A charging pile state prediction method, computer readable medium and device

Abstract

The application provides a charging pile state prediction method, a computer readable medium and equipment, the method is a health index and double attention optimization GCN charging pile state prediction method, including load data acquisition, dynamic wavelet denoising, health index construction, health state prediction. The principle of the application is: firstly, the charging pile load data is collected and processed, and the server platform data and the voltage, current and power data in the charging station are integrated and processed; secondly, the original data is denoised by using dynamic threshold wavelet transform to obtain a charging pile state data matrix; thirdly, based on the voltage, current and power matrix of the charging pile load data, the health factors of the single state indicators of the charging pile are constructed by using the distance sparsity algorithm, and the health factors are fused into a health index according to the improved objective weighting method; finally, the health state of the charging pile is predicted by using a double attention optimization graph convolutional neural network.

Description

Technical Field

[0001] This application relates to the field of charging pile technology, and in particular to a charging pile status prediction method, computer-readable medium and device. Background Technology

[0002] Faced with the rapidly developing and massive electric vehicle charging station market, standardized and regulated management of electric vehicle charging equipment is essential. The scientific management and comprehensive evaluation of existing charging stations have become increasingly prominent issues. Charging station status prediction based on condition monitoring and on-site inspection enables charging station managers and charging station maintenance personnel to promptly understand and grasp the operating status of charging stations. This information serves as a basis for decision-making, guiding daily maintenance work, enabling rapid diagnosis and precise repair, determining the degree of charging station faults, and developing economical, reliable, reasonable, and efficient maintenance and periodic maintenance plans. This helps to achieve the ideal goal of extending the service life of charging stations, and proactively taking measures to prevent faults and identify potential hazards helps reduce the hidden failure rate of existing charging stations.

[0003] However, the accuracy of current predictions regarding the status of charging stations still needs improvement due to factors such as noise interference and the inherent limitations of the algorithm itself. Summary of the Invention

[0004] This application provides a charging pile status prediction method, a computer-readable medium, and a device, which has the advantage of enabling more accurate status prediction of charging piles.

[0005] The technical solution of this application is as follows:

[0006] On the one hand, this application provides a method for predicting the state of a charging pile, including the following steps:

[0007] Load data acquisition: Collect the power load information of individual charging piles to obtain charging pile load data;

[0008] Dynamic wavelet denoising: Wavelet transform is used to denoise the obtained charging pile load data;

[0009] Health Index Construction: After obtaining the charging pile load index data after noise reduction, the distance sparsity algorithm DSA is used to construct the health factor of a single load index based on the distance distribution of the data. The improved objective weighting method CRITIC is used to integrate the health factors of multiple indicators into a health index of the charging pile's operating status.

[0010] Health Status Prediction: After obtaining the health index of the charging pile, the raw data and the health index are fused to construct a multi-dimensional input data matrix D. n×mThe graph topology structure constructed by the graph convolutional neural network (GCN) is used to identify the correlation between different features. The local attention mechanism and the global attention mechanism are combined to extract the long-term and short-term time dependencies of the time series, thereby achieving accurate state prediction of charging piles.

[0011] Furthermore, in the load data acquisition step, the charging pile load data includes the voltage, current, and power data of the charging pile's input / output terminals.

[0012] Furthermore, the dynamic wavelet denoising method is as follows: assuming the noisy signal can be represented as:

[0013] x(t) = s(t) + n(t)

[0014] Where x(t) is the original signal, s(t) is the characteristic signal, and n(t) is the noise signal composed of various interferences;

[0015] Wavelet transform decomposes a signal into a superposition of wavelet functions. Wavelet transform provides decomposition in terms of scale and location, or frequency and time, through a series of convolution operations.

[0016]

[0017] Where W represents the wavelet coefficients, a is a scaling parameter inversely proportional to the center frequency; b is the translation parameter of the positioning signal; ψ * (t) is the complex conjugate of the wavelet basis ψ(t);

[0018] Determine a dynamic hard threshold and apply it to filter the decomposed wavelet coefficients:

[0019]

[0020] Where H∈R L Here, L is the decision coefficient, φ is the length of the wavelet coefficient, and φ is the feature decision output.

[0021] Calculate the feature judgment output φ: For the intra-scale dependency of wavelet coefficients, extract intra-scale feature information through one-dimensional convolution, and then obtain the final feature judgment output φ through Softmax(·). i :

[0022] φ i =Softmax(CNN(W i ))

[0023] Then, the feature judgment output φ i The threshold decision coefficient H is converted into an optimizable threshold. Logical judgment then transforms this feature decision into a hard operation, using a Gumbel distribution and a Softmax function to sample from the probability distribution.

[0024] h i =Softmax(((log(φ) i )+g i ) / ε), i=0,1

[0025]

[0026] Where g0 and g1 are samples drawn from the Gumbel(0,1) distribution, and ε is the control distribution h. i A function of smoothness Returns the sampling results of Gumbel-Softmax; where forward represents the feedforward stage and backward represents the backpropagation stage.

[0027] Furthermore, the Distance Sparsity Algorithm (DSA) determines the probability distribution of the indicator under normal conditions based on the logarithmic distance between the data, thereby judging the operating status of the indicator through the probability range.

[0028] Calculation of probability density for distance sparsity:

[0029]

[0030]

[0031]

[0032]

[0033]

[0034] Where v is the base of the logarithmic function, satisfying v∈R,v>0,v≠1, μ is the mean of the index data after transformation, σ is the standard deviation, and f(z) is the Gaussian probability density function of z.

[0035] Perform numerical conversion calculations:

[0036]

[0037] H i The health factor of a single indicator is ∈(0,1). The closer it is to 1, the more normal the health status of the load indicator; the closer it is to 0, the more abnormal the health status of the load indicator.

[0038] Furthermore, the health status prediction step includes: raw workload data. Integrating with the health index HI to construct a multi-dimensional input matrix D n×m The resulting multidimensional data matrix D n×mInput is fed into a graph convolutional neural network (GCN) for feature extraction, and a multidimensional data matrix D is generated. n×m for:

[0039]

[0040] The original load data has a length of n and a dimension of m-1. This includes data such as collected current, voltage, and power; the charging pile health index HI is represented as [d] 1m ,d 2m ...d nm ] T ;

[0041] The data attribute graph G = (V, E) consists of a node set V and an edge set E, where v i ∈V represents a node of V, e ij =(v i ,v j )∈E, representing the edge from node i to node j, N(v i )={u∈V|(v i Let {u} ∈ E be the neighborhood of node i, and let A ∈ R be the adjacency matrix. N×N As shown in the following equation, the characteristic matrix X∈R n×d ,x i ∈R d , represents the feature vector of node i;

[0042]

[0043] The spatiotemporal diagram is represented as a discrete-time dynamic diagram, i.e., G = {G1, G2, ..., G...} T The computation method of a graph convolutional neural network is as follows:

[0044]

[0045] I represents the input feature of each layer. For the initial input layer, feature I is D, so I l Let W represent the input features of the l-th layer of the network, where σ is a non-linear activation function. l This represents the weight matrix of the layer, D represents the degree matrix of the graph, and A represents the adjacency matrix.

[0046] Secondly, the output of the GCN layer is fed into an LSTM network with a dual attention mechanism to further mine the temporal correlation of features. The dual attention mechanism includes an LSTM network with a local attention mechanism and an LSTM network with a global attention mechanism. The formula for calculating the local attention value can be written as:

[0047]

[0048]

[0049]

[0050] Where s refers to the window size of the local attention mechanism; It is the final output of the LSTM network during state prediction. Hidden state with time j Match scores between them It is a trainable weight matrix. The context vector represents a single-layer neural network that takes a linear activation function as its input.

[0051] The formula for calculating the global attention value can be written as:

[0052]

[0053]

[0054]

[0055] in, It is a trainable weight matrix. It is the final output of the LSTM network during state prediction. Hidden state with time j Match scores between them The context vector represents a single-layer neural network that takes a linear activation function as its input.

[0056] Finally, the local attention values ​​and global attention values ​​are input into a feedforward neural network with a linear activation function to predict the health status. The optimization function is as follows:

[0057]

[0058] g(·) represents the ReLU() activation function, b is the bias, W is the trainable weight matrix in the neural network, L is the loss function in the neural network, and Y is the health index of the charging station.

[0059] On the other hand, this application provides a computer-readable medium storing a computer program, which, when invoked by a computer, executes the charging pile status prediction method described in the above scheme.

[0060] On the other hand, this application provides a device including a memory and a processor, wherein the memory stores a computer program, and when the computer program is called by the processor, it executes the charging pile status prediction method described in the above scheme.

[0061] In summary, the beneficial effects of this application are as follows:

[0062] 1. This invention performs dynamic threshold wavelet transform on voltage, current and power data obtained from charging piles. The threshold function reflects different processing strategies for wavelet coefficients that exceed and fall below the threshold. By dynamically eliminating noise-related components through point-by-point hard thresholding, most noise coefficients are reduced while retaining key information.

[0063] 2. Based on the noise-reduced charging pile load index data, this invention constructs a health factor for a single load index according to the distance distribution of the data using the distance sparsity algorithm (DSA). The improved objective weighting method (CRITIC) is used to integrate multiple health factors into a health index of the charging pile's operating status. The health index can reflect the real-time status of the charging pile.

[0064] 3. This invention uses the graph topology structure constructed by graph convolutional neural network (GCN) to identify the correlation between different features, introduces an LSTM network with dual attention mechanism to predict time series, and combines local attention mechanism and global attention mechanism to extract the long-term and short-term dependencies of time series, thereby achieving accurate state prediction of charging piles. Attached Figure Description

[0065] Figure 1 This is a flowchart illustrating the charging pile status prediction method of this application;

[0066] Figure 2 This is a schematic diagram of the dynamic threshold wavelet denoising model of this application;

[0067] Figure 3 This is a schematic diagram of the attribute graph structure;

[0068] Figure 4 This is a schematic diagram of the dual attention mechanism optimized GCN health status prediction model of this application. Detailed Implementation

[0069] The specific embodiments of this application are described in detail below with reference to the accompanying drawings.

[0070] Example: This application provides a charging pile status prediction method, which is a GCN charging pile status prediction method that integrates health index and dual attention optimization. The process is as follows: Figure 1 As shown, the specific implementation steps are as follows:

[0071] Step (1): Load data acquisition

[0072] The data acquisition of this invention mainly involves collecting the power load of each individual charging pile through a data acquisition device inside the charging pile. The load data includes the voltage, current, and power data at the input / output terminals of the charging pile.

[0073] Step (2): Dynamic wavelet noise reduction

[0074] like Figure 2 As shown, after obtaining the charging pile load data, wavelet transform is used to denoise the data. Wavelet transform is a process of filtering and denoising signals based on the frequency distribution characteristics of signals and noise. The decomposed wavelet coefficients are filtered using a wavelet basis, a decomposition level, an appropriate threshold, and a threshold function. This process maximizes the retention of effective signals and minimizes noise, ultimately resulting in a more effective signal.

[0075] (2.1) Wavelet function

[0076] Suppose that a noisy signal can be represented as:

[0077] x(t) = s(t) + n(t)

[0078] Where x(t) is the original signal, s(t) is the characteristic signal, and n(t) is the noise signal composed of various interferences.

[0079] Wavelet transform decomposes a signal into a superposition of wavelet functions. It provides decomposition in terms of scale and location, or frequency and time, through a series of convolution operations.

[0080]

[0081] Where W represents the wavelet coefficients, a is a scaling parameter inversely proportional to the center frequency; b is the translation parameter of the positioning signal; ψ * (t) is the complex conjugate of the wavelet basis ψ(t).

[0082] (2.2) Dynamic hard threshold

[0083] After wavelet transform, the coefficients of n(t) are small and uniformly distributed, while the coefficients of s(t) are concentrated. Therefore, setting the coefficients of noise to zero is the core idea of ​​thresholding. The key point of thresholding is determining which coefficients should be retained (features) and which coefficients should be removed (noise). The threshold can be expressed as:

[0084]

[0085] Where H∈R L φ is the judgment coefficient, L is the length of the wavelet coefficient, and φ is the feature judgment output.

[0086] First, the feature judgment output φ needs to be calculated. For the scale dependence of wavelet coefficients, this invention extracts scale feature information through one-dimensional convolution, and then obtains the final feature judgment output φ through Softmax(·). i :

[0087] φ i =Softmax(CNN(Wi ))

[0088] Then, the feature judgment output φ i The threshold decision coefficient H is converted into an optimizable threshold. Logical judgment then transforms this feature decision into a hard operation, using a Gumbel distribution and a Softmax function to sample from the probability distribution.

[0089] h i =Softmax(((log(φ) i )+g i ) / ε), i=0,1

[0090]

[0091] Where g0 and g1 are samples drawn from the Gumbel(0,1) distribution, and ε is the control distribution h. i A function of smoothness Returns the sampling results of Gumbel-Softmax; where forward represents the feedforward stage and backward represents the backpropagation stage.

[0092] Step (3): Health Index Construction

[0093] (3.1) Distance Sparsity Algorithm

[0094] After obtaining the charging pile load index data after noise reduction, the distance sparsity algorithm (DSA) is used to construct the health factor of a single load index based on the distance distribution of the data. The improved objective weighting method (CRITIC) is used to integrate the health factors of multiple indicators into a health index of the charging pile's operating status.

[0095] Distance Sparsity Algorithm (DSA) determines the probability distribution of an indicator under normal conditions based on the logarithmic distance between data points, thereby judging the operating status of the indicator through the probability range.

[0096] Calculation of probability density for distance sparsity:

[0097]

[0098]

[0099]

[0100]

[0101]

[0102] Where v is the base of the logarithmic function, satisfying v∈R,v>0,v≠1, μ is the mean of the transformed index data, σ is the standard deviation, and f(z) is the Gaussian probability density function of z.

[0103] Since this invention focuses on abnormal conditions such as overcurrent, overvoltage, overload, undercurrent, and undervoltage of charging piles, the data are all intermediate-level indicators. However, the health index needs to be constructed as a very large-scale indicator, so numerical conversion calculations are required.

[0104]

[0105] H i The health factor of a single indicator is ∈(0,1). The closer it is to 1, the more normal the health status of the load indicator; the closer it is to 0, the more abnormal the health status of the load indicator.

[0106] (3.2) Improved Objective Weighting Method

[0107] After obtaining several health factors, construct a health factor matrix H = [H1, H2...H2]. n ] T This invention employs an improved objective weighting method to weight and fuse H into a health index HI that reflects the overall status of the charging pile.

[0108] The CRITIC method uses contrast strength and index conflict to determine the weight of evaluation indicators. This invention improves upon the traditional CRITIC weighting method by using the coefficient of variation as the contrast strength for calculation. The larger the coefficient of variation, the more significant the fluctuation of the indicator. The conflict between indicators is measured by the distance correlation coefficient between them. If the correlation coefficient between two indicators is larger, it indicates a higher positive correlation between the indicators, and thus a lower conflict between them.

[0109] First, calculate the coefficient of variation:

[0110]

[0111] Where V j σ is the coefficient of variation of the j-th indicator. j It is the standard deviation of the j-th term. It is the average of the j-th term.

[0112] Secondly, calculate the distance correlation coefficient r and the conflict quantification index value:

[0113] dcov(h i ,h j )=E(|h i -h' i ||h j -h' j |)+E(|hi -h' i |)E(|h j -h' j |)-2E(|h i -h” i ||h j -h” j |)

[0114]

[0115] Among them, h i h j Represents a data vector of length n, (h i ,h j ),(h' i ,h' j ),(h” i ,h” j Let be random variables, follow the same distribution, and be independent of each other.

[0116] The quantitative indicator value for the conflict between the j-th indicator and other indicators is:

[0117]

[0118] As can be seen from the weight determination principle of the CRITIC method, the final objective weight of the indicators includes information on the comparative strength and conflict between the indicators. Let C... j Let C represent the amount of information contained in the j-th evaluation index. j It can be represented as:

[0119]

[0120] Finally, calculate the indicator weights, C of the indicators. j The larger the value, the greater the amount of information it contains, and the greater the corresponding indicator weight value. The objective weight calculation formula for the j-th indicator is as follows:

[0121]

[0122] After obtaining the health factor matrix H = [H1, H2... H... n ] T And its corresponding objective weight parameters W = [w1, w2... w n After that, the health index was

[0123] Step (4) Health Status Prediction

[0124] After obtaining the health index HI of the charging pile, the raw data is fused with HI to construct a multi-dimensional input data matrix D. n×mThe graph topology structure constructed by the graph convolutional neural network (GCN) is used to identify the correlation between different features. The local attention mechanism and the global attention mechanism are combined to extract the long-term and short-term time dependencies of the time series, thereby achieving accurate state prediction of charging piles.

[0125] Raw load data Integrating with the health index HI to construct a multi-dimensional input matrix D n×m The resulting multidimensional data matrix D n×m Input is fed into a graph convolutional neural network (GCN) for feature extraction, and a multidimensional data matrix D is generated. n×m for

[0126]

[0127] The original load data has a length of n and a dimension of m-1. This includes data such as current, voltage, and power collected by the system; the charging pile health index HI is represented as [d 1m ,d 2m ...d nm ] T .

[0128] First, the obtained multidimensional data matrix D n×m Inputting data into a Graph Convolutional Neural Network (GCN) for feature extraction is a method of graph analysis. GCN is a neural network suitable for analyzing graph data. It takes graph structure and node information as input to the model. The output type of GCN can be defined according to different downstream application tasks. GCN relies on the relationships between nodes to obtain information in the graph and optimizes the message passing mechanism between graph nodes to obtain the dependencies between different nodes. It has become a widely used graph analysis method.

[0129] like Figure 3 As shown, the attribute graph G = (V, E) consists of a node set V and an edge set E. i ∈V represents a node of V, e ij =(v i ,v j )∈E, representing the edge from node i to node j, N(v i )={u∈V|(v i Let {u} ∈ E be the neighborhood of node i, and let A ∈ R be the adjacency matrix. N×N As shown in the following equation, the characteristic matrix X∈R n×d ,x i ∈R d , where represents the feature vector of node i.

[0130]

[0131] The spatiotemporal diagram is represented as a discrete-time dynamic diagram, i.e., G = {G1, G2, ..., G...} T The computation method of a graph convolutional neural network is as follows:

[0132]

[0133] I represents the input feature of each layer. For the initial input layer, feature I is D, so I l Let W represent the input features of the l-th layer of the network, where σ is a non-linear activation function. l Let A represent the weight matrix of this layer, D represent the degree matrix of the graph, and A represent the adjacency matrix.

[0134] Secondly, such as Figure 4 As shown, the output of the GCN layer is fed into an LSTM network with a dual attention mechanism to further mine the temporal correlation of features. The dual attention mechanism includes an LSTM network with local attention and an LSTM network with global attention. The formula for calculating the local attention value can be written as:

[0135]

[0136]

[0137]

[0138] Where s refers to the window size of the local attention mechanism; It is the final output of the LSTM network during state prediction. Hidden state with time j Match scores between them It is a trainable weight matrix. This represents the context vector of a single-layer neural network that takes a linear activation function as its input.

[0139] The formula for calculating the global attention value can be written as:

[0140]

[0141]

[0142]

[0143] in, It is a trainable weight matrix. It is the final output of the LSTM network during state prediction. Hidden state with time j Match scores between them This represents the context vector of a single-layer neural network that takes a linear activation function as its input.

[0144] Finally, the local attention values ​​and global attention values ​​are input into a feedforward neural network with a linear activation function to predict the health status. The optimization function is as follows:

[0145]

[0146] g(·) represents the ReLU() activation function, b is the bias, W is the trainable weight matrix in the neural network, L is the loss function in the neural network, and Y is the health index of the charging station.

[0147] This application also provides a computer-readable medium storing a computer program, which, when called by a computer, executes the charging pile status prediction method described above.

[0148] This application also provides a device including a memory and a processor. The memory stores a computer program, and when the computer program is called by the processor, it executes the charging pile status prediction method described in the above scheme.

[0149] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several modifications and improvements can be made without departing from the inventive concept of this application, and these all fall within the protection scope of this application.

Claims

1. A charging pile state prediction method, characterized in that, The method comprises the following steps: Load data acquisition: collecting the power load of the charging pile to obtain charging pile load data; Dynamic wavelet denoising: using wavelet transform to denoise the obtained charging pile load data; Health index construction: after obtaining the denoised charging pile load index data, the distance sparsity algorithm DSA is used to construct the health factor of a single load index according to the distance distribution of the data, and the improved objective weighting method CRITIC is used to fuse multiple index health factors into the health index of the charging pile operation state; Health state prediction: after obtaining the health index of the charging pile, the original data and the health index are fused to construct a multi-dimensional input data matrix, the graph topology structure constructed by the graph convolutional neural network GCN is used to identify the correlation between different features, and the local attention mechanism and the global attention mechanism are combined to extract the time dependence of the time series long and short term, so as to realize accurate state prediction of the charging pile.

2. The charging pile state prediction method according to claim 1, characterized in that, In the load data acquisition step, the charging pile load data includes voltage, current and power data of the input / output end of the charging pile.

3. The charging pile state prediction method according to claim 1, characterized in that, The method of dynamic wavelet denoising is that a signal with noise is expressed as: x(t) = s(t) + n(t) Where x(t) is the original signal, s(t) is the characteristic signal, and n(t) is the noise signal composed of various disturbances. Wavelet transform is to decompose the signal into a series of superpositions of wavelet functions. Wavelet transform provides scale and position, or frequency and time decomposition through a series of convolution operations: where W represents the wavelet coefficients, a is a scaling parameter inversely proportional to the center frequency; b is a translation parameter of the positioning signal; ψ * (t) is the complex conjugate of the wavelet basis ψ(t). The dynamic hard threshold value is determined and the wavelet coefficients after decomposition are filtered by the dynamic hard threshold value: where H ∈ R L is a decision coefficient, L is the length of the wavelet coefficients, and φ is the characteristic decision output. The calculation characteristic judgment output φ: for the scale intra-dependence of the wavelet coefficient, the scale intra-feature information is extracted by one-dimensional convolution, and then the final output φ of the characteristic judgment is obtained by Softmax(·) i : φ i = Softmax(CNN(W i )) Then, the feature judgment output φ i Converts to the threshold judgment coefficient H that can be optimized, and directly converts the feature decision to a hard operation through logical judgment, samples on the probability distribution by using the Gumbel distribution and the Softmax function: h i = Softmax(((log(φ i )+ g i ) / ε), i = 0, 1 where g0and g1are samples drawn from a Gumbel(0, 1) distribution, and ε is a control distribution h i a function of smoothness, return the sampling result of Gumbel-Softmax; wherein forward represents the forward stage; backward represents the back propagation stage.

4. The charging pile state prediction method according to claim 3, characterized in that, The distance sparsity algorithm DSA judges the probability distribution of the index under normal conditions according to the logarithmic distance between the data, so as to judge the operation state of the index through the probability range. The probability density calculation of distance sparsity is as follows: Where v is the base of the logarithmic function, v R, v > 0, v ≠ 1, μ is the average value of the transformed index data, σ is the standard deviation, and f(z) is the Gaussian probability density function of z. Numerical conversion calculation is performed: M = max{|Z i - μ|}, H i ∈(0,1) and is a health factor of a single index. The closer to 1, the more normal the health degree of the running state of the load index. The closer the distance 0 is, the more abnormal the operation state of the load index is.

5. The charging pile state prediction method according to claim 4, characterized in that, After obtaining the health factors of several indexes, a health factor matrix H = [H1, H2... H n ] T ; the improved objective weighting method is used to weight and fuse H into a health index HI reflecting the overall state of the charging pile. First, the coefficient of variation is calculated: where V j is the coefficient of variation of the jth metric, σ j is the standard deviation of the jth metric, is the mean of the jth metric; Second, the distance correlation coefficient r and the conflict quantization index value are calculated: dcov(h i ,h j ) = E(|h i -h’ i ||h j -h' j |) + E(|h i -h’ i |) E(|h j -h' j |) - 2E(|h i -h” i ||h j -h” j |) dCov 2 (h i ) = (1 / T 2 )∑ j,k ((h ij -h ik )-∑ k (h ij -h ik )-∑ j (h ij -h ik )+∑ j,k (h ij -h ik )) 2 wherein h i , h j represents a segment of an index data vector of length n, (h i , h j ), (h' i , h' j ), (h" i , h" j ) are random variables, subject to the same distribution, and independent of each other; The conflict quantization index value of the jth index with other indexes is: Let C j represents the amount of information contained in the jth evaluation index, then C j is represented as: Finally, the index weight is calculated, and the C j value is larger, the information quantity is larger, and the corresponding index weight value is larger. The objective weight of the jth index is calculated as follows: After getting the health factor matrix H = [H1, H2... H n ] T and its corresponding objective weight parameter W = [w1, w2... w n ], the health index is 6. The charging pile state prediction method according to claim 1, characterized in that, The health status prediction step comprises: original load data Fusing with the health index HI to construct a multi-dimensional input matrix D n×m The obtained multi-dimensional data matrix D n ×m Input into a graph convolutional neural network GCN for feature extraction, and the multi-dimensional data matrix D n×m is: Wherein, the length of the original load data is n, and the dimension is m-1, including the collected current, voltage and power data; the charging pile health index HI is expressed as [d 1m ,d 2m ...d nm ] T ; The data attribute graph G = (V, E) is composed of a node set V and an edge set E, v i ∈ V represents a node of V, e ij = (v i , v j ) ∈ E represents an edge from node i to node j, N(v i ) = {u ∈ V | (v i , u) ∈ E} is the neighborhood of node i, and the adjacency matrix A ∈ R N×N The feature matrix X ∈ R n×d , x i ∈ R d represents the feature vector of node i; The space-time graph is expressed as a discrete-time dynamic graph, that is, G = {G1, G2,..., G T}, and the calculation manner of the graph convolutional neural network is: I represents the input features for each layer, for the initial input layer, the features I are D, so I l represents the input features of the / th layer of the network, σ is a non-linear activation function, W l represents the weight matrix of the layer, D represents the degree matrix of the graph, A represents the adjacency matrix; Second, the output of the GCN layer is transmitted to the LSTM network with a double attention mechanism to further mine the time correlation of the features; the double attention mechanism includes an LSTM network with a local attention mechanism and an LSTM network with a global attention mechanism, and the formula for calculating the local attention value is written as: where s refers to the window size of the local attention mechanism; is the final output of the LSTM network at state prediction is the matching score between the hidden state at time j , is a trainable weight matrix, denotes the context vector of a single layer neural network with an input linear activation function; The formula for calculating the global attention value is written as: wherein, is a trainable weight matrix, is the final output of the LSTM network at state prediction time is the matching score between the hidden state at time j denotes the context vector of a single layer neural network with an input linear activation function;​ Finally, the local attention value and the global attention value are input into the feedforward neural network with a linear activation function to predict the health state, and the optimization function is as follows: g(·) represents the ReLu() activation function, b is the bias, W is the trainable weight matrix in the neural network, L is the loss function in the neural network, and Y is the health index of the charging pile.

7. A computer readable medium characterized by The computer readable medium stores a computer program, and the computer program is invoked by a computer to execute the charging pile state prediction method in any one of claims 1-6.

8. An apparatus, comprising: The computer readable medium stores a computer program, and the computer program is invoked by a computer to execute the charging pile state prediction method in any one of claims 1-6.

Images