Power cabinet monitoring and early warning method and device based on big data

Abstract

The invention discloses an electric power cabinet monitoring and early warning method and device based on big data, and the method comprises the steps: collecting the data of an electric power cabinet in real time through arranging a multi-measurement-point and multi-variable sensor in the electric power cabinet; performing normalization processing on the power cabinet data to construct a power cabinet health monitoring data set; performing feature extraction by using AFFC to obtain a local feature matrix, performing feature extraction by using global local discriminant analysis GLDA to obtain a global feature matrix, and splicing and fusing the local feature matrix and the global feature matrix into a new matrix convenient to predict; a time sequence prediction model PowerMama is adopted to predict operation data of the power cabinet, and a Bayesian optimization algorithm is adopted to optimize model hyper-parameters; error correction is carried out on a prediction result of the model through multiple linear regression and RVFL, finally, prediction data are calculated through an empirical formula to obtain a health value, and whether early warning is carried out or not is judged according to the health value. The method can be applied to the modeling process of health monitoring and early warning of the electric power cabinet, and the accuracy and reliability of monitoring and early warning of the electric power cabinet are ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of power cabinet monitoring and early warning, and specifically relates to a power cabinet monitoring and early warning method and device based on big data. Background Art

[0002] In the stable operation of the power system, the distribution cabinet plays a vital role. As a key device for power distribution and control, the operating status of the distribution cabinet directly affects the stability and safety of the power system. The traditional distribution cabinet status warning method relies on complex spatial modeling processing, which has problems such as large computational complexity, high complexity, easy to cause errors and low efficiency. These problems not only affect the accuracy and reliability of the warning system, but also limit the widespread application of the warning system in the power system. In early technologies, the real-time monitoring and warning of the distribution cabinet status has technical problems such as low accuracy and efficiency of the warning. In addition, electrical fires in distribution cabinets are an important safety hazard in the operation of distribution cabinets, while traditional electrical fire warning models mainly rely on electrical fire inducing factors such as arc light, smoke, and temperature, and lack comprehensive consideration of electrical parameters such as voltage, current and power quality. These limitations limit the effectiveness of the warning system in practical applications and fail to fully capture the abnormal state of the distribution cabinet, thus affecting the safety and stability of the power system.

[0003] With the development of big data and artificial intelligence technologies, modern power cabinet monitoring and early warning methods have made significant progress. Current research and practice have begun to turn to power cabinet monitoring and early warning methods based on big data. These methods collect key data in real time by arranging multiple measurement points and multivariable sensors inside the power cabinet, and use advanced data processing technology to extract features and build prediction models. The development of these technologies has significantly improved the accuracy and reliability of power cabinet health monitoring and early warning, providing a more solid technical guarantee for the stable operation of the power system. However, there are still problems such as insufficient accuracy and inefficient hyperparameter optimization. Summary of the invention

[0004] Purpose of the invention: The present invention proposes a power cabinet health monitoring and early warning method and device based on big data to ensure the accuracy of power cabinet health monitoring and early warning.

[0005] Technical solution: The invention discloses a method for monitoring and warning the health of a power cabinet based on big data, comprising the following steps:

[0006] (1) Arrange multiple measuring points and multi-variable sensors inside the power cabinet to collect power cabinet data in real time;

[0007] (2) Normalize the power cabinet data to construct a power cabinet health monitoring dataset and divide it into training set, validation set and test set;

[0008] (3) Use the feature fusion convolutional network AFFC to extract features and obtain a local feature matrix. At the same time, use the global local discriminant analysis GLDA to extract features and obtain a global feature matrix. The local feature matrix and the global feature matrix are concatenated and fused into a new matrix that is easy to predict.

[0009] (4) The spliced ​​and fused feature matrix is ​​input into the time series prediction model PowerMamba to predict the operation data of the power cabinet, and the Bayesian optimization algorithm is used to optimize the model hyperparameters;

[0010] (5) Error correction of the model's prediction results was performed using multiple linear regression and RVFL;

[0011] (6) Calculate the predicted data using empirical formulas to obtain the health value, and determine whether to issue an early warning based on the health value.

[0012] Furthermore, the power cabinet data in step (1) includes temperature, voltage, current, concentration of harmful gases and pressure.

[0013] Furthermore, the feature extraction implementation process using the feature fusion convolutional network AFFC in step (3) is as follows:

[0014] Use convolutional layers to extract local features; extract features through convolutional layers for each variable separately:

[0015] f(x)=(W*x)+b

[0016] Among them, W is the convolution kernel, x is the input feature, * represents the convolution operation, and b is the bias term;

[0017] Feature fusion dynamically adjusts the relative importance of different feature maps through learned weights and performs fusion:

[0018] z=w1·x1+w2·x2+…+w n ·x n

[0019] Among them, z is the fused feature, x i are features of different scales, w i is the corresponding weight;

[0020] Scale alignment, upsampling, and upsampling the feature maps using bilinear interpolation so that feature maps of different scales have the same size to ensure alignment during fusion:

[0021] x upsampled =F.interpolate(x,scale f actor=2,mode=bilinear,align corners=False)

[0022] Among them, x is the feature map to be upsampled, x upsampled It is the feature map after upsampling;

[0023] Separable convolution, use separable convolution to fine-tune and enhance the fused feature map:

[0024] x separable =SeparableConvBNReLU(x,dim,dim,5)

[0025] Among them, x is the input feature map, dim is the number of channels, and 5 is the convolution kernel size;

[0026] Output an i-row local feature matrix, where each row represents the feature representation of a variable:

[0027]

[0028] Where M is the i-row local feature matrix, f i is the characteristic representation of the i-th variable.

[0029] Furthermore, the feature extraction implementation process using global local discriminant analysis (GLDA) in step (3) is as follows:

[0030] Compute the global mean of all data:

[0031]

[0032] Among them, μ g is the global mean, N is the total number of samples, x i is the i-th sample;

[0033] Compute the mean for each category:

[0034]

[0035] Among them, μ c is the local mean of category c, N c is the number of samples of category c, is the i-th sample in category c;

[0036] Compute the within-class scatter matrix of the data within each class:

[0037]

[0038] Among them, S W is the within-class scatter matrix, C is the total number of classes;

[0039] Compute the between-class scatter matrix of data between different classes:

[0040]

[0041] Among them, S B is the between-class scatter matrix, N c is the number of samples of category c;

[0042] Solving the Matrix The feature vector is used for feature extraction:

[0043]

[0044] Among them, λ is the eigenvalue and W is the corresponding eigenvector;

[0045] Project the original data onto the eigenvector to obtain the global feature matrix:

[0046] X new =XW

[0047] Where X is the original data matrix, W is the eigenvector matrix, X new is the global feature matrix;

[0048] Output the global feature matrix, and the obtained global feature matrix x new It is an i-row matrix, each row represents the global feature representation of a variable.

[0049] Furthermore, the step (3) of splicing and fusing the local feature matrix and the global feature matrix into a new matrix that is convenient for prediction is specifically to splice the two matrices together using horizontal splicing:

[0050] C=[MX new ]

[0051] Among them, C is the new matrix after concatenation, M is the local feature matrix, X new It is the global feature matrix; the concatenated matrix C contains all the features of the original two matrices and is used in the subsequent prediction model.

[0052] Furthermore, the implementation process of inputting the spliced ​​and fused matrix into the time series prediction model PowerMamba to predict the power cabinet operation data in step (4) is as follows:

[0053] Build the PowerMamba model and initialize the model parameters, including the state matrix A, input matrix B, output matrix C and projection parameter D; the state matrix A∈R N×N , represents state transition; input matrix B∈R N×1 , represents the mapping from input to state; the output matrix C∈R N×1 , represents the mapping from state to output; projection parameter D∈R N, which means a direct mapping from input to output;

[0054] Convert the continuous-time state matrix A to the discrete-time state matrix A, and convert the continuous-time input matrix B to the discrete-time input matrix B:

[0055] A=exp(ΔA)

[0056] B=(ΔA) -1 (exp(ΔA)-I)·AB

[0057] Where Δ is the time scale parameter;

[0058] Use the training set data to train the PowerMamba model and learn parameters A, B, C, and D; the goal of the model is to minimize the difference between the predicted value and the actual value; use the trained model and the concatenated matrix C to predict future time series data, and predict the next state based on the current state and input:

[0059] h k =Ah k-1 +Bw k

[0060] Among them, h k is the state at the kth time point, h k-1 is the state at the k-1th time point, w k is the input at the kth time point; output is predicted based on the state:

[0061] y k =Ch k +Da k

[0062] Among them, y k is the predicted output at the kth time point, a k is the actual input at the kth time point; outputs the prediction result, the model will output the prediction matrix P, each row represents the predicted value of a variable.

[0063] Furthermore, the implementation process of optimizing the model hyperparameters using the Bayesian optimization algorithm in step (4) is as follows:

[0064] Define the objective function, which accepts the model's hyperparameters as input and returns an evaluation metric as output. The evaluation metric will be used to guide the optimization process. The objective function is as follows:

[0065] f(θ)=MSE(y true ,y pred )

[0066] Among them, θ represents the hyperparameters of the model, y true is the true value, ypred is the value predicted by the model based on the hyperparameter θ, and MSE is the mean square error between the predicted value and the true value; an initial point θ0 is selected as the starting point for optimization; a Gaussian process is used as a proxy model to approximate the objective function, and the Gaussian process is as follows:

[0067]

[0068] Among them, m(x) is the mean function, which is usually set to 0; k(x, x′) is the kernel function used to describe the similarity between input x and x′; select the next point and use the acquisition function to select the next evaluation point θ new , the expected improvement in EI is as follows:

[0069]

[0070] Among them, f(x + ) is the best objective function value found so far; the EI function measures the expected improvement of evaluating the objective function at x relative to the current best value; at θ new The objective function is evaluated at each point and the proxy model is updated; if the preset maximum number of evaluations is reached or the objective function converges, the optimization process is stopped; otherwise, the next evaluation point is reselected.

[0071] Furthermore, the implementation process of step (5) is as follows:

[0072] First, calculate the error between the initial predictions of the PowerMamba model and the actual values:

[0073]

[0074] Among them, e is the error, y is the actual value, is the initial prediction value of the PowerMamba model; use ADF to test the stationarity of the error term e; if e is stationary, use multivariate linear regression correction; if e is non-stationary, use RVFL correction;

[0075] A multiple linear regression model is established. For the stable error term e, a multiple linear regression model is used for correction:

[0076] e=Xβ+∈

[0077] Where e is the error term, X is the input matrix containing multiple features, β is the model parameter vector, and ∈ is the error term;

[0078] Use the least squares method to estimate the model parameters β so that the sum of squared errors is minimized:

[0079]

[0080] in, is the estimated parameter vector, X T is the transpose of the input matrix X, and y is the actual error value vector;

[0081] According to the estimated parameters Calculate the error prediction value

[0082]

[0083] For non-stationary error terms, the RVFL model is used for correction. The RVFL model consists of an input layer, a random hidden layer, and an output layer. The weights and biases of the hidden layer nodes are randomly generated, and the output H of the hidden layer is calculated:

[0084] H=φ(W·X+b)

[0085] Among them, φ is the activation function, W is the randomly generated weight matrix, b is the bias vector, and X is the input feature matrix;

[0086] The output weight β is learned so that the model output is close to the actual value of the error term; the output weight is obtained by linear regression:

[0087]

[0088] Among them, H T is the transpose of the hidden layer output matrix H, is the actual error value vector;

[0089] Calculate the error prediction value, using the learned output weight β:

[0090]

[0091] Add the initial predictions from the PowerMamba model to the corrected error predictions to get the final predictions:

[0092]

[0093] in, is the final predicted value, is the initial prediction value of the PowerMamba model, is the error prediction value after the cointegration test.

[0094] Furthermore, the implementation process of step (6) is as follows:

[0095] The prediction matrix P is an i-row, 1-column matrix containing the predicted values ​​of various variables:

[0096]

[0097] Among them, p1, p2, ..., p i Represents each row of the matrix P, each row represents the predicted value of a variable;

[0098] Determine the weight coefficient and assign a weight coefficient w to each predicted value i , to reflect its impact on health value:

[0099]

[0100] Among them, w1, w2, ..., w i Represents each row of the matrix W, each row represents a weight coefficient of a predicted value, which is used to assign a weight to each predicted value to reflect its impact on the health value;

[0101] Calculate the weighted sum, multiply each predicted value by its corresponding weight coefficient, and then sum them to get the health value H:

[0102] H=w1·p1+w2·p2+…+w i ·p i

[0103] Set the warning threshold and determine a warning threshold H threshold , when the health value H exceeds this threshold, the system will issue an early warning.

[0104] A device according to the present invention includes a memory and a processor, wherein:

[0105] A memory for storing computer programs that can be run on the processor;

[0106] The processor is used to execute the steps of the power cabinet monitoring and early warning method based on big data as described above when running the computer program.

[0107] Beneficial effects: Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention adopts feature fusion convolutional network (AFFC) and global local discriminant analysis (GLDA) for feature extraction, obtains local feature matrix and global feature matrix respectively, and splices and fuses the two to form a prediction matrix that integrates local and global information, which greatly improves the expression ability of features and the accuracy of prediction; furthermore, the time series prediction model PowerMamba is combined with the Bayesian optimization algorithm to optimize the model hyperparameters, which further improves the prediction performance and adaptability of the model, ensuring that the operation data of the power cabinet can be accurately predicted under different working conditions; finally, through multivariate linear regression The RVFL is used to perform error correction on the model prediction results, and the predicted data is calculated with the empirical formula to obtain the health value, which realizes the accurate evaluation of the health status of the power cabinet, effectively reduces the risk of false alarm and missed alarm, and improves the accuracy and reliability of the early warning. In summary, the present invention can be effectively applied to the modeling process of health monitoring and early warning of the power cabinet, comprehensively considers the multi-dimensional parameter data of the power cabinet, and realizes the accurate prediction of the health status of the power cabinet. It not only improves the prediction accuracy of the model, but also helps to timely discover the potential risks of the power cabinet through real-time monitoring and early warning mechanism, and take corresponding prevention and control measures to avoid the occurrence of safety accidents. It has important practical significance and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0108] Figure 1 It is a flow chart of the power cabinet monitoring and early warning method based on big data;

[0109] Figure 2 This is a flow chart of optimizing model hyperparameters using the Bayesian optimization algorithm in the present invention;

[0110] Figure 3 This is a flow chart of error correction of multiple linear regression and RVFL in the present invention. DETAILED DESCRIPTION

[0111] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0112] like Figure 1 As shown, the present invention proposes a power cabinet monitoring and early warning method based on big data, which specifically includes the following steps:

[0113] Step 1: Collect various data of the power cabinet in real time.

[0114] Multiple measuring points and multi-variable sensors are arranged inside the power cabinet to collect various data of the power cabinet in real time, including temperature, voltage, current, concentration of harmful gases, pressure, etc.

[0115] Step 2: Normalize the power cabinet data to construct a power cabinet health monitoring dataset and divide it into training set, validation set, and test set.

[0116] Step 2.1: Convert each value of each feature to the interval [0, 1] to obtain a data set with different features:

[0117]

[0118] Where X is the original data point, X min is the minimum value of the feature in the data set, which is used to ensure that the minimum value after normalization is 0. max is the maximum value of the feature in the data set, which is used to ensure that the maximum value after normalization is 1. norm are the normalized data points.

[0119] Step 2.2: After normalization, the power cabinet health monitoring dataset is obtained and divided into training set, validation set and test set in a ratio of 6:2:2.

[0120] Step 3: Use the feature fusion convolutional network (AFFC) to extract features and obtain the local feature matrix. At the same time, use the global local discriminant analysis (GLDA) to extract features and obtain the global feature matrix. The two are spliced ​​and fused into a new matrix that is easy to predict.

[0121] Step 3.1: Construct the AFFC module and build a feature fusion convolutional network module that can dynamically adjust the relative importance of different feature maps and fuse them.

[0122] Use convolutional layers to extract local features. For each variable (temperature, voltage, current, harmful gas concentration, pressure, etc.), extract features through convolutional layers:

[0123] f(x)=(W*x)+b

[0124] Among them, W is the convolution kernel, x is the input feature, * represents the convolution operation, and b is the bias term.

[0125] Feature fusion dynamically adjusts the relative importance of different feature maps through learned weights and performs fusion:

[0126] z=w1·x1+w2·x2+…+w n ·x n

[0127] Among them, z is the fused feature, x i are features of different scales, w i is the corresponding weight.

[0128] Scale alignment, upsampling, and upsampling the feature maps using bilinear interpolation so that feature maps of different scales have the same size to ensure alignment during fusion:

[0129] x upsampled =F.interpolate(x,scale f actor=2,mode='bilinear',align c orners=False)

[0130] Among them, x is the feature map to be upsampled, x upsampled It is the upsampled feature map.

[0131] Separable convolution, use separable convolution to fine-tune and enhance the fused feature map:

[0132] x separable =SeparableConvBNReLU(x,dim,dim,5)

[0133] Among them, x is the input feature map, dim is the number of channels, and 5 is the convolution kernel size.

[0134] Output the local feature matrix. After the above steps, a 5-row local feature matrix is ​​obtained, where each row represents the feature representation of a variable:

[0135]

[0136] Among them, M is a 5-row local feature matrix, f i is the characteristic representation of the i-th variable.

[0137] Step 3.2: Calculate the global mean of all data:

[0138]

[0139] Among them, μ g is the global mean, N is the total number of samples, x i is the i-th sample.

[0140] For each category (e.g., temperature, voltage, current, hazardous gas concentration, pressure), calculate the mean for each category:

[0141]

[0142] Among them, μ c is the local mean of category c, N c is the number of samples of category c, is the i-th sample in category c.

[0143] Within-class scatter matrix: Compute the scatter matrix of the data within each class:

[0144]

[0145] Among them, S W is the within-class scatter matrix, and C is the total number of classes.

[0146] Between-class scatter matrix: Calculate the scatter matrix of data between different classes:

[0147]

[0148] Among them, S B is the between-class scatter matrix, N c is the number of samples of category c.

[0149] Solving the Matrix These feature vectors will be used for feature extraction:

[0150]

[0151] Among them, λ is the eigenvalue and W is the corresponding eigenvector.

[0152] Project the original data onto the eigenvector to obtain the global feature matrix:

[0153] X new =XW

[0154] Where X is the original data matrix, W is the eigenvector matrix, X new is the global feature matrix.

[0155] Output the global feature matrix, and the obtained global feature matrix X new It is a 5-row matrix, each row represents the global feature representation of a variable (temperature, voltage, current, harmful gas concentration, pressure).

[0156] Step 3.3: Concatenate the two matrices together using horizontal concatenation:

[0157] C=[MX new ]

[0158] Among them, C is the new matrix after concatenation, M is the local feature matrix, X new is the global feature matrix.

[0159] The concatenated matrix C will contain all the features of the original two matrices and can be used in subsequent prediction models.

[0160] Step 4: Input the concatenated feature matrix into the time series prediction model PowerMamba to predict the operation data of the power cabinet, and use the Bayesian optimization algorithm to optimize the model hyperparameters.

[0161] Step 4.1: Build the PowerMamba model and initialize the model parameters. The PowerMamba model needs to initialize some parameters, including the state matrix A, input matrix B, output matrix C and projection parameter D. These parameters can be learned from the training set data. The state matrix A∈R N×N , represents the state transition. Input matrix B∈R N×1 , represents the mapping from input to state. Output matrix C∈R N×1 , represents the mapping from state to output. Projection parameter D∈R N , which represents a direct mapping from input to output.

[0162] Discretization: The discretization process converts the continuous state space model into a discrete time model to adapt to the actual deep learning algorithm. The commonly used discretization method is the zero-order hold (ZOH) rule:

[0163] A=exp(ΔA)

[0164] Convert the continuous-time state matrix A to the discrete-time state matrix A, where Δ is the time scale parameter;

[0165] Convert the continuous-time input matrix B to the discrete-time input matrix B:

[0166] B=(ΔA)-1(exp(ΔA)-I)·AB

[0167] Train the model and use the training set data to train the PowerMamba model to learn parameters A, B, C, and D. The goal of the model is to minimize the difference between the predicted value and the actual value.

[0168] Predict future values, using the trained model and the concatenated matrix C to predict future time series data:

[0169] h k =Ah k-1 +Bw k

[0170] Predict the next state based on the current state and input, where h k is the state at the kth time point, h k-1 is the state at the k-1th time point, w k is the input at the kth time point:

[0171] y k =Ch k +Da k

[0172] Predict the output based on the state, where y k is the predicted output at the kth time point, a kis the actual input at the kth time point.

[0173] Output the prediction results. The model will output a prediction matrix P with 5 rows and 1 column, where each row represents the predicted value of a variable.

[0174] Step 4.2: If Figure 2 As shown, define the objective function, which accepts the model's hyperparameters as input and returns an evaluation indicator (such as mean square error MSE) as output. This evaluation indicator will be used to guide the optimization process. The objective function is as follows:

[0175] f(θ)=MSE(y true ,y pred )

[0176] Among them, θ represents the hyperparameters of the model, y true is the true value, y pred is the value predicted by the model based on the hyperparameter θ, and MSE is the mean squared error between the predicted value and the true value.

[0177] Select an initial point θ0 as the starting point for optimization, which can usually be selected randomly or based on experience.

[0178] Use Gaussian process (GP) as a surrogate model to approximate the target function. Gaussian process is a non-parametric Bayesian method that can be used to predict continuous valued functions. The Gaussian process is as follows:

[0179]

[0180] Among them, m(x) is the mean function, which is usually set to 0; k(x, x′) is the kernel function, which is used to describe the similarity between input x and x′. Commonly used kernel functions include the radial basis function (RBF) kernel.

[0181] Select the next point and use the acquisition function (such as expected improvement EI) to select the next evaluation point θ new The expected improvement (EI) is as follows:

[0182]

[0183] Among them, f(x + ) is the best objective function value found so far. The EI function measures the expected improvement of evaluating the objective function at x relative to the current best value.

[0184] In θ new The objective function is evaluated at , and the proxy model is updated. If the preset maximum number of evaluations is reached or the objective function converges, the optimization process is stopped; otherwise, the next evaluation point is selected again.

[0185] The training set is input into the PowerMamba model to build and train the machine learning model so that it can recognize specific features and patterns; the validation set is used to adjust model parameters and prevent overfitting during the training process to ensure the generalization ability of the model; the test set is used to evaluate the performance of the model on unseen data, providing key indicators such as accuracy, precision and recall, thereby verifying the actual application effect of the model.

[0186] Step 5: Figure 3 As shown in the figure, the prediction results of the model are corrected by multivariate linear regression and RVFL. Finally, the predicted data are calculated by empirical formula to get the health value, and whether to issue an early warning is determined based on the health value.

[0187] Step 5.1: First, calculate the error between the initial predicted value of the PowerMamba model and the actual value. The error calculation formula is as follows:

[0188]

[0189] Among them, e is the error, y is the actual value, are the initial predictions of the PowerMamba model.

[0190] Next, we use ADF to test the stationarity of the error term e. If e is stationary, we use multiple linear regression correction, and if e is non-stationary, we use RVFL correction.

[0191] A multiple linear regression model is established. For the stable error term e, we use the multiple linear regression model for correction. The model can be expressed as:

[0192] e=Xβ+∈

[0193] Where e is the error term, X is the input matrix containing multiple features, β is the model parameter vector, and ∈ is the error term.

[0194] Parameter estimation, use the least squares method to estimate the model parameter β so that the sum of squared errors is minimized. The formula for parameter estimation is:

[0195]

[0196] in, is the estimated parameter vector, X T is the transpose of the input matrix X, and y is the actual error value vector.

[0197] Calculate the error prediction value based on the estimated parameters Calculate the error prediction value

[0198]

[0199] A random vector function link (RVFL) model is constructed. For non-stationary error terms, we use the RVFL model for correction. The RVFL model consists of an input layer, a random hidden layer, and an output layer.

[0200] Randomly generate the weights and biases of the hidden layer nodes and calculate the output H of the hidden layer:

[0201] H=φ(W·X+b)

[0202] Among them, φ is the activation function, W is the randomly generated weight matrix, b is the bias vector, and X is the input feature matrix.

[0203] The output weight β is learned so that the model output is close to the actual value of the error term. The output weight can be obtained by simple linear regression:

[0204]

[0205] Among them, H T is the transpose of the hidden layer output matrix H, is the actual error value vector.

[0206] Calculate the error prediction value, using the learned output weight β, calculate the error prediction value β:

[0207]

[0208] Finally, add the initial prediction value of the PowerMamba model to the corrected error prediction value to get the final prediction value. The formula is as follows:

[0209]

[0210] in, is the final predicted value, is the initial prediction value of the PowerMamba model, is the error prediction value after the cointegration test.

[0211] Step 5.2: The prediction matrix P is a 5-row and 1-column matrix containing the predicted values ​​of temperature, voltage, current, harmful gas concentration, and pressure.

[0212]

[0213] Among them, T represents temperature, V represents voltage, I represents current, C represents the concentration of harmful gases, and S represents pressure.

[0214] Determine the weight coefficient and assign a weight coefficient w to each predicted value i , to reflect its impact on health value.

[0215]

[0216] Calculate the weighted sum, multiply each predicted value by its corresponding weight coefficient, and then sum them to get the health value H:

[0217] H=w T ·T+w V ·V+w I I+w C ·C+w S ·S

[0218] This formula combines the impact of each predicted value on health by weighted summation, and the weight coefficient w i It can be obtained based on actual application scenarios and historical data analysis.

[0219] Set the warning threshold H threshold When the health value H exceeds this threshold, the system will issue an early warning. The health value is compared with the preset threshold to determine whether an early warning is needed. The threshold should be set based on historical data and expert experience.

[0220] The present invention also provides a device, including a memory and a processor, wherein: the memory is used to store a computer program that can be run on the processor; the processor is used to execute the steps of the power cabinet monitoring and early warning method based on big data as described above when running the computer program.

[0221] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A power cabinet monitoring and early warning method based on big data, characterized in that: The following steps are involved: (1) Arrange multiple measuring points and multi-variable sensors inside the power cabinet to collect power cabinet data in real time; (2) Normalize the power cabinet data to construct a power cabinet health monitoring dataset and divide it into training set, validation set and test set; (3) Use the feature fusion convolutional network AFFC to extract features and obtain a local feature matrix. At the same time, use the global local discriminant analysis GLDA to extract features and obtain a global feature matrix. The local feature matrix and the global feature matrix are concatenated and fused into a new matrix that is easy to predict. (4) The spliced ​​and fused feature matrix is ​​input into the time series prediction model PowerMamba to predict the operation data of the power cabinet, and the Bayesian optimization algorithm is used to optimize the model hyperparameters; (5) Error correction of the model's prediction results was performed using multiple linear regression and RVFL; (6) Calculate the predicted data using empirical formulas to obtain the health value, and determine whether to issue an early warning based on the health value.

2. According to the big data-based power cabinet monitoring and early warning method of claim 1, it is characterized in that: The power cabinet data in step (1) includes temperature, voltage, current, concentration of harmful gases and pressure.

3. The method for monitoring and early warning of a power cabinet based on big data according to claim 1 is characterized in that: The process of implementing feature extraction using the feature fusion convolutional network AFFC in step (3) is as follows: Use convolutional layers to extract local features; extract features through convolutional layers for each variable separately: f(x)=(W*x)+b Among them, W is the convolution kernel, x is the input feature, * represents the convolution operation, and b is the bias term; Feature fusion dynamically adjusts the relative importance of different feature maps through learned weights and performs fusion: z=w1·x1+w2·x2+…+w n ·x n Among them, z is the fused feature, x i are features of different scales, w i is the corresponding weight; Scale alignment, upsampling, and upsampling the feature maps using bilinear interpolation so that feature maps of different scales have the same size to ensure alignment during fusion: x upsampled =F.interpolate(x,scale f actor=2,mode=‘bilinear’,align c orners=False) Among them, x is the feature map to be upsampled, x upsampled It is the feature map after upsampling; Separable convolution, use separable convolution to fine-tune and enhance the fused feature map: x separable =SeparableConvBNReLU(x,dim,dim,5) Among them, x is the input feature map, dim is the number of channels, and 5 is the convolution kernel size; Output an i-row local feature matrix, where each row represents the feature representation of a variable: Where M is the i-row local feature matrix, f i is the characteristic representation of the i-th variable.

4. The method for monitoring and early warning of a power cabinet based on big data according to claim 1 is characterized in that: The feature extraction implementation process using global local discriminant analysis GLDA in step (3) is as follows: Compute the global mean of all data: Among them, μ g is the global mean, N is the total number of samples, x i is the i-th sample; Compute the mean for each category: Among them, μ c is the local mean of category c, N c is the number of samples of category c, is the i-th sample in category c; Compute the within-class scatter matrix of the data within each class: Among them, S W is the within-class scatter matrix, C is the total number of classes; Compute the between-class scatter matrix of data between different classes: Among them, S B is the between-class scatter matrix, N c is the number of samples of category c; Solving the Matrix The feature vector is used for feature extraction: Among them, λ is the eigenvalue and W is the corresponding eigenvector; Project the original data onto the eigenvector to obtain the global feature matrix: X new =XW Where X is the original data matrix, W is the eigenvector matrix, X new is the global feature matrix; output the global feature matrix, and the obtained global feature matrix X new It is an i-row matrix, each row represents the global feature representation of a variable.

5. The method for monitoring and early warning of a power cabinet based on big data according to claim 1 is characterized in that: Step (3) of splicing and fusing the local feature matrix and the global feature matrix into a new matrix that is easy to predict is specifically to splice the two matrices together using horizontal splicing: C=[M X new ] Among them, C is the new matrix after concatenation, M is the local feature matrix, X new is the global feature matrix; the concatenated matrix C contains all the features of the original two matrices and is used in subsequent prediction models.

6. The method for monitoring and early warning of a power cabinet based on big data according to claim 1 is characterized in that: The implementation process of inputting the spliced ​​and fused matrix into the time series prediction model PowerMamba to predict the power cabinet operation data in step (4) is as follows: Build the PowerMamba model and initialize the model parameters, including the state matrix A, input matrix B, output matrix C and projection parameter D; the state matrix A∈R N×N , indicating state transition; Input matrix B∈R N×1 , represents the mapping from input to state; Output matrix C∈R N×1 , represents the mapping from state to output; Projection parameter D∈R N , which means a direct mapping from input to output; Convert the continuous-time state matrix A to the discrete-time state matrix A, and convert the continuous-time input matrix B to the discrete-time input matrix B: A=exp(ΔA) B=(ΔA) -1 (exp(ΔA)-I)·AB Where Δ is the time scale parameter; Use the training set data to train the PowerMamba model and learn parameters A, B, C, and D; the goal of the model is to minimize the difference between the predicted value and the actual value; use the trained model and the concatenated matrix C to predict future time series data, and predict the next state based on the current state and input: h k =Ah k-1 +Bw k Among them, h k is the state at the kth time point, h k-1 is the state at the k-1th time point, w k is the input at the kth time point; output is predicted based on the state: y k =Ch k +Yes k Among them, y k is the predicted output at the kth time point, a k is the actual input at the kth time point; outputs the prediction result, the model will output the prediction matrix P, each row represents the predicted value of a variable.

7. The method for monitoring and early warning of a power cabinet based on big data according to claim 1 is characterized in that: The implementation process of optimizing the model hyperparameters using the Bayesian optimization algorithm in step (4) is as follows: Define the objective function, which accepts the model's hyperparameters as input and returns an evaluation metric as output. The evaluation metric will be used to guide the optimization process. The objective function is as follows: f(θ)=MSE(y true ,y pred ) Among them, θ represents the hyperparameters of the model, y true is the true value, y pred is the value predicted by the model based on the hyperparameter θ, and MSE is the mean square error between the predicted value and the true value; an initial point θ0 is selected as the starting point for optimization; a Gaussian process is used as a proxy model to approximate the objective function, and the Gaussian process is as follows: Among them, m(x) is the mean function, which is usually set to 0; k(x,x') is the kernel function used to describe the similarity between input x and x'; select the next point and use the acquisition function to select the next evaluation point θ new , the expected improvement in EI is as follows: Among them, f(x + ) is the best objective function value found so far; the EI function measures the expected improvement of evaluating the objective function at x relative to the current best value; at θ new The objective function is evaluated at each point and the proxy model is updated; if the preset maximum number of evaluations is reached or the objective function converges, the optimization process is stopped; otherwise, the next evaluation point is reselected.

8. The method for monitoring and early warning of a power cabinet based on big data according to claim 1 is characterized in that: The implementation process of step (5) is as follows: First, calculate the error between the initial predictions of the PowerMamba model and the actual values: Among them, e is the error, y is the actual value, is the initial prediction value of the PowerMamba model; use ADF to test the stationarity of the error term e; if e is stationary, use multivariate linear regression correction; if e is non-stationary, use RVFL correction; A multiple linear regression model is established. For the stable error term e, a multiple linear regression model is used for correction: e=Xβ+∈ Where e is the error term, X is the input matrix containing multiple features, β is the model parameter vector, and ∈ is the error term; Use the least squares method to estimate the model parameters β so that the sum of squared errors is minimized: in, is the estimated parameter vector, X T is the transpose of the input matrix X, and y is the actual error value vector; According to the estimated parameters Calculate the error prediction value For non-stationary error terms, the RVFL model is used for correction. The RVFL model consists of an input layer, a random hidden layer, and an output layer. The weights and biases of the hidden layer nodes are randomly generated, and the output H of the hidden layer is calculated: H=φ(W·X+b) Among them, φ is the activation function, W is the randomly generated weight matrix, b is the bias vector, and X is the input feature matrix; The output weight β is learned so that the model output is close to the actual value of the error term; the output weight is obtained by linear regression: Among them, H T is the transpose of the hidden layer output matrix H, is the actual error value vector; Calculate the error prediction value, using the learned output weight β: Add the initial predictions from the PowerMamba model to the corrected error predictions to get the final predictions: in, is the final predicted value, is the initial prediction value of the PowerMamba model, is the error prediction value after the cointegration test.

9. The method for monitoring and early warning of a power cabinet based on big data according to claim 1 is characterized in that: The implementation process of step (6) is as follows: The prediction matrix P is an i-row, 1-column matrix containing the predicted values ​​of various variables: Among them, p1, p2, …, p i Represents each row of the matrix P, each row represents the predicted value of a variable; Determine the weight coefficient and assign a weight coefficient w to each predicted value i , to reflect its impact on health value: Among them, w1,w2,…,w i Represents each row of the matrix W, each row represents a weight coefficient of a predicted value, which is used to assign a weight to each predicted value to reflect its impact on the health value; Calculate the weighted sum, multiply each predicted value by its corresponding weight coefficient, and then sum them up to get the health value H: H=w1·p1+w2·p2+…+w i ·p i Set the warning threshold and determine a warning threshold H threshold , when the health value H exceeds this threshold, the system will issue an early warning.

10. A device, characterized in that: comprising a memory and a processor, wherein: A memory for storing computer programs that can be run on the processor; A processor is used to execute the steps of a power cabinet monitoring and early warning method based on big data as described in any one of claims 1 to 9 when running the computer program.

Images