time series prediction according to claim 6, wherein Inversely calculate the insulation resistance of the positive and negative electrodes to the ground according to the capacitor charge and discharge characteristic equation, specifically including: Close the main control switch S1 to connect the insulation detection circuit; Disconnect the on-off switch S2 and close the on-off switch S3. At this time, the Y-capacitor C p discharges, and the Y-capacitor C n charges and waits for the time T1; The voltage at the known sampling point A1 is U A1 , calculate the voltages U across the resistors R1 and R2 12 , given the total battery voltage U A2 , the following formula can be obtained: Close the on-off switch S2 and open the on-off switch S3. At this time, the Y capacitor C p is charged, and the Y capacitor C n is discharged, and wait for the time T2; The second voltage at the known voltage sampling point A1 is U A ′ 1. Calculate the voltages U1 at the resistors R1 and R2 ′ 2. The total battery voltage U is known from the second acquisition A ′ 2. The following formula is obtained: The high-voltage positive-to-ground insulation resistance value R is obtained by solving the simultaneous equations p and the high-voltage negative-to-ground insulation resistance value R n , and the results are as follows: Among them, the resistors R1, R2, R3, and R4 are adjustable bridge arm resistors and are configured asymmetrically.

Images