Electric vehicle charging pile operation status prediction method based on CNN and LSTM hybrid network

Abstract

The present invention proposes a method for predicting the operating status of an electric vehicle charging pile based on a hybrid network of CNN and LSTM, which solves the problems of low efficiency and low accuracy in fault detection of charging piles. First, by acquiring the relevant historical typical fault data, information parameters, operating environment and other operating status data of the charging pile, the influencing factors of the fault pile are analyzed and the key indicators of the operating status of the charging pile are determined. Then, the key indicators are quantitatively analyzed and processed, and the weights of the key indicators are determined using the improved hierarchical analysis method. A hybrid network architecture of CNN and LSTM is built, and the training samples are trained using the hybrid network of CNN and LSTM, and the relevant parameters are adjusted in combination with the test samples to improve the classification accuracy. The model of the present invention is highly efficient and practical.

Description

Technical Field

[0001] The present invention relates to the technical field of electric vehicles, and in particular to a method for predicting the operating state of an electric vehicle charging pile based on a CNN and LSTM hybrid network. Background Art

[0002] Promoting the use of new energy vehicles, especially electric vehicles, is an inevitable choice for my country's current energy conservation and emission reduction, promoting industrial upgrading and achieving economic transformation. Charging facilities are an important foundation for the development of new energy vehicles, and the matching rate of charging facilities is one of the key factors affecting the development of new energy vehicles. However, there are widespread problems in the operation and maintenance of charging piles, such as slow fault type identification and long operation and maintenance cycle, resulting in an overall utilization rate of less than 10%. The charging pile fault detection system is the most core link in the current electric vehicle service system and is the decisive factor in determining the quality of service. In order to ensure the healthy development of the charging market, it is urgent to conduct in-depth research on charging pile fault identification technology, establish a charging pile operation status evaluation system, provide technical guidance for charging pile operation and maintenance, quickly locate charging pile faults, and improve maintenance efficiency.

[0003] Most fault detection methods are based on empirical judgment, and there are currently few methods that can be identified through deep learning technology. Limited fault detection of charging piles has been proposed, and early methods were to detect faults through manual detection and the establishment of analytical models. These methods and models can determine charging pile faults to a certain extent, but the judgment accuracy is not high. Similarly, with the increase in the number of charging piles and the increasing complexity of equipment systems, the above methods are also difficult to detect faults. Therefore, more effective methods are needed for fault detection.

[0004] Artificial intelligence and machine learning have been widely used and have become a research hotspot. This method of using neural networks to learn features and classification has achieved very good results in many fields such as pattern recognition and image classification. Among them, the convolutional neural network (hereinafter referred to as "CNN") is a variant of the feedforward neural network, first proposed by LeCun et al. It has the advantage of feature extraction and has become a powerful model for solving classification problems. In the past two decades, CNN has been successfully used to solve classification problems in many research fields. Its advantage is that it can capture the local dependencies of features in the input data. The long short-term memory network (hereinafter referred to as "LSTM") is a variant of the recurrent neural network (hereinafter referred to as "RNN"). Since its invention, it has been widely used in time series prediction to solve the problem of gradient explosion or reduction caused by long-term lag in back propagation during RNN learning, and has been proven to perform well in long-term and short-term predictions. Summary of the invention

[0005] The purpose of the present invention is to quickly predict and locate charging pile faults and improve maintenance efficiency.

[0006] In order to achieve the above object, the technical solution of the present invention is to provide a method for predicting the operating status of an electric vehicle charging pile based on a CNN and LSTM hybrid network, which is characterized by comprising the following steps:

[0007] S1. Establish a charging pile operation status indicator system

[0008] The operation status of the charging pile is reflected by the quality of the performance parameters of the charging pile, which is divided into normal state, general defect state, serious defect state and critical state. The corresponding state set is V = {v1, v2, v3, v4}, where: v1 is the normal state, v2 is the general defect state, v3 is the serious defect state, and v4 is the critical state;

[0009] S2. For the status information data of the charging pile, establish the weight of the charging pile operation status indicator

[0010] Step 1: Determine the comparison matrix: For n status indicators, first compare the importance of each status indicator two by two to form a comparison matrix X:

[0011]

[0012] In formula (1), X ij represents the comparison result of the importance between the i-th state indicator and the j-th state indicator, i = 1, 2, ..., n, j = 1, 2, ..., n;

[0013] Step 2: Determine the ranking number: Solve the comparison results of the i-th state index and the n-th state index, and sum them up to obtain the ranking index b of the i-th state index i :

[0014]

[0015] Step 3: Determine the evaluation matrix Y: Solve each element in the evaluation matrix by the sorting number, where the element y in the i-th row and j-th column is ij Calculated by the following formula (3):

[0016]

[0017] In formula (3), b max =max(b i ), b min =min(b i ),

[0018] According to the calculated evaluation matrix Y, a quasi-optimal consistent matrix Y′ is constructed. The element y in the i-th row and j-th column of the quasi-optimal consistent matrix Y′ is ij ' is calculated by the following formula (4) and (5):

[0019]

[0020] d ij =InY ij ( 5 )

[0021] S3. Establish and train a hybrid network of CNN and LSTM, select the membership of the operating time of the key indicator of the charging pile as input, and use the hybrid network of CNN and LSTM to output the state confidence interval to be predicted at the next moment. CNN extracts the characteristic attributes of the charging pile operation indicators, and then uses LSTM to associate the characteristic attributes extracted by CNN with time as a prediction tool.

[0022] Preferably, in step S1, v1 is a normal state, indicating that the charging pile equipment operates stably and all state parameters of the charging pile meet the standards; v2 is a general defect state, indicating that the charging pile equipment can operate, but the change trend of some parameters of the charging pile is developing in the direction close to the warning value and has not exceeded the warning value; v3 is a serious defect state, indicating that some characteristic parameters of the charging pile have changed greatly and are close to the standard limit, or some general parameters have exceeded the standard value, which manifests as one or more defects. The operation should be monitored and a power outage for maintenance should be arranged as soon as possible; v4 is a critical state, indicating that some state parameters of the charging pile have exceeded the warning value, which manifests as one or more critical faults, and a power outage for maintenance must be carried out immediately.

[0023] Preferably, in step s2, the status information data of the charging pile includes general selection configuration, quality of key components, operating environment and other evaluations.

[0024] The present invention proposes a method for predicting the operating state of an electric vehicle charging pile based on a CNN and LSTM hybrid network, which solves the problems of low efficiency and low accuracy of charging pile fault detection. First, by obtaining the relevant historical typical fault data, information parameters, operating environment and other operating state data of the charging pile, the influencing factors of the fault pile are analyzed and the key indicators of the operating state of the charging pile are determined. Then, the key indicators are quantitatively analyzed and processed, and the weights of the key indicators are determined by using the improved hierarchical analysis method. A CNN and LSTM hybrid network architecture is built. The CNN network mainly performs feature mining and extraction on the key indicators of the operating state of the charging pile. The LSTM network associates the key feature quantities with the time dimension. The CNN and LSTM hybrid network is used to train the training samples. The relevant parameters are adjusted in combination with the test samples to improve the classification accuracy. Finally, the MAE and RMSE are used to verify the model performance. At the same time, the model accuracy is used to verify the accuracy of the classification. The experimental results show that the model is efficient and practical. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is the charging pile operation status information model;

[0026] Figure 2 It is the basic structure of LSTM;

[0027] Figure 3 It is the basic structure of the hybrid network of CNN and LSTM;

[0028] Figure 4 It is a charging pile status prediction architecture based on CNN and LSTM hybrid network. DETAILED DESCRIPTION

[0029] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0030] The present invention provides a method for predicting the operating state of an electric vehicle charging pile based on a CNN and LSTM hybrid network, comprising the following steps:

[0031] S1. Establish a charging pile operation status indicator system

[0032] The operating status of the charging pile is reflected by the quality of the performance parameters of the charging pile, which can be divided into four risk levels: normal state, general defect state, serious defect state and critical state. The corresponding state set is V = {v1, v2, v3, v4}.

[0033] v1 is a normal state, indicating that the charging pile equipment is running stably and all state parameters of the charging pile meet the standards. v2 is a general defect state, indicating that the charging pile equipment can operate, but the change trend of some parameters of the charging pile is developing in the direction close to the warning value and has not exceeded the warning value. v3 is a serious defect state, indicating that some characteristic parameters of the charging pile have changed greatly and are close to the standard limit, or some general parameters have exceeded the standard value, showing one or more defects. The operation should be monitored and power outages and maintenance should be arranged as soon as possible. v4 is a critical state, indicating that some state parameters of the charging pile have exceeded the warning value, showing one or more critical faults, and power outages and maintenance must be carried out immediately. The operating status of the charging pile and the corresponding maintenance strategy are shown in Table 1 below.

[0034] Charging station status normal Notice serious critical Maintenance strategy Planned maintenance Priority maintenance Repair as soon as possible Repair immediately

[0035] Table 1: Charging pile status and corresponding maintenance strategy

[0036] The status information data of charging piles mainly includes the following aspects: general selection configuration, key component quality, operating environment and other evaluations. Figure 1 The multi-source state information model of the charging pile in different dimensions shown can be used to mine characteristic information representing the operating state of the charging pile.

[0037] Since complex equipment status evaluation problems are often multi-level and multi-layered, in order to better clarify the structure of the indicator hierarchy, the evaluation indicator system is generally divided into three layers: the first layer is the overall evaluation target layer; the second layer is the sub-target layer, also known as the criterion layer; the third layer is the indicator layer, also known as the target layer. In actual evaluation work, the overall goal of the research object should be determined first, and then the research object should be decomposed level by level, and the factors of each decomposition should be analyzed.

[0038] S2. Establish the weight of charging pile operation status index

[0039] The present invention improves the traditional hierarchical analysis method. In view of the strong subjectivity of the nine-scale method, a three-scale method is adopted, that is, 1, 5, and 9 scales represent equal importance, obvious importance, and extremely important, respectively. To reduce its subjectivity and enhance objectivity. Secondly, when the consistency test is not satisfied, the judgment matrix needs to be continuously corrected, and the optimal transfer matrix is ​​used. Its advantage is that the weight vector can be directly obtained, avoiding the consistency test process, thereby improving the calculation efficiency. The specific process is as follows:

[0040] Step 1: Determine the comparison matrix: For n state indicators z n First, compare the importance of each state indicator two by two, that is, indicator z j With z j Compare with each other to form a comparison matrix X:

[0041]

[0042] In formula (1), X ij It represents the comparison result of the importance between the i-th state indicator and the j-th state indicator, and the comparison result of the importance is obtained by comparing the two indicators with each other, i, j = 1, 2, ..., n.

[0043] Step 2: Determine the ranking number: Solve the comparison results of the i-th state index and the n-th state index, and sum them up to obtain the ranking index b of the i-th state index i :

[0044]

[0045] Step 3: Determine the evaluation matrix Y: Solve each element in the evaluation matrix by the sorting number, where the element y in the i-th row and j-th column is ij Calculated by the following formula (3):

[0046]

[0047] In formula (3), b max =max(b i ), b min =min(b i ),

[0048] According to the calculated evaluation matrix Y, a quasi-optimal consistent matrix Y′ is constructed. The element y in the i-th row and j-th column of the quasi-optimal consistent matrix Y′ is ij ' is calculated by the following formula (4):

[0049]

[0050] d ij =InY ij (5)

[0051] S3. Establish a hybrid network based on CNN and LSTM

[0052] Given a training data set {x n , n=1,2,...,} and the corresponding label set {t n In the present invention, the quasi-optimal consistent matrix Y′ obtained through historical data is used as the training data set, that is, the nth sample data x in the training data set n That is, it is a quasi-optimal consistent matrix Y′ obtained by the aforementioned method, and the charging pile operating states corresponding to each quasi-optimal consistent matrix Y′ constitute a corresponding label set.

[0053] CNN is trained by minimizing the mean square error or E(w) function as shown in equation (6):

[0054]

[0055] In formula (6): y(x n , w) is the given vector input x n and the vector weights w and t generated by the training algorithm n The corresponding actual label (class) is used as the predicted label (class) output by CNN.

[0056] CNN is used to extract the characteristic attributes of the charging pile operation indicators, and then LSTM is used to associate the characteristic attributes extracted by CNN with time as a prediction tool.

[0057] LSTM can be seen as a network of LSTM units. Each LSTM unit is equipped with three gates to control the flow of information: input gate, forget gate, and output gate. The structure is as follows: Figure 2 shown.

[0058] The LSTM network consists of several gates that track sequence information. The network contains three gates, the input gate i t 、Forget gate t and output gate o t . Input gate i t The function of the gate is to select the data to be stored in the next state. t Select the data that will not be stored from the state data. For example, if the calculation result is 0, it means to completely retain the information. If the calculation result is 1, it means to completely discard the information. Output gate o t Determine the state information to be routed to the output. Symbol c represents the cell state. At any time t, the equations shown in the following equations (7) to (10) are used to calculate i t 、f t , o t and

[0059] f t =σ(w f ×[h t1 , x t ]+b f ) (7)

[0060] i t =σ(w i ×[h t-1 , x t ]+b i ) (8)

[0061] o t =σ(w o x t +b o ) (9)

[0062]

[0063] In the formula, σ represents the activation function sigmoid, w f 、w i 、w o and c are weight matrices, b f 、b i 、b o and b c represents the bias vector, x t is the input at time t, which is the characteristic attribute of the charging pile operation index extracted by CNN, h t-1 is the LSTM output calculated at the previous moment, is the vector of intermediate candidates created by the tanh layer and σ is the symbolic sigmoid activation function.

[0064] The forget gate rule then computes the previous memory value to be removed from the cell state, where the previous memory value is the previous cell state.

[0065] In addition, the input gate determines the new input of the cell state. The calculation of the cell state is shown in the following equation (11):

[0066]

[0067] Then, the LSTM output As shown in the following formula (12):

[0068]

[0069] Given w y Yes t The reduced projection matrix.

[0070] LSTM output As input for calculating the forecast value at the next moment.

[0071] The prediction model of the CNN and LSTM hybrid network has two components, namely CNN and LSTM, which play their respective roles in the fault prediction model. CNN can efficiently mine the potential information of the input sequence. LSTM can better predict the input sequence. Its structure is as follows: Figure 3 shown.

[0072] S4. Charging pile operation status prediction based on CNN and LSTM hybrid network

[0073] The key indicators of charging piles are selected: general selection configuration, quality of key components, operating environment and other evaluations, and membership of operating time as input, and the confidence interval of the state to be predicted at the next moment is taken as output.

[0074] The charging pile status prediction architecture based on CNN and LSTM hybrid network is as follows Figure 4 As shown, the specific steps are as follows.

[0075] (1) Collect and select samples and divide them into training set and test set.

[0076] (2) The general selection configuration information and key component quality, operating environment, and other evaluations in the training samples are input with the membership of the operating time. In order to reduce the impact of the dispersion of the general selection configuration information data on the model, the deviation normalization method is used for normalization, as shown in the following formula (13):

[0077]

[0078] In formula (13), z min is the minimum value of the sample data, x max is the maximum value of the sample data, The data is normalized.

[0079] (3) The support vector machine is used to fit the membership function of the triangle and semi-trapezoidal combination of the key component quality to obtain the distribution function corresponding to different states.

[0080] (4) The operating environment and other evaluations are evaluated based on fuzzy statistical experiments.

[0081] (5) Determine the weight of each indicator in the general selection configuration, key component quality, operating environment and other evaluation parameters based on the hierarchical analysis method.

[0082] (6) The results in steps (3) and (4) are weighted according to the weight allocation in step (5) to obtain the comprehensive fuzzy evaluation results corresponding to v1, v2, v3, and v4, and the comprehensive evaluation results are used as data labels.

[0083] (7) Train the CNN and LSTM hybrid network model, extract the characteristic connection between key parameters and the predicted charging pile status, and obtain the prediction model parameters.

[0084] (8) The CNN and LSTM hybrid network prediction model parameters are used to predict the status of charging piles in the test set and verify the model accuracy.

[0085] S5. Use Mean Absolute Error (MAE) MAEAnd root mean square error (Root Mean Square Error, RMSE) E RMSE To study the prediction results. Its definition is shown in the following formula (14) and the following formula (15):

[0086]

[0087]

[0088] Where N is the number of predictions; Y i is the standardized actual observed value at that time; Y i is the predicted standardized value for the same period; i is the sequence number of the prediction point.

[0089] S6. 626 charging pile data confirmed to have abnormal defect faults collected and sorted from the project site and 174 charging pile status data tracked and observed after the warning / alarm on the online operation platform are selected to form the sample library of the CNN-LSTM model. 640 charging pile status data are randomly selected to form the training set, and the remaining 160 charging pile data are used to form the test set. The CNN network is used to extract the correlation characteristics between the predicted charging pile status and the general selection configuration, key component quality, operating environment and other evaluations. In order to increase the learning speed and effect and reduce the risk of the network falling into a local minimum during learning, the weight matrix in the CNN network is first randomly initialized by obeying a Gaussian distribution with a mean of 0 and a variance of 1, and then the singular value decomposition is used to obtain the cross basis matrix as the initialization value. The CNN bias term and the output layer bias term are initialized to 0, and the output layer weight matrix is ​​a random number obeying a Gaussian distribution with a mean of 0 and a variance of 1 multiplied by 0.01. The prediction model consists of an LSTM network and a Softmax network layer. The input layer scale is 100. The three Conv layers are all set to 16 groups of convolution kernels, and the convolution kernel sizes are set to 8×8, 5×5 and 3×3 respectively, followed by a Batch Normalization function and a ReLU activation function. The number of hidden layer neurons in LSTM is 128. To prevent overfitting, the signal loss rate is set to 0.2, the output layer size is 4, and the number of iterations is 100.

[0090] As shown in Table 2, the MAE and RMSE results of the BP, ARIMA and CNN-LSTM methods for predicting the operation status of the charging pile for one month. For the same test set, compared with the BP neural network and ARIMA models, the CNN-LSTM prediction model is close to the actual charging pile failure situation, and its MAE and RMSE results are the lowest.

[0091] method MAE RMSE CNN-LSTM 0.875 1.186 ARIMA 0.968 1.347 BP 1.069 1.464

[0092] Table 2 Error results of different prediction methods

Claims

1. A method for predicting the operating status of an electric vehicle charging pile based on a CNN and LSTM hybrid network, characterized in that: The following steps are involved: S1. Establish a charging pile operation status indicator system The operation status of the charging pile is reflected by the quality of the performance parameters of the charging pile, which is divided into normal state, general defect state, serious defect state and critical state. The corresponding state set is V = {v1, v2, v3, v4}, where: v1 is the normal state, v2 is the general defect state, v3 is the serious defect state, and v4 is the critical state; S2. For the status information data of the charging pile, establish the weight of the charging pile operation status indicator Step 1: Determine the comparison matrix: For n status indicators, first compare the importance of each status indicator two by two to form a comparison matrix X: In formula (1), X ij It represents the comparison result of the importance between the i-th state index and the j-th state index, i = 1, 2, ..., n, j = 1, 2, ..., n; Step 2: Determine the ranking number: Solve the comparison results of the i-th state index and the n-th state index, and sum them up to obtain the ranking index b of the i-th state index i : Step 3: Determine the evaluation matrix Y: Solve each element in the evaluation matrix by the sorting number, where the element y in the i-th row and j-th column is ij Calculated by the following formula (3): In formula (3), b max =max(b i ), b min =min(b i ), According to the calculated evaluation matrix Y, a quasi-optimal consistent matrix Y' is constructed. The element y in the i-th row and j-th column of the quasi-optimal consistent matrix Y' is ij 'Calculated by the following formula (4) and (5) the ij =InY ij (5) S3. Establish and train a hybrid network of CNN and LSTM, select the membership of the running time of the key indicator of the charging pile as input, and use the hybrid network of CNN and LSTM to output the state confidence interval to be predicted at the next moment. CNN extracts the characteristic attributes of the charging pile operation indicators, and then associates the characteristic attributes extracted by CNN with time through LSTM as a prediction tool, where: Given a training data set {x n ,n=1,2,…,N} and the corresponding label set {t n } as the goal, and use the N quasi-optimal consistent matrices Y' obtained from historical data as the training data set, that is, the nth sample data x in the training data set n That is, it is a quasi-optimal consistent matrix Y' obtained through step S1 to step S2, and the charging pile operation status corresponding to each quasi-optimal consistent matrix Y' constitutes a corresponding label set.

2. A method for predicting the operating status of an electric vehicle charging pile based on a CNN and LSTM hybrid network as claimed in claim 1, characterized in that: In step S1, v1 is a normal state, indicating that the charging pile equipment is running stably and all status parameters of the charging pile meet the standards; v2 is a general defect state, indicating that the charging pile equipment can operate, but the change trend of some parameters of the charging pile is developing in the direction of approaching the warning value and has not exceeded the warning value; v3 is a serious defect state, indicating that some characteristic parameters of the charging pile have changed greatly and are close to the standard limit, or some general parameters have exceeded the standard value, showing one or more defects, and the operation should be monitored, and power outage and maintenance should be arranged as soon as possible; v4 is a critical state, indicating that some status parameters of the charging pile have exceeded the warning value, showing one or more critical faults, and power outage and maintenance must be carried out immediately.

3. The method for predicting the operating status of an electric vehicle charging pile based on a CNN and LSTM hybrid network as claimed in claim 1, characterized in that: In step S2, the status information data of the charging pile includes general selection configuration, key component quality, operating environment and other evaluations.

Images