Substation evaluation method and system fused with attention mechanism neural network model

Abstract

The invention discloses a substation evaluation method and system fused with an attention mechanism neural network model, and the method comprises the steps: organizing data related to the operation of a substation according to a time sequence, forming multi-dimensional time sequence data, building a data prediction model with a prediction set of the substation at a future moment as a target function, and carrying out the prediction of the data. The method comprises the following steps: training a data prediction model by taking multi-dimensional time sequence data as a sample data set, inputting real-time data of a transformer substation into the data prediction model for solving to obtain a prediction set of the transformer substation at a future moment, and constructing a multi-index scoring model by taking a transformer substation transformation time sequence score as a target function, and inputting the prediction set of the transformer substation at the future moment into the multi-index scoring model to obtain a transformer substation transformation time sequence score, and predicting the transformation demand degree of the equipment. The method not only can effectively capture the characteristics of the complex time sequence data and improve the accuracy of substation transformation evaluation, but also can improve the network expression ability through the convolutional neural network and reduce the overall calculation cost of the neural network.

Description

Technical Field

[0001] The present invention relates to a substation evaluation method, and particularly to a substation evaluation method and system integrating an attention mechanism neural network model. Background Art

[0002] With the continuous development of the power grid, as an important part of the power grid, substations play an important role in aspects such as power transmission, distribution, regulation, and protection. The operation and maintenance of substation equipment are an important part of the normal operation of the power system and an important link for power enterprises to ensure normal power supply.

[0003] As substation equipment gradually ages and its performance declines, especially the aging problems of core equipment such as transformers, switches, and protection equipment pose great hidden dangers to the stable operation of the power grid, making it an urgent need for the power industry to technically transform existing substations.

[0004] Substation transformation requires the analysis and prediction of a large amount of time-series data to formulate a reasonable transformation plan. Existing analysis and prediction methods usually adopt statistical analysis models. The statistical analysis model collects and analyzes historical monitoring data to dynamically adjust the statistical scope and compares abnormal data with a multi-source fusion model to determine whether the abnormal data is related to a power event; a secondary inspection is carried out according to the matching situation between the abnormal data and known events to confirm whether the abnormal data indicates equipment or system failures.

[0005] Although this statistical analysis model can improve the flexibility and accuracy of data processing, it still has the following defects:

[0006] Traditional statistical analysis models are difficult to effectively capture the characteristics of complex time-series data, resulting in inaccurate and unreliable analysis results.

[0007] Disclosing the information of this background art section is only intended to increase the understanding of the overall background of the present application, and should not be regarded as an admission or any form of implication that this information constitutes prior art already known to those of ordinary skill in the art. Summary of the Invention

[0008] The object of the present invention is to overcome the defect in the prior art that it is difficult to effectively capture the characteristics of complex time-series data, and provides a substation evaluation method and system integrating an attention mechanism neural network model that can effectively capture the characteristics of complex time-series data.

[0009] To achieve the above object, the technical solution of the present invention is:

[0010] A substation evaluation method integrating an attention mechanism neural network model, the evaluation method includes:

[0011] S1. Collect data related to the operation of the substation from the substation monitoring system and equipment management system, organize the data related to the substation operation in time series to form multi-dimensional time series data, and preprocess the multi-dimensional time series data to obtain a sample data set;

[0012] S2. Construct a data prediction model. The prediction target of the data prediction model is the prediction set of the substation data at future times. Train the data prediction model through the sample data set, input the real-time data of the substation into the trained data prediction model, solve the data prediction model, and obtain the prediction set of the substation at future times;

[0013] S3. Construct a multi-index scoring model with the substation renovation time series score as the objective function, input the prediction set of the substation at future times into the trained multi-index scoring model to obtain the substation renovation time series score, and predict the renovation requirement degree of the equipment according to the substation renovation time series score and in combination with the scoring interval;

[0014] The prediction set of the substation at future times is a health status index.

[0015] In the above S1, the data related to the substation operation includes health status indicators, defect and fault indicators, maintenance records, environmental indicators, equipment usage load, time characteristics, and other related characteristics;

[0016] The health status indicators include the equipment operation status, equipment insulation status, and equipment aging degree. The defect and fault indicators include historical fault records, fault warning signals, and characteristic changes before the occurrence of faults. The maintenance records include maintenance time points, maintenance contents, and changes in the equipment status after maintenance. The environmental indicators include environmental parameters, weather conditions, and geographical locations. The equipment usage load includes electrical parameters, load rates, and load change trends. The time characteristics include time stamps, seasonal characteristics, and holidays or special events. The other related characteristics include equipment types, equipment manufacturers and models, and historical operation data.

[0017] The above S2 includes:

[0018] The data prediction model includes a convolutional neural network based on the attention mechanism and a long short-term memory network. The convolutional neural network based on the attention mechanism is used to extract the original features in the training data. The convolutional neural network based on the attention mechanism includes an attention module, a convolutional unit, and a pooling unit. The branch input sequence of the attention module covers the input of the convolutional neural network, and deep convolutional networks and pooling layers are stacked, so that the input receptive field corresponding to the features is expanded, and the model obtains more comprehensive context information, thereby learning the importance degree of the current local sequence features;

[0019] The objective function of the data prediction model includes:

[0020]

[0021] In the above formula, is the model prediction value, and the model prediction value is the prediction set of the substation at a future time. CNN is a convolutional neural network, LSTM is a long short-term memory recurrent neural network, Attn represents the attention mechanism, and y t is the real device status index, which is extracted from the sample dataset. λ ∈ [0.01, 0.1] represents the L2 regularization coefficient, which is used to prevent overfitting. θ represents all the trainable parameters of the model. All the trainable parameters of the model include the convolutional kernel size, the number of convolutional kernels, the number of LSTM layers, and the model learning rate of the CNN. x t-n:t represents the real-time data of the substation input from time t - n to time t.

[0022] The output features of the convolutional neural network based on the attention mechanism are multiplied element-wise by the saliency features output by the corresponding attention mechanism module, and the importance degree of the features is reflected by the high and low values of the numerical values, thus completing the discrimination of important features.

[0023] The long short-term memory network is used to control the transmission and update of information through neurons and gate structures, mine the long-term dependencies in the substation time-series data, and predict the prediction set of the substation at a future time.

[0024] In S2, the original input of the CNN module is the sample dataset. The CNN module includes a convolutional layer, a pooling layer, and a fully connected layer. The convolutional layer performs a convolution operation by sliding the convolutional kernel on the data to learn local feature patterns. The pooling layer is used to reduce the data dimension, reduce the amount of calculation, and at the same time retain important feature information. The fully connected layer is used to integrate the extracted features.

[0025] The convolutional layer of the CNN model uses a "convolutional kernel" for local perception. The size of the convolutional kernel defines the receptive field of the convolution operation, and the ReLU function is selected as the activation function, including:

[0026] (Convolutional layer) H (l) = f(W (l) *H (l-1) + b (l) );

[0027] (Activation function) ReLU(x) = max(0, x);

[0028] In the above formula, H (l) is the output feature of the l-th layer convolution, l is the depth of the convolutional neural network, f() represents the activation function, and W (l)is the weight matrix of the convolutional kernel in the l-th layer, * represents the convolution operation, and H (l-1) is the convolutional output feature in the (l-1)-th layer, and b (l) represents the bias term, and x is the input vector from the previous layer of the neural network;

[0029] Then, the local features extracted by the CNN are input into the LSTM module of the fusion attention mechanism. The LSTM module controls the transmission and update of information through neurons and gate structures, mines the long-term dependencies in the substation time series data, and captures the dynamic change laws of the data in the time series; in the hidden layer of the LSTM model, through the attention mechanism, the attention weights are calculated on the hidden state processed by the LSTM to highlight the key parts of the features and learn the global dynamic dependencies of the time series data;

[0030] Select the Adam optimization algorithm to train the model, and set appropriate hyperparameters such as the learning rate and decay rate, which can enable the model to converge quickly;

[0031] Use the training set data to train the CNN-LSTM-attention time series analysis model. Through the collaborative work of the CNN network, LSTM network and attention mechanism, learn the patterns and laws in the data, verify the performance of the model, and continuously adjust the parameters of the model to minimize the loss function;

[0032] The LSTM network mentioned above includes four parts: a forget gate, an input gate, an output gate, and an update module, which cooperate to store, update, and output information. The input gate determines how much of the input information at the current time step should be remembered, the forget gate determines how much of the memory state at the previous time step should be forgotten, the output gate determines how much of the memory state at the current time step should be output to the next layer or the output layer, and the memory unit is responsible for storing and transmitting context information;

[0033]

[0034] In the above formula, f t is the output of the forget gate, i t is the output of the input gate, is the candidate memory state, C t is the updated memory state, o t is the output gate, h t is the output of the hidden state, σ is the activation function, W f is the forget gate, W i is the input gate, W C is the candidate memory state, W o is the weight term of the output gate, b f is the forget gate, b i is the input gate, b Cis the candidate memory state, b o is the bias term of the output gate, is the Hadamard product, and the Hadamard product represents the multiplication of vector elements, h t-1 is the hidden state at time t-1, x t is the input at time t, and the x t 's original input is the feature vector output by the CNN;

[0035] The attention mechanism is used to dynamically allocate weights according to the importance degree of different parts of the input data to the output. The expression of the attention mechanism is:

[0036] Improved formula: Feature fusion:

[0037] In the above formula, α ij is the attention weight, which reflects the correlation strength of the feature at the i-th moment to the j-th moment. The attention weight is calculated by the normalized exponential function, e ij is the attention score, N is the total number of categories, e ik is the attention score of the K-th component, Q (t) is the query vector, and the query vector is obtained from the LSTM hidden state obtained, is the query vector at the i-th moment, d k is the key vector dimension, K (t) is the key vector, is the key vector at the j-th moment, and the key vector dimension d k = 64, W q and W k are the weight vectors, is the state of the LSTM at time t, is the feature of the CNN at time t, F fusion is, LayerNorm represents the layer normalization operation, α is the attention weight, represents the channel concatenation operation,

[0038] represents the exclusive NOR operation.

[0039] In the S3, the comprehensive scoring function of the multi-index scoring model includes:

[0040]

[0041] In the above formula, Score(t) is the substation renovation timing score of the device at time t, and the substation renovation timing score represents the possibility that the substation needs to be renovated. φ represents the index function, and φ = [Aging, Health, F, M, E, L] T , ω represents the weight coefficient, ε represents the score threshold. The threshold θ ∈ [0.6, 0.8] is set as the renovation trigger boundary, and the sigmoid function is introduced to map the score to the interval (0, 1).

[0042] In S3, the device operation age index includes:

[0043] Aging(t) = t / T max ;

[0044] In the above formula, Aging(t) is the operation age of the device at time t, t is the current time, and T max is the design life of the device;

[0045] The health status index includes:

[0046]

[0047] In the above formula, Health(t) represents the health status evaluation of the device at time t, w a , w s , w f all represent dynamic weights, and the weight distribution satisfies w a + w s + w f = 1, e -λt is the aging attenuation term. The aging attenuation term is used to strengthen the non-linear degradation effect of the device service time through the exponential function, which is set according to the device situation. λ ∈ [0.01, 0.1] is the aging attenuation coefficient, and the aging attenuation coefficient is used to control the attenuation speed. S(t) is the evaluation status of the device, which is obtained by weighting each index in the prediction set of the future time of the substation. ∑faults(t) is the weighted sum of the number of device failures in the historical time period;

[0048] The historical defect and fault index includes:

[0049]

[0050] In the above formula, F(t) is the historical defect and fault index of the device at time t, F hist(t - 1) is the historical defect failure index at the previous evaluation, and S(t - 1) is the equipment status at the previous evaluation; g1(·) is a function considering the historical defect failure and equipment status, used to calculate the defect failure index at the current moment; w1 and w2 are the weight parameters of the historical defect failure index, used to adjust the importance of the historical defect failure and equipment status in the current defect failure index;

[0051] The historical maintenance situation index includes:

[0052]

[0053] In the above formula, M(t) is the historical maintenance situation index of the equipment at time t, and g2(·) is a function considering the historical maintenance situation and equipment maintenance situation, used to calculate the maintenance situation index at the current moment; M hist (t - 1) is the historical maintenance situation index at the previous evaluation, and R(t - 1) is the equipment maintenance situation at the previous evaluation; w7 and w8 are the weight parameters of the historical maintenance situation index, used to adjust the importance of the historical maintenance situation and equipment maintenance situation in the current maintenance situation index;

[0054] The environmental factor index includes:

[0055] E(t) = w9 * T(t) + w 10 * H(t) + w 11 * P(t);

[0056] Among them, E(t) is the environmental factor index of the equipment at time t, T(t) is the temperature of the environment where the equipment is located, which can be the data collected by the temperature sensor; H(t) is the humidity of the environment where the equipment is located; P(t) is the air pressure or other influencing factors of the environment where the equipment is located; w9, w 10 and w 11 are the weight coefficients corresponding to the environmental factor index.

[0057] The equipment usage load index includes:

[0058] L(t) = w 12 * I(t) + w 13 * U(t) + w 14 * R(t);

[0059] In the above formula, L(t) is the equipment usage load index of the equipment at time t, I(t) is the load current of the equipment at time t, indicating the amount of current carried by the equipment; U(t) is the operating voltage of the equipment; R(t) is the cumulative operating duration of the equipment, indicating the total time the equipment has been running; w 12 、w 13 and w 14 are the weight coefficients of the equipment usage load index.

[0060] In S3, when 0.80 < Score(t) ≤ 1.00, there are significant problems in the aging and fault history of this type of equipment, with high renovation requirements, and renovation measures need to be taken immediately; when 0.60 < Score(t) ≤ 0.80, this type of equipment shows medium-level aging or faults, and renovation needs to be considered as a priority; when 0.40 < Score(t) ≤ 0.60, this type of equipment shows minor aging and relatively good operating conditions, and no renovation is required for the time being; when 0.00 < Score(t) ≤ 0.40, the operation is stable and no renovation is needed.

[0061] A substation evaluation system integrating an attention mechanism neural network model, which is used to execute the substation evaluation method of the integrated attention mechanism neural network model described above, specifically including: a data collection module, a data prediction module, and an evaluation module;

[0062] The data collection module is used to collect data related to the operation of the substation from the monitoring system and equipment management system of the substation, organize the data related to the operation of the substation in time series to form multi-dimensional time series data, and preprocess the multi-dimensional time series data to obtain a sample data set;

[0063] The data prediction module is used to construct a data prediction model. The prediction target of the data prediction model is the prediction set of the substation data at future moments. The data prediction model is trained through the sample data set, and the real-time data of the substation is input into the trained data prediction model to solve the data prediction model to obtain the prediction set of the substation at future moments;

[0064] The evaluation module is used to construct a multi-index scoring model with the substation renovation time series score as the objective function, input the prediction set of the substation at future moments into the trained multi-index scoring model to obtain the substation renovation time series score, and predict the renovation requirement degree of the equipment according to the substation renovation time series score and in combination with the scoring interval.

[0065] A substation evaluation device integrating an attention mechanism neural network model, including a memory and a processor. The memory is used to store computer program code and transmit the computer program code to the processor;

[0066] The processor is used to execute the substation evaluation method of the integrated attention mechanism neural network model according to the instructions in the computer program code.

[0067] A computer program product includes a computer program, and the computer program is executed by the processor to perform the substation evaluation method of the integrated attention mechanism neural network model described above.

[0068] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0069] 1. In the substation evaluation method integrating an attention mechanism neural network model of the present invention, it includes a convolutional neural network and a long short-term memory network integrating an attention mechanism. The present invention combines the advantages of both to accurately and efficiently extract local features and global dependencies of time-series data, effectively capture the features of complex time-series data, and obtain the time-series score for substation transformation by inputting the real-time data of the substation into the CNN-LSTM-attention model. The time-series score for substation transformation is a probability value, and the higher the score, the higher the possibility that the substation needs to be transformed. When the time-series score for substation transformation is greater than the threshold, it indicates that the substation needs to be transformed. Therefore, this design can effectively capture the features of complex time-series data and effectively improve the accuracy of substation transformation evaluation.

[0070] 2. In the substation evaluation method integrating an attention mechanism neural network model of the present invention, the convolutional neural network uses the rectified linear unit (ReLU) as the activation function of the convolutional neural network. The ReLU function is in the form of max(0,x), which has the advantages of being similar to the activation mechanism of biological neurons, simple and fast derivative calculation, etc. It avoids the problem of gradient disappearance, helps the training of deep networks, better retains the non-linear relationship of the input data, improves the expression ability of the network, and reduces the overall computational cost of the neural network. Therefore, this design can improve the expression ability of the network through the convolutional neural network and effectively reduce the overall computational cost of the neural network.

[0071] 3. In the substation evaluation method integrating an attention mechanism neural network model of the present invention, the long short-term memory network uses Tanh as the state transition function, which can better capture the positive and negative values of information and reduce the impact of output imbalance. At the same time, it uses Sigmoid as the gating function to precisely control the flow of information. This combined activation function design enables the LSTM to more effectively process and store long-term dependency information, while adapting to complex time-series data processing tasks, ensuring that the model can converge more quickly and effectively during the training process, thereby improving the overall performance. Therefore, this design can ensure that the model can converge more quickly and effectively during the training process through the long short-term memory network and effectively improve the overall performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 is the flowchart of the method of the present invention.

[0073] Figure 2 is the structural diagram of the CNN-LSTM-attention model in the present invention.

[0074] Figure 3It is the model architecture diagram of the CNN model in the present invention.

[0075] Figure 4 It is the basic unit structure diagram of the LSTM model in the present invention.

[0076] Figure 5 It is the working principle diagram of the attention mechanism in the present invention.

[0077] Figure 6 It is the structure diagram of the system described in the present invention.

[0078] Figure 7 It is the structure diagram of the device described in Embodiment 3. Detailed implementation manners

[0079] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0080] Embodiment 1:

[0081] Refer to Figure 1 , a substation evaluation method integrating an attention mechanism neural network model, the evaluation method includes:

[0082] S1. Collect data related to the operation of the substation from the monitoring system and equipment management system of the substation, organize the data related to the operation of the substation in time series to form multi-dimensional time series data, and preprocess the multi-dimensional time series data to obtain a sample data set;

[0083] S2. Construct a data prediction model, the prediction target of the data prediction model is the prediction set of the substation data at future moments, train the data prediction model through the sample data set, input the real-time data of the substation into the trained data prediction model, solve the data prediction model, and obtain the prediction set of the substation at future moments;

[0084] S3. Construct a multi-index scoring model with the substation transformation time series score as the objective function, input the prediction set of the substation at future moments into the trained multi-index scoring model to obtain the substation transformation time series score, and predict the degree of transformation demand of the equipment according to the substation transformation time series score in combination with the scoring interval.

[0085] In the above S1, the data related to the operation of the substation includes health status indicators, defect and fault indicators, maintenance records, environmental indicators, equipment usage load, time characteristics, and other related characteristics;

[0086] The health status indicators include the operating status of the equipment (such as temperature, vibration, noise, etc.), the insulation status of the equipment (such as insulation resistance, partial discharge, etc.), and the degree of equipment aging (such as operating time, number of historical faults, etc.). The defect and fault indicators include historical fault records (such as fault type, occurrence time, duration, etc.), fault warning signals (such as abnormal current, voltage fluctuation, etc.), and characteristic changes before the occurrence of faults (such as sudden temperature rise, abnormal vibration, etc.). The maintenance records include maintenance time points (such as the last maintenance time, maintenance frequency, etc.), maintenance contents (such as replaced components, types of faults repaired, etc.), and changes in the equipment status after maintenance (such as performance recovery, changes in operating parameters, etc.). The environmental indicators include environmental parameters (such as temperature, humidity, air pressure, etc.), weather conditions (such as extreme weather conditions like lightning strikes, heavy rain, high temperature, etc.), and geographical locations (such as altitude, geological conditions, etc.). The equipment usage load includes electrical parameters (such as current, voltage, power, load rate (such as the ratio of the equipment operating load to the rated load), and load change trends (such as peak load, average load, etc.). The time characteristics include time stamps (such as year, month, day, hour, etc.), seasonal characteristics (such as high load in summer, low load in winter, etc.), and holidays or special events (such as power grid dispatching, major events, etc.). The other relevant characteristics include equipment types (such as transformers, circuit breakers, cables, etc.), equipment manufacturers and models, and historical operation data (such as long-term operation trends, periodic changes, etc.);

[0087] Organize the above features in a time series to form multi-dimensional time series data. The features at each time step can be represented as a vector, and the data of multiple time steps can form a matrix as the input of the convolutional neural network;

[0088] The S1 includes:

[0089] S1.1. The time series data of the operation of substation equipment is obtained through the monitoring system and equipment management system of the substation. The time series data of the operation of substation equipment includes historical electrical parameter data, equipment operation status data, operation environment data, and previous substation transformation records;

[0090] S1.2. Remove the noise and outliers in the time series data, and identify and correct or delete the values that significantly deviate from the normal range in the data;

[0091] S1.3. Normalize the original data with different dimensions and value ranges so that it is mapped to a specific interval;

[0092] S1.4. Divide the preprocessed time series data into a training set and a test set. The training set is used for training the model, and at the same time, the performance of the model is evaluated during the training process to adjust the hyperparameters of the model. The test set is used to finally evaluate the generalization ability and prediction accuracy of the model.

[0093] The S2 includes:

[0094] The data prediction model includes a convolutional neural network and a long short-term memory network based on the attention mechanism. The convolutional neural network based on the attention mechanism is used to extract the original features in the training data. The convolutional neural network based on the attention mechanism includes an attention module, a convolutional unit, and a pooling unit. The input sequence of the branch of the attention module covers the input of the convolutional neural network, and a deep convolutional network and a pooling layer are stacked, so that the input receptive field corresponding to the feature is enlarged, and the model obtains more comprehensive context information, thereby learning the importance degree of the current local sequence feature.

[0095] The objective function of the data prediction model includes:

[0096]

[0097] In the above formula, is the model prediction value, and the model prediction value is the prediction set of the substation at a future moment. CNN is a convolutional neural network, LSTM is a long short-term memory recurrent neural network, Attn represents the attention mechanism, and y t is the true device status index, and the true device status index is extracted from the sample dataset. λ∈[0.01,0.1] represents the L2 regularization coefficient, and the L2 regularization coefficient is used to prevent overfitting. θ represents all trainable parameters of the model. All trainable parameters of the model include the convolutional kernel size, the number of convolutional kernels, the number of LSTM layers, and the model learning rate of the CNN. x t-n:t represents the real-time data of the substation input from time t - n to time t.

[0098] The output feature of the convolutional neural network based on the attention mechanism is multiplied element by element with the saliency feature output by its corresponding attention mechanism module, and the importance degree of the feature is reflected by the level of the value, thereby completing the discrimination of the important features.

[0099] The long short-term memory network is used to control the transmission and update of information through neurons and gate structures, mine the long-term dependence relationship in the substation time-series data, and predict the prediction set of the substation at a future moment.

[0100] In the S2, the original input of the CNN module is the sample dataset. The CNN module includes a convolutional layer, a pooling layer, and a fully connected layer. The convolutional layer performs a convolution operation by sliding a convolutional kernel on the data to learn local feature patterns. The pooling layer is used to reduce the data dimension, reduce the amount of calculation, and at the same time retain important feature information. The fully connected layer is used to integrate the extracted features.

[0101] The convolutional layer of the CNN model uses a "convolution kernel" for local perception. The size of the convolution kernel defines the receptive field of the convolution operation, and the ReLU function is selected as the activation function, including:

[0102] (Convolutional layer) H (l) = f(W (l) * H (l-1) + b (l) );

[0103] (Activation function) ReLU(x) = max(0, x);

[0104] In the above formula, H (l) is the output feature of the l-th layer convolution, l is the depth of the convolutional neural network, f() represents the activation function, W (l) is the weight matrix of the l-th layer convolution kernel, * is the convolution operation, H (l-1) is the output feature of the (l - 1)-th layer convolution, b (l) represents the bias term, and x is the input vector from the previous layer of the neural network;

[0105] Then, the local features extracted by the CNN are input into the LSTM module of the fusion attention mechanism. The LSTM module controls the transmission and update of information through neurons and gate structures, mines the long-term dependencies in the substation time series data, and captures the dynamic change laws of the data in the time series; in the hidden layer of the LSTM model, through the attention mechanism, the attention weights are calculated on the hidden state processed by the LSTM to highlight the key parts of the features and learn the global dynamic dependencies of the time series data;

[0106] Select the Adam optimization algorithm to train the model. Setting appropriate hyperparameters such as the learning rate and decay rate can enable the model to converge quickly;

[0107] Use the training set data to train the CNN-LSTM-attention time series analysis model. Through the collaborative work of the CNN network, LSTM network, and attention mechanism, learn the patterns and laws in the data, verify the performance of the model, and continuously adjust the parameters of the model to minimize the loss function;

[0108] The LSTM network mentioned above includes four parts: a forget gate, an input gate, an output gate, and an update module, which cooperate to store, update, and output information. The input gate determines how much of the input information at the current time step should be remembered. The forget gate determines how much of the memory state at the previous time step should be forgotten. The output gate determines how much of the memory state at the current time step should be output to the next layer or the output layer. The memory unit is responsible for storing and transmitting context information;

[0109]

[0110] In the above formula, f t is the output of the forget gate, i t is the output of the input gate, is the candidate memory state, C t is the updated memory state, o t is the output gate, h t is the hidden state output, σ is the activation function, W f is the forget gate, W i is the input gate, W C is the candidate memory state, W o is the weight term of the output gate, b f is the forget gate, b i is the input gate, b C is the candidate memory state, b o is the bias term of the output gate, is the Hadamard product, and the Hadamard product represents the multiplication of vector elements, h t-1 is the hidden state at time t-1, x t is the input at time t, and the original input of the t x is the feature vector output by the CNN;

[0111] The attention mechanism is used to dynamically assign weights according to the importance degree of different parts of the input data to the output. The expression of the attention mechanism is:

[0112] Improved formula:

[0113] Feature fusion:

[0114] In the above formula, α ij is the attention weight, which reflects the correlation strength of the feature at the i-th moment to the j-th moment. The attention weight is calculated through the normalized exponential function, e ij is the attention score, N is the total number of categories, e ik is the attention score of the K-th component, Q (t) is the query vector, and the query vector is obtained from the LSTM hidden state and is the query vector at the i-th moment, d k is the key vector dimension, K (t) is the key vector, is the key vector at the j-th moment, and the key vector dimension d k =64, W q and W k are the weight vectors, is the state of the LSTM at time t, The feature of the CNN at time t, F fusion where LayerNorm represents the layer normalization operation, α is the attention weight, represents the channel concatenation operation, and ⊙ represents the exclusive NOR operation.

[0115] In S3, the comprehensive scoring function of the multi-index scoring model includes:

[0116]

[0117] In the above formula, Score(t) is the substation renovation timing score of the device at time t, and the substation renovation timing score represents the possibility that the substation needs to be renovated. φ represents the index function, φ = [Aging, Health, F, M, E, L] T , ω represents the weight coefficient, ε represents the scoring threshold. The threshold θ ∈ [0.6, 0.8] is set as the renovation trigger boundary, and the sigmoid function is introduced to map the score to the interval (0, 1).

[0118] In S3, the device operation age index includes:

[0119] Aging(t) = t / T max ;

[0120] In the above formula, Aging(t) is the operation age of the device at time t, t is the current time, and T max is the design life of the device;

[0121] The health status index includes:

[0122]

[0123] In the above formula, Health(t) represents the health status evaluation of the device at time t, w a , w s , w f all represent dynamic weights, and the weight distribution satisfies w a + w s + w f = 1, e -λt is the aging attenuation term. The aging attenuation term is used to strengthen the non-linear degradation effect of the device service time through the exponential function, which is set according to the device situation. λ ∈ [0.01, 0.1] is the aging attenuation coefficient, and the aging attenuation coefficient is used to control the attenuation speed. S(t) is the evaluation state of the device, which is obtained by weighting each index in the prediction set of the substation at future times. ∑faults(t) is the weighted sum of the number of device failures in the historical time period;

[0124] The historical defect and fault index includes:

[0125]

[0126] In the above formula, F(t) is the historical defect fault index of the device at time t, and F hist (t - 1) is the historical defect fault index at the previous evaluation, and S(t - 1) is the device state at the previous evaluation; g1(·) is a function considering the historical defect fault and the device state, used to calculate the defect fault index at the current moment; w1 and w2 are the weight parameters of the historical defect fault index, used to adjust the importance of the historical defect fault and the device state in the current defect fault index;

[0127] The historical maintenance situation index includes:

[0128]

[0129] In the above formula, M(t) is the historical maintenance situation index of the device at time t, and g2(·) is a function considering the historical maintenance situation and the device maintenance situation, used to calculate the maintenance situation index at the current moment; M hist (t - 1) is the historical maintenance situation index at the previous evaluation, and R(t - 1) is the device maintenance situation at the previous evaluation; w7 and w8 are the weight parameters of the historical maintenance situation index, used to adjust the importance of the historical maintenance situation and the device maintenance situation in the current maintenance situation index;

[0130] The environmental factor index includes:

[0131] E(t) = w9 * T(t) + w 10 * H(t) + w 11 * P(t);

[0132] Among them, E(t) is the environmental factor index of the device at time t, T(t) is the temperature of the environment where the device is located, which can be the data collected by the temperature sensor; H(t) is the humidity of the environment where the device is located; P(t) is the air pressure or other influencing factors of the environment where the device is located; w9, w 10 and w 11 are the weight coefficients corresponding to the environmental factor index.

[0133] The device usage load index includes:

[0134] L(t) = w 12 * I(t) + w 13 * U(t) + w 14 * R(t);

[0135] In the above formula, L(t) is the device usage load index of the device at time t, I(t) is the load current of the device at time t, representing the amount of current carried by the device; U(t) is the operating voltage of the device; R(t) is the cumulative operating duration of the device, representing the total time the device has been operating; w 12 、w 13 and w 14 are the weight coefficients of the device usage load index.

[0136] In step S3, when 0.80 < Score(t) ≤ 1.00, there are significant problems in the aspects of aging and fault history of this type of device, with high renovation requirements, and renovation measures need to be taken immediately; when 0.60 < Score(t) ≤ 0.80, this type of device shows medium-level aging or faults and renovation needs to be considered preferentially; when 0.40 < Score(t) ≤ 0.60, this type of device shows slight aging and its operating condition is relatively good, and renovation is not required temporarily; when 0.00 < Score(t) ≤ 0.40, it operates stably and renovation is not required.

[0137] Embodiment 2:

[0138] A substation evaluation system integrating an attention mechanism neural network model, the system is used to execute the substation evaluation method integrating the attention mechanism neural network model as described in Embodiment 1, and specifically includes: a data collection module, a data prediction module, and an evaluation module;

[0139] The data collection module is used to collect data related to the operation of the substation from the monitoring system and the device management system of the substation, organize the data related to the operation of the substation in time series to form multi-dimensional time series data, and preprocess the multi-dimensional time series data to obtain a sample data set;

[0140] The data prediction module is used to construct a data prediction model, the prediction target of the data prediction model is the prediction set of the substation data at future times, train the data prediction model through the sample data set, input the real-time data of the substation into the trained data prediction model, solve the data prediction model, and obtain the prediction set of the substation at future times;

[0141] The evaluation module is used to construct a multi-index scoring model with the substation renovation time series score as the objective function, input the prediction set of the substation at future times into the trained multi-index scoring model, obtain the substation renovation time series score, and predict the renovation requirement degree of the device according to the substation renovation time series score and in combination with the scoring interval.

[0142] Embodiment 3:

[0143] A substation evaluation device integrating an attention mechanism neural network model, comprising a memory and a processor, wherein the memory is used for storing computer program codes and transmitting the computer program codes to the processor;

[0144] The processor is used for executing the substation evaluation method of the integrated attention mechanism neural network model as described in Embodiment 1 according to the instructions in the computer program codes.

[0145] A computer program product, comprising a computer program, and the computer program is executed by a processor to perform the substation evaluation method of the integrated attention mechanism neural network model as described in Embodiment 1.

[0146] The above is only the preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiment. Any equivalent modification or change made by those of ordinary skill in the art according to the disclosed content of the present invention shall be included in the protection scope recorded in the claims.

Claims

1. A substation evaluation method integrating an attention mechanism neural network model, characterized in that: The evaluation method includes: S1. Collect data related to substation operation from the substation monitoring system and equipment management system, organize the data related to substation operation in time series to form multi-dimensional time series data, and preprocess the multi-dimensional time series data to obtain a sample data set; S2. Construct a data prediction model. The prediction target of the data prediction model is the prediction set of substation future moment data. Train the data prediction model through the sample data set, input the substation real-time data into the trained data prediction model, solve the data prediction model, and obtain the prediction set of substation future moment; S3. Construct a multi-index scoring model with the substation transformation time sequence score as the objective function, input the prediction set of substation future moment into the trained multi-index scoring model to obtain the substation transformation time sequence score, and predict the equipment transformation demand degree according to the substation transformation time sequence score and in combination with the scoring interval.

2. The substation evaluation method integrating an attention mechanism neural network model according to claim 1, characterized in that: In S1, the data related to substation operation includes health status indicators, defect and fault indicators, maintenance records, environmental indicators, equipment usage load, time features, and other related features; The health status indicators include equipment operation status, equipment insulation status, and equipment aging degree. The defect and fault indicators include historical fault records, fault warning signals, and characteristic changes before the occurrence of faults. The maintenance records include maintenance time points, maintenance contents, and equipment status changes after maintenance. The environmental indicators include environmental parameters, weather conditions, and geographical locations. The equipment usage load includes electrical parameters, load rates, and load change trends. The time features include time stamps, seasonal features, and holidays or special events. The other related features include equipment types, equipment manufacturers and models, and historical operation data.

3. The substation evaluation method integrating an attention mechanism neural network model according to claim 1, characterized in that: S2 includes: The data prediction model includes a convolutional neural network based on an attention mechanism and a long short-term memory network. The convolutional neural network based on an attention mechanism is used to extract the original features in the training data. The convolutional neural network based on an attention mechanism includes an attention module, a convolutional unit, and a pooling unit. The input sequence of the branch of the attention module covers the input of the convolutional neural network, and deep convolutional networks and pooling layers are stacked, so that the input receptive field corresponding to the features is expanded, and the model obtains more comprehensive context information, thereby learning the importance degree of the current local sequence features; The objective function of the data prediction model includes: In the above formula, is the model prediction value, and the model prediction value is the prediction set of the substation at a future moment. CNN is a convolutional neural network, LSTM is a long short-term memory recurrent neural network, Attn represents the attention mechanism, and y t is the true device status index, which is extracted from the sample dataset. λ∈[0.01,0.1] represents the L2 regularization coefficient, which is used to prevent overfitting. θ represents all trainable parameters of the model. All trainable parameters of the model include the convolutional kernel size, the number of convolutional kernels, the number of LSTM layers, and the model learning rate of the CNN. x t-n:t represents the real-time data of the substation input from time t-n to time t; The output features of the convolutional neural network based on the attention mechanism are multiplied element by element with the significant features output by the corresponding attention mechanism module, and the importance degree of the features is reflected by the high and low values, so as to complete the discrimination of important features. The long short-term memory network is used to control the transmission and update of information through neurons and gate structures, mine the long-term dependencies in the substation time-series data, and predict the prediction set at future moments of the substation.

4. The substation evaluation method of a neural network model integrating an attention mechanism according to claim 3, wherein: In S2, the original input of the CNN module is a sample data set. The CNN module includes a convolutional layer, a pooling layer, and a fully connected layer. The convolutional layer performs a convolutional operation by sliding a convolutional kernel over the data to learn local feature patterns; the pooling layer is used to reduce the data dimension, reduce the computational amount, and at the same time retain important feature information; the fully connected layer is used to integrate the extracted features; The convolutional layer of the CNN model uses a "convolutional kernel" for local perception. The size of the convolutional kernel defines the receptive field of the convolutional operation, and the ReLU function is selected as the activation function, including: (Convolutional layer) H (l) = f(W (l) * H (l-1) + b (l) ); (Activation function) ReLU(x) = max(0, x); In the above formula, H (l) is the output feature of the l-th layer convolution, l is the depth of the convolutional neural network, f() represents the activation function, and W (l) is the weight matrix of the convolutional kernel of the l-th layer, * is the convolution operation, and H (l-1) is the output feature of the (l-1)-th layer convolution, b (l) represents the bias term, and x is the input vector from the previous layer of the neural network; Then, the local features extracted by the CNN are input into the LSTM module of the fusion attention mechanism. The LSTM module controls the transmission and update of information through neurons and gate structures, mines the long-term dependencies in the substation time-series data, and captures the dynamic change rules of the data in the time series; in the hidden layer of the LSTM model, through the attention mechanism, the attention weights are calculated on the hidden state processed by the LSTM to highlight the key parts of the features and learn the global dynamic dependencies of the time-series data; The Adam optimization algorithm is selected to train the model, and appropriate hyperparameters such as the learning rate and decay rate are set, which can enable the model to converge quickly; The training set data is used to train the CNN-LSTM-attention time series analysis model. Through the collaborative work of the CNN network, the LSTM network, and the attention mechanism, the patterns and rules in the data are learned, and the performance of the model is verified. The parameters of the model are continuously adjusted to minimize the loss function; The LSTM network includes four parts: a forget gate, an input gate, an output gate, and an update module, which cooperate to store, update, and output information. The input gate determines how much of the input information at the current time step is to be remembered. The forget gate determines how much of the memory state at the previous time step is to be forgotten. The output gate determines how much of the memory state at the current time step is to be output to the next layer or the output layer. The memory unit is responsible for storing and transmitting context information; f t = σ(W f · [h t-1 , x t + b f ) i t = σ(W i · [h t-1 , x t + b i ) o t = σ(W o · [h t-1 , x t + b o ) In the above formula, f t is the output of the forget gate, i t is the output of the input gate, is the candidate memory state, C t is the updated memory state, o t is the output gate, h t is the hidden state output, σ is the activation function, W f is the forget gate, W i is the input gate, W C is the candidate memory state, W o is the weight term of the output gate, b f is the forget gate, b i is the input gate, b C is the candidate memory state, b o is the bias term of the output gate, is the Hadamard product, and the Hadamard product represents the multiplication of vector elements, h t-1 is the hidden state at time t - 1, x t is the input at time t, and the original input of the t is the feature vector output by the CNN;​ The attention mechanism is used to dynamically allocate weights according to the importance degree of different parts of the input data to the output. The expression of the attention mechanism is: Improved formula: Feature fusion: In the above formula, α ij is the attention weight, reflecting the association strength of the feature at the i-th moment with respect to the j-th moment. The attention weight is calculated through the softmax function, e ij is the attention score, N is the total number of categories, e ik is the attention score of the K-th component, Q (t) is the query vector, and the query vector is obtained from the LSTM hidden state . is the query vector at the i-th moment, d k is the key vector dimension, K (t) is the key vector, is the key vector at the j-th moment, and the key vector dimension d k = 64, W q and W k are weight vectors, is the state of the LSTM at time t, is the feature of the CNN at time t, F fusion is, LayerNorm represents the layer normalization operation, α is the attention weight, represents the channel concatenation operation, and ⊙ represents the exclusive NOR operation.

5. The substation evaluation method of a neural network model integrating an attention mechanism according to claim 1, wherein: In S3, the comprehensive scoring function of the multi-index scoring model includes: In the above formula, Score(t) is the substation renovation timing score of the device at time t, and the substation renovation timing score represents the possibility that the substation needs to be renovated. φ represents the index function, and φ = [Aging, Health, F, M, E, L] T , ω represents the weight coefficient, ε represents the score threshold. The threshold θ ∈ [0.6, 0.8] is set as the renovation trigger boundary, and the sigmoid function is introduced to map the score to the interval (0, 1).

6. The substation evaluation method of a neural network model integrating an attention mechanism according to claim 4, wherein: In S3, the equipment commissioning years index includes: Aging(t)=t / T max ; In the above formula, Aging(t) is the operation years of the device at time t, t is the current time, and T max is the design life of the device; The health status index includes: In the above formula, Health(t) represents the health status evaluation of the device at time t, w a , w s , w f all represent dynamic weights, and the weight allocation satisfies w a + w s + w f = 1, e -λt is the aging attenuation term. The aging attenuation term is used to strengthen the non-linear degradation effect of the device's service time through an exponential function, and is set according to the device situation. λ ∈ [0.01, 0.1] is the aging attenuation coefficient, and the aging attenuation coefficient is used to control the attenuation speed. S(t) is the evaluation status of the device, which is obtained by weighting each index in the prediction set of the substation at future times. ∑faults(t) is the weighted sum of the number of device failures occurring during the historical time period; The historical defect and fault index includes: F(t) = g1(F hist (t - 1), S(t - 1)) = w1 * F hist (t - 1)+ w2 * S(t - 1); In the above formula, F(t) is the historical defect failure index of the device at time t, F hist (t - 1) is the historical defect failure index at the previous evaluation, and S(t - 1) is the device state at the previous evaluation; g1(·) is a function considering the historical defect failure and the device state, which is used to calculate the defect failure index at the current moment; w1 and w2 are the weight parameters of the historical defect failure index, which are used to adjust the importance of the historical defect failure and the device state in the current defect failure index; The historical maintenance situation indicators include: M(t) = g2(M hist (t - 1), R(t - 1)) = w7 * M hist (t - 1)+ w8 * R(t - 1); In the above formula, M(t) is the historical maintenance situation index of the device at time t, g2(·) is a function considering the historical maintenance situation and the device maintenance situation, and is used to calculate the maintenance situation index at the current moment; M hist (t - 1) is the historical maintenance situation index at the previous evaluation, R(t - 1) is the device maintenance situation at the previous evaluation; w7 and w8 are the weight parameters of the historical maintenance situation index, and are used to adjust the importance of the historical maintenance situation and the device maintenance situation in the current maintenance situation index; The environmental factor indicators include: E(t) = w9 * T(t) + w 10 * H(t) + w 11 * P(t); Among them, E(t) is the environmental factor index of the device at time t, T(t) is the temperature of the environment where the device is located, which can be collected by the data of the temperature sensor; H(t) is the humidity of the environment where the device is located; P(t) is the air pressure or other influencing factors of the environment where the device is located; w9, w 10 and w 11 are the weight coefficients corresponding to the environmental factor indexes. The equipment usage load indicators include: L(t) = w 12 *I(t) + w 13 *U(t) + w 14 *R(t); In the above formula, L(t) is the device usage load index of the device at time t, I(t) is the load current of the device at time t, representing the amount of current carried by the device; U(t) is the operating voltage of the device; R(t) is the cumulative operating duration of the device, representing the total time the device has been operating; w 12 、w 13 and w 14 are the weight coefficients of the device usage load index.

7. The substation evaluation method of a fusion attention mechanism neural network model according to claim 1, characterized in that: In S3, when 0.80 < Score(t) ≤ 1.00, there are significant problems in terms of the history of aging and faults in this type of equipment, with a high transformation requirement, and transformation measures need to be taken immediately; when 0.60 < Score(t) ≤ 0.80, this type of equipment shows medium-level aging or faults and transformation needs to be considered preferentially; when 0.40 < Score(t) ≤ 0.60, this type of equipment shows slight aging and the operating condition is relatively good, and transformation is not required temporarily; when 0.00 < Score(t) ≤ 0.40, it operates stably and no transformation is needed.

8. A substation evaluation system integrating an attention mechanism neural network model, characterized in that, The system is used to execute the substation evaluation method of the fusion attention mechanism neural network model according to any one of claims 1 to 7, specifically including: a data collection module, a data prediction module, and an evaluation module; The data collection module is used to collect data related to the operation of the substation from the monitoring system and the equipment management system of the substation, organize the data related to the operation of the substation in time series to form multi-dimensional time series data, and preprocess the multi-dimensional time series data to obtain a sample data set; The data prediction module is used to construct a data prediction model. The prediction target of the data prediction model is the prediction set of the substation data at future moments. The data prediction model is trained through the sample data set, and the real-time data of the substation is input into the trained data prediction model to solve the data prediction model to obtain the prediction set of the substation at future moments; The evaluation module is used to construct a multi-index scoring model with the substation transformation time sequence score as the objective function, input the prediction set of the substation at future moments into the trained multi-index scoring model to obtain the substation transformation time sequence score, and predict the degree of transformation requirement of the equipment according to the substation transformation time sequence score and in combination with the scoring interval.

9. A substation evaluation device integrating an attention mechanism neural network model, characterized in that, It includes a memory and a processor. The memory is used to store computer program code and transmit the computer program code to the processor; The processor is used to execute the substation evaluation method of the fusion attention mechanism neural network model according to any one of claims 1 to 7 according to the instructions in the computer program code.

10. A computer program product, comprising a computer program, characterized in that, The computer program is executed by the processor to perform the substation evaluation method of the fusion attention mechanism neural network model according to any one of claims 1 to 7.

Images