CN111736084A - Health state prediction method of valve-regulated lead-acid battery based on improved LSTM neural network - Google Patents

Abstract

Based on the improved LSTM neural network for valve-regulated lead-acid battery health state prediction method, the floating charge voltage, equalization current, equalization duration, discharge cut-off voltage, and discharge duration of the battery are measured daily by the online monitoring device. Measured every two months at check balance charge. Taking n days as the time span, build an n-dimensional sample input x(t _i ) . Taking the battery capacity data sequence h(t _i ) as the output and x(t _i ) as the input, a neural network model including multiple LSTM neural network units is established. In the initial state, the weight matrix W and the bias matrix b in the network are assigned values by randomly generating decimals between 0 and 1. The Dropout algorithm is introduced to improve the LSTM neural network model and its training process is improved. The invention can reduce the problems of low prediction accuracy and underfitting caused by insufficient data samples, accurately predict the health state of the battery in the substation, and improve the utilization rate of the battery.

Description

Health state prediction method of valve-regulated lead-acid battery based on improved LSTM neural network

technical field

The invention belongs to the technical field of artificial intelligence control of valve-regulated lead-acid batteries in substations, and in particular relates to a method for predicting the health state of valve-regulated lead-acid batteries based on an improved LSTM neural network.

Background technique

The VRLA battery is the core of the DC power system, and its performance quality is related to the safe and stable operation of the entire substation. However, it is difficult to estimate the state of health of the battery in the substation in actual operation. In order to improve the reliability of the power supply of the battery in the event of an accident. The sealed VRLA battery has many advantages such as superior performance, simple maintenance, convenient installation, high reliability, and no environmental pollution, so it has many applications in the DC system of substations. As a backup power source, the VRLA battery pack is affected by the operation mode of the substation and has unique operating characteristics: (1) During the normal operation of the substation, the VRLA battery pack is in a floating state, and it is not actually (2) When an accident in the power grid causes the AC system of the substation to lose power, the valve-regulated lead-acid battery bank serves as the emergency power supply of the substation to provide DC power for the equipment. Therefore, the valve-regulated lead-acid battery pack in the DC system of the substation has been in the floating charging state for a long time, and the measured value of its maximum energy storage capacity can only be measured by checking the balance charging once every two months. It can be seen that the battery in the substation scenario is in a floating state for a long time, and it is difficult to collect data such as the actual capacity of the battery, and there are problems such as insufficient data samples and low accuracy of prediction results.

The current approach is to replace VRLA batteries in substations at regular intervals of two years. However, this model has shortcomings in terms of excessive consumption and environmental pollution. Therefore, finding a practical method to estimate the health status of VRLA batteries in substations, effectively improve the utilization efficiency of VRLA batteries, and reduce the occurrence of power grid accidents caused by battery failure is the current need to solve technical issues.

SUMMARY OF THE INVENTION

The invention proposes a method for predicting the state of health of a valve-regulated lead-acid battery based on an improved LSTM neural network, which can more accurately and rapidly predict the state of health of the battery.

The technical scheme adopted in the present invention is:

The health state prediction method of VRLA battery based on improved LSTM neural network includes the following steps:

Step 1. Collection of sample data:

The floating charge voltage, equalization current, equalization duration, discharge cut-off voltage, and discharge duration input data of the battery are measured daily by the online monitoring device, and the battery capacity is measured by checking the equalization charge every two months;

Step 2. Preprocessing of sample data:

其中，

分别表示n天内的蓄电池的浮充电压、均充电流、均充时长、放电截止电压、放电时长的向量，蓄电池容量数据序列h(t_i)，即多次容量实测结果；With n days as the time span, establish an n-dimensional sample input

in,

respectively represent the float voltage, equalization current, equalization duration, discharge cut-off voltage, and discharge duration of the battery within n days, and the battery capacity data sequence h(t _i ) is the result of multiple capacity measurements;

Step 3. Build the LSTM neural network model:

Taking the battery capacity data sequence h(t _i ) as the output and x(t _i ) as the input, a neural network model containing multiple LSTM neural network units is established. Each LSTM neural network unit can be regarded as an LSTM neural network at different times. The state on the span, in the initial state, by randomly generating decimals between 0 and 1, assign values to the weight matrix W and the bias matrix b in the network;

Step 4. Introduce the Dropout algorithm to improve the LSTM neural network model and improve its training process.

Step 5. Substitute the input samples in the test set into the trained model, and then 12 capacity prediction values of the battery can be obtained, and the interval between each value is 2 months.

The present invention is a method for predicting the state of health of a valve-regulated lead-acid battery based on an improved LSTM neural network, and the technical effects are as follows:

1) The present invention proposes a method for predicting the state of health of a valve-regulated lead-acid battery based on an improved LSTM neural network. This method introduces artificial intelligence technology into the health state prediction of VRLA batteries in substations. In the case that the actual capacity and other data are difficult to collect, and the accuracy of the prediction results of the general artificial intelligence method is low, the time span of each group of input data samples in the present invention is two months, and the input samples are float voltage, equalizing current, The information of equalizing charge time, discharge cut-off voltage and discharge time are all 60-dimensional vectors. And since the battery operation time of the substation is 2 years, the number of time series samples collected by a single battery is only 12 groups. That is, the actual capacity measurement results of the battery for 12 times during the two-year period are used as the sample output data. Establish a multi-level LSTM prediction model, relying on the long and short-term memory characteristics of LSTM and increasing the complexity of the network model to improve the accuracy of the prediction results.

2) In order to prevent the over-fitting problem caused by the increase of model complexity, Dropout optimization algorithm is introduced to improve its training process, and the activation state of neurons is determined according to the connection strength of each neuron, that is, the neuron with higher connection strength changes. The probability of being in the inactive state is greater. In this way, the dependence of the LSTM prediction model on some input features is reduced. Improve the generalization ability of the model, so that the model has a high accuracy and good adaptability.

3) The method proposed by the present invention can reduce the problem of low prediction accuracy and underfitting caused by insufficient data samples, and at the same time avoid the problem of overfitting caused by increasing the complexity of the neural network model, and improve the generalization of the model. It can accurately predict the health status of the battery in the substation. It can provide a basis for timely maintenance or replacement of batteries by substation staff, thereby improving battery utilization while ensuring battery reliability, ensuring the reliable operation of substations and power grid security. When the invention is used in the direct current system of the substation, compared with the existing method, the health state of the storage battery can be predicted more accurately and quickly.

Description of drawings

Figure 1 is a flowchart of the improved Dropout optimization method.

Figure 2 is the network training flow chart of the improved LSTM.

Fig. 3(a) is the prediction result of battery health state of station A;

Fig. 3(b) is the prediction result of battery health state of station B;

Fig. 3(c) is the result of prediction result of battery health state of station C;

Fig. 3(d) is a graph showing the prediction result of the battery health state of station D.

Figure 4(a) is the result of the prediction result of the battery state of health of the E station;

Fig. 4(b) is the predicted result of the battery state of health of the F station;

Figure 4(c) is a graph showing the prediction result of the battery health state of station G.

Detailed ways

LSTM neural network is a deep neural network with long and short-term memory. It is mainly composed of forgetting gate, input gate and output gate, which determines the degree to which new input or historical information is forgotten or retained.

The health state prediction method of VRLA battery based on improved LSTM neural network includes the following steps:

Step 1. Collection of sample data:

The floating charge voltage, equalization current, equalization duration, discharge cut-off voltage, and discharge duration input data of the battery are measured daily by the online monitoring device. The battery capacity is measured by checking the equalization charge every two months.

Online monitoring device configuration list: The configuration is based on 2V300AH, 104 pieces, and 2 sets of battery packs.

Step 2. Preprocessing of sample data:

其中，

分别表示60天内的蓄电池的浮充电压、均充电流、均充时长、放电截止电压、放电时长的向量，蓄电池容量数据序列h(t_i)，即两年期间每间隔2个月蓄电池所进行的共12次容量实测结果。Create a 60-dimensional sample input with a time span of 60 days

in,

The vectors representing the float voltage, equalization current, equalization duration, discharge cut-off voltage, and discharge duration of the battery within 60 days, respectively, and the battery capacity data sequence h(t _i ), that is, the battery is charged every 2 months during the two-year period. A total of 12 capacity measurement results.

Step 3. Build the LSTM neural network model:

Taking the battery capacity data sequence h(t _i ) as the output and x(t _i ) as the input, a neural network model containing 12 LSTM neural network units is established. Each LSTM neural network unit can be regarded as an LSTM neural network at different times. The state on the span, in the initial state, by randomly generating decimals between 0 and 1, assign values to the weight matrix W and the bias matrix b in the network;

Step 4. Introduce the Dropout algorithm to improve the LSTM neural network model and improve its training process.

The Dropout algorithm is a solution to prevent overfitting in neural network training, and is mainly suitable for situations with high neural network complexity and large network scale. Its core is to change the activation state of neurons in the process of network model training, thereby reducing the dependence of neural network prediction results on some local neurons, similar to the operation of avoiding the local optimal trap, preventing the overfitting of the model. , to improve the generalization ability of the model.

The principle of the Dropout algorithm is to randomly select neurons in the network to change their activation states during each iteration of training, and gradually complete the training of the network model. However, using the Dropout algorithm will increase the training time of the model by 2-3 times, and in the case of a complex network model, it will even exceed the number of iterations and cannot converge.

The Dropout optimization algorithm proposed by the present invention takes the connection strength of the neuron as the probability of changing its activation state, so as to improve the convergence speed of the training process. That is, neurons with higher connection strength have a higher probability of transitioning to an inactive state. In this way, the dependence of the LSTM prediction model on some input features is reduced.

Step 5. Substitute the input samples in the test set into the trained model, and then 12 capacity prediction values of the battery can be obtained, and the interval between each value is 2 months.

In the step 2, in the preprocessing of the sample data,

x(t _i ) is the network input of the LSTM neural network at time t _i , h(t _i ) is the network output at time t _i , and C(t _i ) is the unit state output of the network at time t _i ;

Among them, the network input includes the float voltage, equalization current, equalization duration, discharge cut-off voltage and discharge duration of the battery. The network output is the maximum energy storage capacity of the battery, namely:

为[t_i-60,t_i]期间的浮充电压，

分别为表征均衡充电电流大小和充电时长的向量，若第j天没有均衡充电，则对应向量中的元素取值为0，即

则为[t_i-60,t_i]期间蓄电池放电的记录，

为放电截止电压向量，

为放电时长的向量，若蓄电池第j天没有放电，则截止电压数值上与浮充电压相等，即

SOH(t_i)为蓄电池第t_i天测得的电池储能容量。Among them, each element in x(t _i ) is a vector with a dimension of 60, representing the charge and discharge information from day t _i and 60 days back, where

is the float voltage during [t _i -60,t _i ],

are the vectors representing the equalizing charging current and charging duration, respectively. If there is no equalizing charging on the jth day, the element in the corresponding vector takes the value 0, that is,

is the record of battery discharge during [t _i -60,t _i ],

is the discharge cut-off voltage vector,

is the vector of discharge time. If the battery is not discharged on the jth day, the cut-off voltage is numerically equal to the float voltage, that is

SOH(t _i ) is the battery energy storage capacity measured on day t _i of the battery.

The step 3 includes the following steps:

3.1. Initialization of network hyperparameters: The set hyperparameters include: the number of input nodes m, the number of hidden nodes k, the number of output nodes n, the learning rate y _ita , the error threshold σ, and the number of LSTM nuclei w.

3.2. Weight bias initialization: In the initial state, by randomly generating decimals between 0 and 1, assign values to the weight matrix W and bias matrix b in the network.

The step 4 includes the following steps:

Step 4.1, forward operation to predict battery capacity:

According to the initially set parameters, the parameters of each gate in the LSTM model are updated according to the following formula (1), and the output results of the network are further obtained according to the formula (8):

Among them, f(t _i ), i(t _i ), o(t _i ), C(t _i ) represent forget gate output, input gate output, output gate output and unit state respectively, h(t _i ) is t _i network output at time. Both σ and tanh are activation functions, where σ is the sigmoid function and tanh is the hyperbolic tangent function. The calculation formulas of the two are as follows:

Here e is a natural constant, and z is used as a variable to express the formulas of these two activation functions.

W _f , Wi , W _c , and W _o represent the weight matrix of the forget gate, input gate, current input unit state and output gate, respectively, and b _f , _bi , b _c , and _{bo represent} the forget gate, input gate, current Input unit state and bias matrix of output gate. These 8 parameter matrices are parameter matrices to be obtained, which are gradually optimized and updated during the training process of the model.

表示按元素乘，当

作用于两个向量时，运算如下：

means element-wise multiplication, when

When acting on two vectors, the operation is as follows:

作用于一个向量和一个矩阵时，运算如下：when

When acting on a vector and a matrix, the operation is as follows:

作用于两个矩阵时，两个矩阵对应位置的元素相乘即可。when

When acting on two matrices, the elements at the corresponding positions of the two matrices can be multiplied.

W _f , W _i , W _c , W _o , b _f , b _i , b _c , b _o , these 8 parameters are trained by the network, no need to manually set specific values, but need to manually specify the matrix dimension, and set by The computer generates a random number between 0 and 1 as its initial value.

Step 4.2. Correct the weight and bias parameters of the neural network according to the error of the prediction result:

After the output value of the network is calculated according to formula (2), the error C between the predicted value and the actual value is calculated according to formula (6). Update parameters and biases in the network.

C=|h'(t _i )-h(t _i )| (6)

In the above formula, h'(t _i ) represents the predicted capacity value output by the LSTM network, h(t _i ) represents the actual capacity value, α represents the learning rate, W=[W _f , Wi , W _c , W _{o ]} and b=[b _f , b _i , b _c , b _o ] represents the weight and bias before updating, and W′ and b′ represent the weight and bias after updating.

4.3. Neuron activation state update:

The connection strength of all neurons is calculated according to formula (8), and the activation state of neurons is updated according to the probability calculated according to formula (9).

A Dropout optimization algorithm is proposed, which uses the link strength of neurons as the probability of changing their activation state to improve the convergence speed of the training process. Let the state of the neuron be activated and inactive, and the value of S _i (t) is 1 and 0, respectively, indicating that the neuron i is in the active state and the inactive state in the t-th iteration. The calculation formula for defining the connection strength R _i (t) of neuron i is:

Among them, S _j (t) is the activation state of any neuron except i in the network, and w _ij (t)∈W is the weight between neurons i and j in the t-th iteration. The activation state of neurons in the iterative process is updated according to the following formula (9):

That is, neurons with higher connection strength have a higher probability of transitioning to an inactive state. In this way, the dependence of the LSTM prediction model on some input features is reduced.

4.4. Whether there is still data in the sample time series data that needs to be trained. If the training is completed, go to step 4.5. If there is more to be trained, substitute the corresponding input x(t _i+1 ) and go to step 4.1. If the sample time series data are all involved in the training, go to the next step.

4.5. Check whether there is any battery sample data that has not participated in the training. If so, substitute a new battery sample, that is, 12-dimensional battery time series data, and go to step 4.1. If there is no new battery sample data, stop the LSTM network parameter update iteration, and output the prediction model obtained from the training.

Example:

Referring to Figure 1, the network training based on LSTM includes:

S11. Calculate the connection strength Ri (t) of the neuron _i . When the error between the model predicted value and the real value is less than the set threshold, the connection strength is calculated by formula (8).

S12, update the activation state S _i (t+1) of the neuron in the iterative process according to formula (9). That is, neurons with higher connection strength have a higher probability of transitioning to an inactive state. In this way, the dependence of the LSTM prediction model on some input features is reduced.

Referring to Figure 2, the network training of LSTM based on Dropout optimization method includes:

S21, LSTM network initialization. Given the number of input nodes m, the number of hidden nodes k, the number of output nodes n, the learning rate yita, the error threshold σ, specify the dimensions of each weight matrix and bias matrix, and use the computer to generate random numbers between 0 and 1 to assign them to each. weight matrix and bias matrix;

S22, preprocessing the original data. Take the collected data of battery float voltage, equalizing current, equalizing time, discharge cut-off voltage and discharge time every two months as a set of 60-dimensional network input data samples x(t), and the measured battery capacity data as 12 The dimensional battery time series data network outputs h(t), and starts to train the LSTM network model;

S23, the forward operation obtains the predicted capacity value h'(t _i ). Use formulas (1) and (2) to perform forward operations according to formulas (1) and (2) to obtain the predicted capacity value h'(t _i );

S24. Calculate the prediction error. The error C is calculated by formula (6). ;

S25, compare the size of the error C and the error threshold σ. If the error C is less than the threshold σ, proceed to the next step S27; if the error C is greater than the threshold σ, the network parameters are corrected by formula (7).

S26, the activation state of the neuron is updated. The connection strength of all neurons is calculated according to formula (8), and the activation state of neurons is updated according to the probability calculated according to formula (9).

S27. Whether there is still data in the sample time series data that needs to participate in training. If the training is completed, go to step S28, if there is still need for training, substitute the corresponding input x(t _i+1 ), and go to step S23. If the sample time series data all participate in the training, the next step is S28;

S28. Check whether there is battery sample data that has not participated in the training. If so, substitute a new battery sample, that is, 12-dimensional battery time series data, and go to step S23. If there is no new battery sample data, stop the LSTM network parameter update iteration, and output the prediction model obtained from the training.

Example analysis:

1) Scenario parameter setting:

The present invention tests the effectiveness of the improved prediction model proposed by the present invention by testing the actual measured data of batteries of seven 110 kV substations in a certain jurisdiction of a city in Hubei Province. Among them, the measured data of A, B, C, and D substations are used as training sets, and E, F, and G are test sets. The batteries used in the 7 substations are all 2V/300Ah batteries of "Shandong Shengyang GFMD-300C". Its specific parameter information is shown in Table 1 below:

Table 1 Substation battery sample parameter information

Table 1 battery sample parameter information

Through the above collection of samples, a total of 51 years of sample data of 7 substations were obtained. Among them, there are 4 substations in the training set and 3 substations in the test set.

2) Comparison of the prediction accuracy of different models:

This example uses BP neural network (BP), long short-term memory neural network (LSTM) and improved neural network (LS-imp) to train the battery health state prediction model respectively. The BP neural network predicts the numerical value of the capacity degradation according to the operating data, while the LSTM neural network before and after the improvement uses the initial state and historical operating data of the battery as input to directly predict its health state. The above methods are respectively used to perform step-by-step prediction, that is, to predict backward step by step according to the prediction result of the first step. In this example, four batteries are selected from the four substations in the training sample set to test the accuracy of the prediction model obtained by training. The results are shown in Figure 3(a), Figure 3(b), Figure 3(c), Figure 3(d) and Table 2.

Due to the more complex structure of LSTM-imp and LSTM models and the existence of "state memory", they have advantages in the prediction of long-term series data. The accuracy of the prediction results of these two models is significantly higher than that of the BP neural network model. Table 2 shows the statistical results of the absolute error percentage at different time steps:

Table 2 Statistical results of the mean absolute error percentage on the training sample set

Table 2 Statistical results of the percentage of absolute errors on the training sample set

As shown in Table 2, when the prediction step size is less than 2, the absolute error rates of the four different models are all lower than 5%. Compared with the BP neural network model, the LSTM-imp and LSTM models have no obvious accuracy in the prediction results. Advantage. But as the prediction time step increases, when the prediction step is 7/8/9, the mean absolute error of the BP neural network model reaches 9.71%, and the maximum absolute error is 10.45%. However, the network model with state memory of the improved LSTM model proposed in the present invention has advantages in processing time series data with a long time span. When the prediction step size is 7/8/9, the mean absolute error of the prediction result is only 2.73% , 5.53%, the maximum absolute error is 6.22%.

3) Comparison of generalization ability of different prediction models:

Models with complex structures can improve the accuracy of prediction results, but when the number of samples is insufficient, the problem of over-fitting is very likely to occur, that is, the generalization ability of the model is insufficient, so that the prediction model obtained from training cannot be applied to the test samples. In the present invention, a storage battery is randomly selected from the three substation types E to G as the test of the prediction result. The results are shown in Fig. 4(a), Fig. 4(b), and Fig. 4(c). The trained model is predicted and tested in three test sample substations, and only the improved LSTM model proposed in the present invention always maintains a high accuracy rate. Table 3 below shows the statistical results of the prediction error of the prediction model on different verification sample substations:

Table 3 Statistical results of the mean absolute error percentage on the test sample set

Table 3 Statistical results of the percentage of absolute errors on the test sample set

as shown in Table 3. The BP neural network has low accuracy in long-term prediction. Although the traditional LSTM neural network model can maintain a high prediction accuracy on the training samples, there is an over-fitting phenomenon in the test sample substations E and F. , the mean absolute error percentages are 14.87% and 11.79%, respectively, and the maximum absolute error percentages reach 18.58% and 17.67%. However, the model proposed in the present invention is improved by adopting the Dropout optimization algorithm in training, and the corresponding generalization ability is stronger. The absolute error percentage of the predicted health state of the improved model is lower than 3.0% and the maximum error percentage is lower than 5.0% in the three test substations.

The invention proposes an improved LSTM neural network for predicting the state of health of valve-regulated lead-acid batteries used in substations. Predict the storage capacity of the battery. This method uses long-term data as the input of the model, which greatly increases the complexity of the LSTM neural network and improves the accuracy of the prediction results. At the same time, in order to avoid the over-fitting phenomenon of the model obtained from training, the Dropout algorithm is improved, and the Dropout optimization algorithm is used to improve the LSTM neural network, which improves the generalization ability of the improved model. The following conclusions can be drawn from the experimental comparative analysis:

(1) The improved model has higher prediction accuracy. Due to the state memory function of LSTM, the improved model proposed in the present invention has advantages in long-term data prediction. The mean absolute error percentage of the prediction results is lower than 3.5%, and the maximum absolute error percentage is lower than 5.0%.

(2) The improved model effectively improves the generalization ability. Compared with other models, the model proposed in the present invention can obtain accurate prediction results on both the training set and the test set. The conventional LSTM network model has different degrees of overfitting.

The invention provides a multi-level LSTM prediction model based on the LSTM neural network, combined with the charging and discharging characteristics of the battery in the substation, using long-time span data as the input of the model, and improving the accuracy of the prediction results by increasing the complexity of the network model. At the same time, in order to prevent the over-fitting problem caused by the increase of model complexity, the Dropout method is introduced to improve the training process to improve the generalization ability of the model, but the use of the Dropout algorithm will increase the training time of the model by 2-3 times. In more complex cases, it may even exceed the number of iterations and fail to converge. Therefore, a dropout optimization algorithm based on neuron connection strength is proposed in the present invention. Therefore, the model has high accuracy, high efficiency, and good adaptability.

Claims (5)

1. the VRLA battery state of health prediction method based on improved LSTM neural network, is characterized in that comprising the following steps: Step 1. Collection of sample data: The floating charge voltage, equalization current, equalization duration, discharge cut-off voltage, and discharge duration input data of the battery are measured daily by the online monitoring device, and the battery capacity is measured by checking the equalization charge every two months; Step 2. Preprocessing of sample data: With n days as the time span, establish an n-dimensional sample input

in,

respectively represent the float voltage, equalization current, equalization duration, discharge cut-off voltage, and discharge duration of the battery within n days, and the battery capacity data sequence h(t _i ) is the result of multiple capacity measurements; Step 3. Build the LSTM neural network model: Taking the battery capacity data sequence h(t _i ) as the output and x(t _i ) as the input, a neural network model containing multiple LSTM neural network units is established. Each LSTM neural network unit can be regarded as an LSTM neural network at different times. The state on the span, in the initial state, by randomly generating decimals between 0 and 1, assign values to the weight matrix W and the bias matrix b in the network; Step 4. Introduce the Dropout algorithm to improve the LSTM neural network model and improve its training process. 2. the VRLA battery state of health prediction method based on improved LSTM neural network according to claim 1, is characterized in that also comprising: Step 5. Substitute the input samples in the test set into the trained model, and then 12 capacity prediction values of the battery can be obtained, and the interval between each value is 2 months. 3. the VRLA battery state of health prediction method based on improved LSTM neural network according to claim 1, is characterized in that: in described step 2, in the preprocessing of sample data, x (t _i ) is LSTM neural network The network input at time t _i , h(t _i ) is the network output at time t _i , and C(t _i ) is the unit state output of the network at time t _i ; Among them, the network input includes the float voltage, equalizing current, equalizing time, discharge cut-off voltage and discharge time of the battery; the network output is the maximum energy storage capacity of the battery, namely:

Among them, each element in x(t _i ) is a vector with a dimension of 60, representing the charge and discharge information from day t _i and 60 days back, where

is the float voltage during [t _i -60,t _i ],

are the vectors representing the equalizing charging current and charging duration, respectively. If there is no equalizing charging on the jth day, the element in the corresponding vector takes the value 0, that is,

is the record of battery discharge during [t _i -60,t _i ],

is the discharge cut-off voltage vector,

is the vector of discharge time. If the battery is not discharged on the jth day, the cut-off voltage is numerically equal to the float voltage, that is

SOH(t _i ) is the battery energy storage capacity measured on day t _i of the battery. 4. the VRLA battery state of health prediction method based on improved LSTM neural network according to claim 1, is characterized in that: described step 3 comprises the following steps: 3.1. Initialization of network hyperparameters: The set hyperparameters include: the number of input nodes m, the number of hidden nodes k, the number of output nodes n, the learning rate y _ita , the error threshold σ, and the number of LSTM nuclei w; 3.2. Weight bias initialization: In the initial state, by randomly generating decimals between 0 and 1, assign values to the weight matrix W and bias matrix b in the network. 5. the VRLA battery state of health prediction method based on improved LSTM neural network according to claim 1, is characterized in that: described step 4 comprises the following steps: Step 4.1, forward operation to predict battery capacity: According to the initially set parameters, the parameters of each gate in the LSTM model are updated according to the following formula (1), and the output results of the network are further obtained according to the formula (8):

Among them, f(t _i ), i(t _i ), o(t _i ), and C(t _i ) represent forget gate output, input gate output, output gate output and unit state respectively; σ and tanh are activation functions, Among them, σ is the sigmoid function, and tanh is the hyperbolic tangent function. The calculation formulas of the two are as follows:

W _f , Wi , W _c , and W _o represent the weight matrix of the forget gate, input gate, current input unit state and output gate, respectively, and b _f , _bi , b _c , and _{bo represent} the forget gate, input gate, current The input unit state and the bias matrix of the output gate, the 8 parameter matrices are the parameter matrices to be obtained, which are gradually optimized and updated during the training process of the model;

means element-wise multiplication, when

When acting on two vectors, the operation is as follows:

when

When acting on a vector and a matrix, the operation is as follows:

when

When acting on two matrices, the elements of the corresponding positions of the two matrices can be multiplied; W _f , W _i , W _c , W _o , b _f , b _i , b _c , b _o , these 8 parameters are trained by the network, no need to manually set specific values, but need to manually specify the matrix dimension, and set by The computer generates a random number between 0 and 1 as its initial value; Step 4.2. Correct the weight and bias parameters of the neural network according to the error of the prediction result: After the output value of the network is calculated according to formula (2), the error C between the predicted value and the actual value is calculated according to formula (6). update the parameters and biases in the network; C=|h'(t _i )-h(t _i )| (6)

In the above formula, h'(t _i ) represents the predicted capacity value output by the LSTM network, h(t _i ) represents the actual capacity value, α represents the learning rate, W=[W _f , Wi , W _c , W _{o ]} and b=[b _f , b _i , b _c , b _o ] represents the weight and bias before updating, W′ and b′ represent the weight and bias after updating; 4.3. Neuron activation state update: Calculate the connection strength of all neurons according to formula (8), and update the activation state of neurons according to the probability calculated according to formula (9); A Dropout optimization algorithm is proposed, which uses the link strength of neurons as the probability of changing its activation state to improve the convergence speed of the training process; the states of neurons are activated and non-activated, and S _i (t) takes the values 1 and 0 indicates that the neuron i is in the active state and the inactive state in the t-th iteration, respectively; the calculation formula of the connection strength R _i (t) of the neuron i is defined as:

Among them, S _j (t) is the activation state of any neuron except i in the network, and w _ij (t)∈W is the weight between neurons i and j in the t-th iteration; The activation state is updated according to the following formula (9):

That is, the higher the connection strength of neurons, the greater the probability of transitioning to an inactive state; in this way, the dependence of the LSTM prediction model on some input features is reduced; 4.4. Whether there is still data in the sample time series data that needs to participate in training; if the training is completed, go to step 4.5, if there is still need for training, substitute the corresponding input x(t _i +1), and go to step 4.1, If the sample time series data are all involved in training, go to the next step; 4.5. Check whether there is any battery sample data that has not participated in the training; if so, substitute a new battery sample, that is, the 12-dimensional battery time series data, and go to step 4.1; if there is no new battery sample data, then Stop the LSTM network parameter update iteration and output the trained prediction model.

Images