Knowledge distillation-based voltage measurement reliability evaluation method in complex environment

Abstract

The invention discloses a knowledge distillation-based voltage metering reliability evaluation method in a complex environment. The method comprises the following steps of collecting related data of voltage metering; preprocessing the data; constructing a deep neural network model for voltage reliability evaluation; performing model compression through knowledge distillation; training the deep neural network by using the training set data, and optimizing the network weight; testing the model by using the test set data; and judging whether the target voltage is reliable or not through threshold comparison. According to the invention, a plurality of units such as the voltage transformer, the electric energy meter, the secondary circuit and the like are incorporated into the evaluation model, so that the reliability of voltage metering is evaluated more comprehensively; deep neural network modeling is adopted, the model is compressed through a knowledge distillation method, the calculation efficiency and precision of the system are effectively improved, and the evaluation model can adapt to an industrial personal computer platform with limited resources; the method can adapt to various environment changes, provides more accurate and stable voltage metering evaluation, and remarkably improves the reliability of a voltage metering system.

Description

Technical Field

[0001] This application relates to the technical field of power data acquisition, and particularly to a method for evaluating the reliability of voltage measurement in a complex environment based on knowledge distillation. Background Art

[0002] The evaluation of the reliability of voltage measurement is a process of evaluating the ability of voltage measurement devices or systems to maintain accuracy and stability during long-term use. As the core component of modern power systems, the operation of smart grids relies on accurate and real-time power data acquisition and analysis. Voltage measurement is not only the basis for monitoring the operation status of the power grid, but also the key basis for multiple functions such as fault diagnosis, load management, and power quality optimization. With the gradual construction and modernization of smart grids, the reliability of voltage measurement systems directly affects the safety and stability of power systems and the efficient utilization of electric energy; in order to cope with problems such as voltage fluctuations and voltage dips that may occur during power grid operation, it is particularly important to evaluate the accuracy, timeliness, and reliability of voltage measurement. By introducing high-precision measurement devices, advanced monitoring technologies, and data analysis methods, the reliability of voltage measurement can be improved, potential problems in the power system can be detected in a timely manner, thereby ensuring the efficient operation of the power grid and the stability of power supply.

[0003] Existing voltage measurement methods mainly include traditional analog measurement methods and digital voltage measurement methods. Traditional analog measurement methods mainly rely on mechanical instruments (such as voltmeters and ammeters) to directly measure voltage values through electrical principles. Their accuracy and response speed are relatively poor, and they are easily affected by external environmental factors, resulting in low reliability of measurement results; digital voltage measurement methods use digital signal processing technology to convert voltage signals into digital signals and process and store them through computer systems. This method has high measurement accuracy, fast response, and strong anti-interference ability, but it has high requirements for hardware devices and may be affected by factors such as system delay and data transmission errors. Therefore, most existing methods ignore the influence of environmental factors (such as atmospheric humidity, altitude, temperature, etc.) on voltage accuracy and cannot comprehensively handle the interaction of multiple links such as voltage transformers, energy meters, and secondary circuits; in addition, existing technologies often lack effective modeling and optimization means when dealing with high-dimensional and multi-factor complex data, especially difficult to achieve real-time and efficient calculations on resource-constrained industrial computer platforms. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for evaluating the reliability of voltage measurement in a complex environment based on knowledge distillation. This method comprehensively considers the influence of multiple environmental factors through a deep neural network model, and combines knowledge distillation technology to compress the model. While ensuring high-precision evaluation, it adapts to complex working conditions and improves calculation efficiency, thus solving the problem of reliability evaluation in the prior art under environmental changes and computational resource limitations.

[0005] To solve the above technical problems, the present invention is implemented as follows:

[0006] A method for evaluating the reliability of voltage measurement in a complex environment based on knowledge distillation, including the following steps:

[0007] S1. Collect relevant data of voltage measurement;

[0008] S2. Data preprocessing;

[0009] S3. Construct a deep neural network model for voltage reliability evaluation;

[0010] S4. Compress the model through knowledge distillation;

[0011] S5. Use the training set data to train the deep neural network and optimize the network weights;

[0012] S6. Use the test set data to test the model;

[0013] S7. Judge whether the target voltage is reliable through threshold comparison.

[0014] Further, the specific method of step S1 is as follows:

[0015] Collect the original data from three core units of the voltage measurement system. The core units include voltage transformers, energy meters, and secondary circuits. Organize the original data into a data set X, and its expression is as follows:

[0016] X = [D1, D2, D3, D4, D5, D6, D7, D8, D9, D 10

[0017] Among them, D1, D2, D3, D4, and D5 respectively represent the atmospheric humidity, altitude, atmospheric pollution degree, operation time, and maximum operating current of the voltage transformer; D6 and D7 respectively represent the operation time and maximum operating current of the energy meter, and D8, D9, and D 10 respectively represent the temperature, operation time, and maximum operating current of the secondary circuit;

[0018] The i-th sample in the data set X is X i , and the j-th feature is D j .​

[0019] Further, the specific method of step S2 is as follows:

[0020] If there are missing values in the dataset X, the missing values are processed by an imputation method to ensure that each feature is used during training. The features in the dataset X are standardized so that their values are within the same range. The standardization process is achieved through the following expression:

[0021]

[0022] where μ j and σ j represent the mean and standard deviation of the j-th feature respectively, represents the value after standardization;

[0023] The standardized dataset is divided into a training set and a test set, where the proportion of the training set is 70% and the proportion of the test set is 30%.

[0024] Further, the specific method of step S3 is as follows:

[0025] By constructing a deep neural network (DNN) model, a multi-layer neural network is used to perform a non-linear mapping on the input feature data. The neural network includes an input layer, multiple hidden layers, and an output layer;

[0026] The input layer is 10 features that have been standardized;

[0027] The hidden layer contains k fully connected layers, and the output expression of each layer is as follows:

[0028] h l = σ(W l ·h l-1 + b l )

[0029] where W l represents the weight matrix of the l-th layer, b l represents the bias term, and σ represents the activation function;

[0030] The output layer calculates the reliability score of each sample through the softmax activation function. The specific expression is as follows:

[0031]

[0032] where p(X i ) represents the reliability score of the sample, h k (X i ) represents the output of the sample X i in the network, h k (Xj ) represents the output of sample X j in the network.

[0033] Furthermore, the specific method of step S4 is as follows:

[0034] During the training process, the knowledge learned by the teacher model is transferred to the student model through the knowledge distillation technique. The teacher model is obtained by a deep neural network using the above dataset and training method, and the probability distribution output by the softmax activation function provides soft labels. The student model learns these soft labels to imitate the decision-making process of the teacher model;

[0035] The loss function of the student model includes the knowledge loss from the teacher model and the cross-entropy loss of the standard labels. The specific expression of its loss function is as follows:

[0036]

[0037] where λ represents the weight parameter, D KL represents the Kullback-Leibler divergence, which measures the difference between two probability distributions, P teacher (X) and P student (X) represent the output probability distributions of the teacher and student models respectively, L CE is the cross-entropy loss function, which represents the prediction ability of the student model for the true labels, and y and represent the true label of the sample and the predicted label of the student model respectively.

[0038] Furthermore, the specific method of step S5 is as follows:

[0039] Update and optimize the network weight parameters by the gradient descent method. The specific update and optimization expression is as follows:

[0040]

[0041] where θ represents the parameter to be optimized in the network, η represents the learning rate, represents the gradient of the loss function with respect to the parameter.

[0042] Furthermore, the specific method of step S6 is as follows:

[0043] After the training is completed, input the test set into the updated and optimized model, calculate the output probability p of the model, and compare it with the true label y. The expression for calculating the accuracy of the model is as follows:

[0044]

[0045] where, represents the indicator function, y i and respectively represent the true label of the i-th sample and the prediction result of the model.

[0046] Furthermore, the specific method of step S7 is as follows:

[0047] For the model to output a probability value p(X i ), by setting a threshold, its conversion into a class label expression is as follows:

[0048]

[0049] The final output indicates whether the voltage is within the standard value range, directly evaluates whether the voltage of the model system meets the design requirements, and reflects the operating conditions of the power grid; if the voltage deviates from the standard value range, it means that there is an error or failure in the metering device, or there is an abnormal fluctuation in the power grid.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] By comprehensively considering multiple units such as voltage transformers, energy meters, secondary circuits, and their environmental factors and incorporating them into the evaluation model, the present invention can more comprehensively evaluate the reliability of voltage metering; by using a deep neural network (Network-T) for modeling and compressing the model through a knowledge distillation method (Network-S), the computing efficiency and accuracy of the system can be effectively improved, enabling the evaluation model to adapt to the resource-limited industrial control computer computing platform and having a stronger application prospect; by training a deep neural network and combining distillation technology, it can adapt to various environmental changes and provide more accurate and stable voltage metering evaluation, operate stably in a complex power environment and continuously optimize, and significantly improve the reliability of the voltage metering system. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a schematic flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0053] The following further elaborates in detail the specific embodiments of the present invention in conjunction with the drawings and specific examples.

[0054] The present invention aims to evaluate the reliability of a voltage metering system in a complex environment, comprehensively considers the influence of multiple factors (such as atmospheric humidity, altitude, energy meter operation time, temperature, etc.) on the accuracy of voltage metering, ensures that the voltage metering can continuously and accurately reflect the actual situation of the power grid, and provides a basis for system maintenance and optimization. The present invention uses a deep neural network (DNN) for modeling and optimizes the model through knowledge distillation to ensure efficient operation on a resource-constrained industrial control computer.

[0055] Such as Figure 1As shown in the figure, a method for evaluating the reliability of voltage measurement in a complex environment based on knowledge distillation includes the following steps:

[0056] S1. Collect relevant data on voltage measurement, and the specific method is as follows:

[0057] Collect raw data from three core units of the voltage measurement system. The core units include voltage transformers, energy meters, and secondary circuits. Organize the raw data into a data set X, and its expression is as follows:

[0058] X = [D1, D2, D3, D4, D5, D6, D7, D8, D9, D 10

[0059] Among them, D1, D2, D3, D4, and D5 respectively represent the atmospheric humidity, altitude, atmospheric pollution degree, operating time, and maximum operating current of the voltage transformer; D6 and D7 respectively represent the operating time and maximum operating current of the energy meter, and D8, D9, and D 10 respectively represent the temperature, operating time, and maximum operating current of the secondary circuit;

[0060] The i-th sample in the data set X is X i , and the j-th feature is D j .

[0061] S2. Data preprocessing, and the specific method is as follows:

[0062] If there are missing values in the data set X, perform missing value processing through the imputation method to ensure that each feature is used during training. Since the dimensions of various influencing factors are different, it is necessary to standardize each feature so that its value is within the same range (such as zero mean and unit variance). Standardize each feature in the data set X so that its value is within the same range, and the standardization process is achieved through the following expression:

[0063]

[0064] Among them, μ j and σ j respectively represent the mean and standard deviation of the j-th feature, represents the value after standardization;

[0065] Divide the standardized data set into a training set and a test set, where the proportion of the training set is 70% and the proportion of the test set is 30%.

[0066] S3. Construct a deep neural network model for voltage reliability evaluation, and the specific method is as follows:

[0067] ​By constructing a deep neural network (DNN) model, a multi-layer neural network is used to perform non-linear mapping on the input feature data. The neural network includes an input layer, multiple hidden layers, and an output layer;

[0068] The input layer consists of 10 features that have been standardized;

[0069] The hidden layer contains k fully connected layers, and the output expression of each layer is as follows:

[0070] h l =σ(W l ·h l-1 +b l )

[0071] where W l represents the weight matrix of the l-th layer, b l represents the bias term, and σ represents the activation function;

[0072] The output layer calculates the reliability score of each sample through the softmax activation function. The specific expression is as follows:

[0073]

[0074] where p(X i ) represents the reliability score of the sample, h k (X i ) represents the output of the sample X i in the network, and h k (X j ) represents the output of the sample X j in the network.

[0075] S4. Model compression is performed through knowledge distillation. The specific method is as follows:

[0076] During the training process, the knowledge learned by the teacher model (Network-T) is transferred to the student model (Network-S) through knowledge distillation technology. The teacher model is obtained through a deep neural network and using the above dataset and training method. The student model is a smaller neural network that uses the same input features but has fewer layers and neurons. Through knowledge distillation, the student model can operate efficiently in an environment with limited computing resources. The probability distribution output by the softmax activation function provides soft labels, and the student model learns these soft labels to imitate the decision-making process of the teacher model;

[0077] The loss function of the student model includes the knowledge loss from the teacher model and the cross-entropy loss of the standard labels. The specific expression of its loss function is as follows:

[0078]

[0079] Among them, λ represents the weight parameter, and D KL represents the Kullback-Leibler divergence, which measures the difference between two probability distributions. P teacher (X) and P student (X) represent the output probability distributions of the teacher and student models respectively. L CE is the cross-entropy loss function, which represents the prediction ability of the student model for the true labels. y and represent the true label of the sample and the predicted label of the student model respectively. Through the distillation process, the student model can not only learn the standard labels but also absorb the rich inter-class relationship information provided by the teacher model, thereby improving its performance.

[0080] S5. Use the training set data to train the deep neural network and optimize the network weights. The specific method is as follows:

[0081] Update and optimize the network weight parameters through the gradient descent method. The specific update and optimization expression is as follows:

[0082]

[0083] Among them, θ represents the parameter to be optimized in the network, η represents the learning rate, represents the gradient of the loss function with respect to the parameter.

[0084] S6. Use the test set data to test the model. The specific method is as follows:

[0085] After training is completed, input the test set into the updated and optimized model, calculate the output probability p of the model, and compare it with the true label y. The accuracy expression of the model is calculated as follows:

[0086]

[0087] Among them, represents the indicator function, y i and represent the true label of the i-th sample and the prediction result of the model respectively.

[0088] S7. Determine whether the target voltage is reliable through threshold comparison. The specific method is as follows:

[0089] For the model to output a probability value p(X i ), by setting a threshold, it is converted into a class label as follows:

[0090]

[0091] Final output Indicates whether the voltage is within the standard value range, directly evaluates whether the voltage of the model system meets the design requirements, and reflects the operating conditions of the power grid; if the voltage deviates from the standard value range, it means that there is an error or failure in the metering equipment, or there is an abnormal fluctuation in the power grid.

[0092] The above are only the implementation manners of the present invention. Once again, for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements can still be made to the present invention, and these improvements are also included in the protection scope of the claims of the present invention.

Claims

1. A method for evaluating the reliability of voltage measurement in complex environments based on knowledge distillation, characterized in that: The steps are as follows: S1. Collect relevant data for voltage measurement; S2. Data preprocessing; S3. Construct a deep neural network model for voltage reliability assessment; S4. Compress the model through knowledge distillation; S5. Use the training set data to train the deep neural network and optimize the network weights; S6. Use the test set data to test the model; S7. Determine whether the target voltage is reliable through threshold comparison.

2. The method for evaluating the reliability of voltage measurement in a complex environment based on knowledge distillation according to claim 1, wherein: The specific method of step S1 is as follows: Collect raw data from three core units of the voltage measurement system. The core units include voltage transformers, energy meters, and secondary circuits. Organize the raw data into a data set X, and its expression is as follows: X = [D1, D2, D3, D4, D5, D6, D7, D8, D9, D 10 ​ Among them, D1, D2, D3, D4, D5 respectively represent the atmospheric humidity, altitude, atmospheric pollution degree, operation time and maximum operating current of the voltage transformer; D6 and D7 respectively represent the operation time and maximum operating current of the electricity meter, and D8, D9, D 10 respectively represent the temperature, operation time and maximum operating current of the secondary circuit; The i-th sample in dataset X is X i , and the j-th feature is D j .

3. The method for evaluating the reliability of voltage measurement in a complex environment based on knowledge distillation according to claim 1, wherein: The specific method of step S2 is as follows: If there are missing values in the data set X, perform missing value processing through an imputation method to ensure that each feature is used during training. Standardize the features in the data set X so that their values are within the same range. Then the standardization process is achieved through the following expression: Among them, μ j and σ j respectively represent the mean and standard deviation of the j-th feature, represents the value after standardization; Divide the standardized data set into a training set and a test set, where the proportion of the training set is 70% and the proportion of the test set is 30%.

4. The method for evaluating the reliability of voltage measurement in a complex environment based on knowledge distillation according to claim 1, wherein: The specific method of step S3 is as follows: By constructing a deep neural network model, use a multi-layer neural network to perform non-linear mapping on the input feature data. The neural network includes an input layer, multiple hidden layers, and an output layer; The input layer is 10 features that have been standardized; The hidden layer contains k fully connected layers, and the output expression of each layer is as follows: h l = σ(W l ·h l-1 + b l ) Among them, W l represents the weight matrix of the l-th layer, and b l represents the bias term, and σ represents the activation function; The output layer calculates the reliability score of each sample through the softmax activation function, and the specific expression is as follows: Among them, p(X i ) represents the reliability score of the sample, and h k (X i ) represents the output of the sample X i in the network, and h k (X j ) represents the output of the sample X j in the network.

5. The method for evaluating the reliability of voltage measurement in a complex environment based on knowledge distillation according to claim 1, wherein: The specific method of step S4 is as follows: During training, transfer the knowledge learned by the teacher model to the student model through knowledge distillation technology. The teacher model is obtained through a deep neural network and the above data set and training method. Provide soft labels through the probability distribution output by the softmax activation function, and the student model learns these soft labels to imitate the decision-making process of the teacher model; The loss function of the student model includes the knowledge loss from the teacher model and the cross-entropy loss of the standard label, and its loss function specific expression is as follows: Among them, λ represents the weight parameter, D KL represents the Kullback-Leibler divergence, which measures the difference between two probability distributions, P teacher (X) and P student (X) represent the output probability distributions of the teacher and student models respectively, L CE is the cross-entropy loss function, which represents the prediction ability of the student model for the true labels, y and represent the true label of the sample and the predicted label of the student model respectively.

6. The method for evaluating the reliability of voltage measurement in a complex environment based on knowledge distillation according to claim 1, wherein: The specific method of step S5 is as follows: Update and optimize the network weight parameters through the gradient descent method, and the specific update and optimization expression is as follows: where θ represents the parameter to be optimized in the network, η represents the learning rate, represents the gradient of the loss function with respect to the parameter.

7. The method for evaluating the reliability of voltage measurement in a complex environment based on knowledge distillation according to claim 1, wherein: The specific method of step S6 is as follows: After the training is completed, the test set is input into the updated and optimized model, the output probability p of the model is calculated, and it is compared with the true label y. The accuracy expression of the model is calculated as follows: Among them, represents the indicator function, and y i and represent the true label of the i-th sample and the prediction result of the model, respectively.

8. The method for evaluating the reliability of voltage measurement in a complex environment based on knowledge distillation according to claim 1, characterized in that: The specific method of step S7 is as follows: For the model to output a probability value p(X i ), by setting a threshold, its conversion into a class label is expressed as follows: Final output Indicates whether the voltage is within the standard value range, directly evaluates whether the voltage of the model system meets the design requirements, and reflects the operating conditions of the power grid.

Images