CN107274067B - A risk assessment method for distribution transformer overload - Google Patents

Abstract

A distribution transformer overload risk assessment method, comprising collecting distribution transformer operating data, preprocessing the original data according to overload attributes, and obtaining a distribution transformer overload attribute set; the method uses neural network technology to analyze the distribution transformer overload time attributes Network modeling and simulation training are set to obtain the network structure of distribution transformer overload risk; according to the network structure, the degree of risk of distribution transformer outage in different operating states during overload is evaluated, and the probability of distribution transformer being out of operation due to overload is simulated; the greater the probability It means that the possibility of transformer return is greater. The distribution transformer overload risk assessment proposed by the present invention can accurately determine the risk degree of distribution transformer overload outage or damage, and can find distribution transformers with relatively high overload risks and take targeted measures when unexpected changes in electricity load occur. Provide evidence, and then contribute to the controllability of transformer overload operation, the accuracy rate can reach more than 87%.

Description

A method for overload risk assessment of distribution transformers

Technical Field

The invention relates to a method for assessing overload risk of a distribution transformer, and belongs to the technical field of transformer operation.

Background Art

Overloaded distribution transformers account for a certain proportion in the distribution network. In some areas, the overload ratio of distribution transformers reaches 15%. The load of distribution transformers changes all the time in different time periods. During peak power consumption periods, distribution transformers are prone to overload. The imbalance of power load requires that distribution transformers have a certain overload capacity to meet users' safe and reliable power consumption.

At present, there are few studies on mining overload implicit patterns from the massive operating data of distribution transformers. Assessing the risk of outage when the distribution transformer is overloaded is a prerequisite for fully utilizing the overload capacity of the distribution transformer and ensuring safe and reliable power use. If the distribution transformer is overloaded and out of operation, there is a strong correlation between the operating state and the operating mode of the distribution transformer when it is overloaded. Therefore, the correlation between the operating state and the operating mode of the distribution transformer when it is overloaded can be mined to help quantify the impact of different operating conditions on the operating mode of the distribution transformer.

Summary of the invention

The purpose of the present invention is to solve the problem that the risk degree of overload power outage of the distribution transformer cannot be determined when the distribution transformer is overloaded. The present invention proposes a distribution transformer overload risk assessment method.

The technical solution of the present invention is as follows: a method for assessing the overload risk of a distribution transformer, which collects operating data of the distribution transformer, processes the original data according to the overload attribute, and obtains an overload attribute set of the distribution transformer; then uses neural network technology to model and train the attribute set of the distribution transformer when it is overloaded, and obtains an overload risk network structure of the distribution transformer; the degree of risk of outage of the distribution transformer in different operating states when overloaded is assessed according to the network structure, and the probability of overload retirement of the distribution transformer is simulated; the greater the probability, the greater the possibility of transformer retirement.

The distribution transformer overload risk network structure is an artificial neural network. Artificial neurons are mathematical models that simulate biological neurons. The output of neurons is a function of weighted input, as shown in the following formula:

Among them, y is the output of the neuron; _xi is the input of the neuron; _wi is the weight corresponding to the input of the neuron;

BP neural network is a multi-layer perceptron structure, including input, output layer and several hidden layers, which is divided into forward propagation stage and backward propagation stage. In the forward propagation stage, information is transmitted from the input layer to the output layer through the hidden layer step by step, as shown in the following formula:

O _i =f _n (f _n-1 ((...f ₁ ([X _i ][W ₁ ])+[B ₁ ]...)[W _n-1 ])+[B _n-1 ] )

_Xi represents the i-th sample input, _W1 …Wn _-1 represents the weight matrix of the hidden layer, _B1 … _Bn-1 represents the bias matrix of the hidden layer, _f1 … _fn-1 represents the weight function of the hidden layer, _fn represents the output function, and _0i represents the expected value;

In the back propagation stage, the weights and biases of the network are repeatedly adjusted and trained. The training process requires the input vector X and the target value Y. The network training process is the process of minimizing the mean square error. Suppose the error accuracy E _i of the i-th sample is as shown in the following formula:

The error E of the entire m sample set is shown as follows:

The simulation training sets a network training error, such as 10 ^-2 , and determines the number of neurons in the input layer, hidden layer, and output layer through multiple tests, wherein the input layer is an overload attribute vector, and the output layer is whether the corresponding distribution transformer is shut down. If it is shut down, it is set to 1, and if it is not shut down, it is set to 0; the training algorithm adopts the BP algorithm and the activation function; 120 overload shutdown samples and 200 overload non-shutdown samples are randomly selected as training samples, and the weights and biases are continuously changed along the direction in which the error function decreases fastest until the training error is less than the specified value, that is, the training error is stopped, that is, the trained weight matrix and bias matrix are obtained, and then the network structure that can be used to predict the overload risk of the distribution transformer is obtained.

The collected distribution transformer operation data includes the following information contained in the substation operation records in the distribution transformer monitoring related system: capacity, voltage value, current value, collection time, active power, reactive power; the weather conditions when the distribution transformer is overloaded are obtained through the meteorological monitoring system to constitute the original distribution transformer operation data.

The raw data is preprocessed according to the overload attribute. The discrete operation data of the distribution transformer cannot be directly used as the neural network input. The operation data needs to be processed according to the overload attribute of the distribution transformer, and the noise data and empty data in the raw data need to be eliminated.

1) Distribution transformer overload shutdown criterion: if there are two consecutive load factors less than 0.01 during the overload of the distribution transformer, the distribution transformer is considered to be out of service;

2) The coefficient K1 before overload: the average value of the n load coefficient K collection points before overload is taken as the coefficient K1 before overload;

3) Overload coefficient K2: the average value of n load coefficient K collection points during overload is taken as the overload coefficient K2 value;

4) Overload coefficient K3: the average value of n load coefficient K collection points after overload is taken as the overload coefficient K3 value;

5) Overload duration T, the time interval from the first sampling point to the last sampling point during overload;

6) Current unbalance rate β1 during overload, the average value of the three-phase unbalance rate during overload is taken as the unbalance rate β1 value during overload;

Through the determination of overload attributes, all overload attribute data are expressed in the form of a matrix, as follows:

Where D is the overload attribute matrix; Λ _m is a certain attribute vector; a _nm represents the overload attribute unit.

The selection of the samples will directly affect the prediction results of the distribution transformer overload risk assessment. Some singular samples should be eliminated, including those with a capacity of 5kVA, those that have just been put into operation, and special transformers. The sample selection principles of this article are as follows:

1) Select S9 model and upper transformer;

2) Select a transformer with a capacity greater than or equal to 30kVA and less than or equal to 800kVA;

3) Only conventional transformers are considered, and high overload distribution transformers and on-load capacity adjustment transformers are not considered;

4) Operation data of distribution transformer for two years or more;

5) Select samples based on the ratio of overload distribution transformers of different capacities.

The determination of the overload attribute requires correlation calculation to verify whether it meets the neural network input requirements; when two or more independent variables injected into the neural network are highly correlated, it will have a negative impact on the learning ability of the neural network. Removing redundant variables will achieve faster training time. Adaptive neural networks can be used to streamline redundant connections and neurons. The correlation analysis of the correlation degree of the two overload attributes is as follows:

Correlation coefficient matrix:

Where, r _ij represents the correlation coefficient between overload attribute vectors _Xi and _Xj , and its calculation formula is as follows:

The definition of correlation degree is shown in Table 1.

Table 1 Definition of correlation

Correlation coefficient Degree of relevance 0.00～±0.3 Micro-correlation ±0.3～±0.5 Low correlation ±0.5～±0.8 Moderately correlated ±0.8～±1.0 Significant correlation

According to the correlation calculation of the overload attribute vector r _ij , it is verified whether the neural network input requirements are met according to Table 1.

The beneficial effect of the present invention is that the distribution transformer overload risk assessment proposed by the present invention can accurately determine the risk level of distribution transformer overload shutdown or damage, and can provide a basis for discovering distribution transformers with relatively high overload risks and taking targeted measures when unexpected changes in power load occur, thereby contributing to the controllability of transformer overload operation, with an accuracy rate of over 87%.

Compared with the thermal model of winding hot spot temperature, the overload risk assessment method of distribution transformer is to obtain a common implicit model for transformer groups through a large amount of overload operation data mining, rather than constructing a thermal model from the transformer overload temperature characteristics. This assessment method has strong practicality.

When using neural networks to evaluate the risk of overload and shutdown of distribution transformers, there is no need to consider the internal parameters of the distribution transformers; the traditional method does not take into account that the distribution transformers have a certain overload capacity, and once the distribution transformers are overloaded, corresponding treatment measures are taken. The evaluation method achieves the purpose of controlling the overload operation risk of the transformer while ensuring the overload operation of the transformer.

A distribution transformer overload risk assessment method proposed in the present invention extracts historical operation discrete data from the perspective of distribution transformer operation, analyzes the correlation between various attributes during overload, uses neural network technology to analyze and mine the overload operation data, and evaluates the influence of different operating states on the operation mode when the distribution transformer is overloaded. The assessment result is helpful for top-level decision-making and actual production, and effectively controls the overload operation of the distribution transformer while accurately assessing the overload risk degree of the distribution transformer.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of an overload risk assessment model of the present invention.

DETAILED DESCRIPTION

The specific implementation of the present invention is shown in Figure 1, and the specific steps of implementing the present invention are as follows:

(1) Data collection

In the distribution transformer monitoring related system, the substation operation records contain rich information, such as capacity, voltage value, current value, collection time, active power, reactive power, etc. The weather conditions when the distribution transformer is overloaded are obtained through the meteorological monitoring system to constitute the original data of the distribution transformer operation.

(2) Data preprocessing

The discrete operation data of the distribution transformer cannot be directly used as the neural network input. The operation data needs to be processed according to the overload properties of the distribution transformer, and the noise data and empty data in the original data need to be eliminated.

1) Distribution transformer overload shutdown criterion: If there are two consecutive load factors less than 0.01 in the distribution transformer overload, the distribution transformer is considered to be out of service.

2) Coefficient K1 before overload: The average value of the n load coefficient K collection points before overload is taken as the coefficient K1 before overload.

3) Overload coefficient K2: Take the average value of n load coefficient K collection points during overload as the overload coefficient K2 value.

4) Overload coefficient K3: The average value of the n load coefficient K collection points after overload is taken as the overload coefficient K3 value.

5) Overload duration T. The time interval from the first sampling point to the last sampling point during overload.

6) Current unbalance rate β1 during overload: The average value of the three-phase unbalance rate during overload is taken as the unbalance rate β1 value during overload.

By determining the overload attributes, all overload attribute data can be expressed in the form of a matrix, as shown in the following example:

As shown in formula 1:

(3) Sample selection

The selection of samples will directly affect the prediction results of the distribution transformer overload risk assessment. Some singular samples should be eliminated, such as those with a capacity of 5kVA, those that have just been put into operation, and special transformers. The sample selection principles are as follows:

1) Select S9 model and upper transformer.

2) Select a transformer with a capacity greater than or equal to 30kVA and less than or equal to 800kVA.

3) Only conventional transformers are considered, and high overload distribution transformers and on-load capacity regulation transformers are not considered.

4) Distribution transformer operation data for two years or more.

5) Select samples based on the ratio of overload distribution transformers of different capacities.

(4) Correlation Verification

The correlation of the overload attribute vector is calculated to verify whether it meets the neural network input requirements.

By analyzing the degree of association between the two overload attributes through correlation, when two or more independent variables injected into the neural network are highly correlated, it will have a negative impact on the learning ability of the neural network. Removing redundant variables will result in faster training time. Adaptive neural networks can be used to streamline redundant connections and neurons.

1) Correlation coefficient matrix

Where, r _ij represents the correlation coefficient between overload attribute vectors _Xi and _Xj , and its calculation formula is as follows:

2) Calculation formula

3) Correlation degree. The definition of correlation degree is shown in Table 1.

Table 1 Definition of correlation

The correlation of the overload attribute vector is calculated, and the corresponding table 1 is used to verify whether it meets the neural network input requirements.

(5) Network modeling

The distribution transformer overload risk network structure is an artificial neural network. Artificial neurons are mathematical models that simulate biological neurons. They are nonlinear elements with multiple inputs and a single output. Each input quantity x _i of a neuron has a corresponding weight w _i . The processing unit quantizes the weighted inputs, then adds them together to obtain the sum of the weighted values, which is then summed with the deviation to form the input of the neuron transfer function.

The output of a neuron is a function of its weighted input, as shown in Formula 4:

BP neural network is a multi-layer perceptron structure, including input, output layer and several hidden layers, which is mainly divided into forward propagation stage and backward propagation stage. In the forward propagation stage, information is transmitted from the input layer to the output layer through the hidden layer step by step, as shown in formula (5):

O _i =f _n (f _n-1 ((...f ₁ ([X _i ][W ₁ ])+[B ₁ ]...)[W _n-1 ])+[B _n-1 ] ) (5)

_Xi represents the i-th sample input, _W1 …Wn _-1 represents the weight matrix of the hidden layer, _B1 … _Bn-1 represents the bias matrix of the hidden layer, _f1 … _fn-1 represents the weight function of the hidden layer, _fn represents the output function, and _Oi is the expected value.

In the back propagation stage, the weights and biases of the network are repeatedly adjusted and trained. The training process requires the input vector X and the target value Y. The network training process is the process of minimizing the mean square error. Suppose the error accuracy of the i-th sample is as shown in formula (6):

The error of the entire m sample set is shown in formula (7):

(6) Simulation training

The network training error is set, such as 10 ^-2 , and the number of neurons in the input layer, hidden layer, and output layer is determined through multiple tests. The input layer is the overload attribute vector, and the output layer is whether the corresponding distribution transformer is out of service. If it is out of service, it is set to 1, and if it is not out of service, it is set to 0. The training algorithm uses the BP algorithm, and the activation function uses the tansig function. The training samples randomly select 120 overload outage samples and 200 overload non-outage samples, and continuously change the weights and biases along the direction where the error function decreases fastest until the training error is less than the specified value, that is, the trained weight matrix and bias matrix are obtained, and then the network structure that can be used to predict the overload risk of distribution transformers is obtained.

(7) Prediction results

The test samples are 30 overload shutdown samples and 50 overload non-shutdown samples. After training, the corresponding neural network weights and deviations are obtained. The trained network is used to test the test samples. If the output result is greater than 0.4, it is a shutdown area, otherwise it is a non-shutdown area. Then the trained neural network structure is used in the actual operation of the distribution transformer. The test results are shown in Table 2:

Table 2 Test results

。

.

Claims (4)

1. A distribution transformer overload risk assessment method, comprising collecting distribution transformer operating data, preprocessing the original data according to the overload attribute, and obtaining the distribution transformer overload attribute set, it is characterized in that the method utilizes neural network technology to Network modeling and simulation training are carried out on the attribute set of distribution transformer overload, and the network structure of distribution transformer overload risk is obtained; according to the network structure, the degree of risk of distribution transformer outage in different operating states during overload is evaluated, and the distribution transformer overload is simulated. The probability; the greater the probability, the greater the possibility of transformer return; The operation data includes the information contained in the operation records of the distribution transformer in the distribution transformer monitoring related system: capacity, voltage value, current value, collection time, active power, reactive power; Condition; The distribution transformer overload risk network structure is an artificial neural network, and the artificial neuron is a mathematical model simulating a biological neuron, and the output of the neuron is a function of the weighted input, as follows:

Among them, y is the output of the neuron; x _i is the input of the neuron; w _i is the weight corresponding to the input of the neuron; BP neural network is a multi-layer perceptron structure, including input, output layer and several hidden layers, which are divided into forward propagation stage and backward propagation stage. In the forward propagation stage, information is transformed step by step from the input layer through the hidden layer. is passed to the output layer, as shown in the following formula: O _i ＝f _n (f _n-1 ((...f ₁ ([X _i ][W ₁ ])+[B ₁ ]...)[W _n-1 ])+[B _n-1 ] ) X _i represents the i-th sample input, W ₁ ... W _n-1 represents the weight matrix of the hidden layer, B ₁ ... B _n-1 represents the bias matrix of the hidden layer, f ₁ ... f _n-1 represents the hidden layer Weight function, f _n represents the output function, 0 _i is the expected value; In the backward propagation stage, the weights and deviations of the network are repeatedly adjusted and trained. The training process needs to provide the input vector X and the target value Y. The process of network training is the process of minimizing the mean square error error. Let the i-th sample The error precision E _i of is shown in the following formula:

The error E of the entire m sample set is shown in the following formula:

The original data is preprocessed according to the overload attribute. The discrete operating data of the distribution transformer cannot be directly input as a neural network. The operating data needs to be processed according to the overload attribute of the distribution transformer. The noise data and empty data in the original data need to be eliminated. ; 1) Distribution transformer overload outage criterion, if there are two consecutive load factors less than 0.01 in the distribution transformer overload, the distribution transformer is considered out of service; 2) Coefficient K1 before overload, take the average value of n load coefficient K collection points before overload as the value of coefficient K1 before overload; 3) Overload coefficient K2, take the average value of n load coefficient K collection points during overload as the overload coefficient K2 value; 4) Coefficient K3 after overload, take the average value of n load coefficient K collection points after overload as the value of coefficient K3 during overload; 5) Overload duration T, the time interval from the first collection point to the last collection point during overload; 6) The current unbalance rate β1 when overloaded, take the average value of the three-phase unbalanced rate when overloaded as the unbalanced rate β1 value when overloaded; Through the determination of overload attributes, all overload attribute data are expressed in the form of matrix, as follows:

Among them, D is the overload attribute matrix; Λ _m is a certain attribute vector; a _nm represents the overload attribute unit; The determination of the overload attribute requires correlation calculation to verify whether the input requirements of the neural network are met; when two or more independent variables injected into the neural network are highly correlated, it will have a negative impact on the learning ability of the neural network, and remove Redundant variables will obtain faster training time, and the adaptive neural network is used to streamline redundant connections and neurons; the correlation analysis of the correlation degree of the two overload attributes is as follows: Correlation coefficient matrix:

In the formula, r _ij represents the correlation coefficient between overload attribute vector X _i and X _j , and its calculation formula is as follows:

2. a kind of distribution transformer overload risk assessment method according to claim 1, is characterized in that, described simulation training, network training error is set, determines input layer, hidden layer, output layer neuron number by multiple tests , where the input layer is the overload attribute vector, and the output layer is whether the corresponding distribution transformer is out of service. If it is out of service, it is set to 1, and if it is not out of service, it is set to 0; Randomly select 120 overloaded shutdown samples and 200 overloaded non-stopped samples, and continuously change the weights and deviations along the direction of the fastest decrease in the error function until the training error is less than the specified value and then stop, that is, the trained weights are obtained Matrix and bias matrix, and then get the network structure for predicting the overload risk of distribution transformers. 3. A distribution transformer overload risk assessment method according to claim 1, characterized in that the collection of distribution transformer operation data includes the information contained in the distribution transformer monitoring related system, station area operation records: capacity , voltage value, current value, collection time, active power, and reactive power; the weather conditions when the distribution transformer is overloaded are obtained through the meteorological monitoring system to form the original data of the distribution transformer operation. 4. A distribution transformer overload risk assessment method according to claim 1, characterized in that the selection of the samples will directly affect the prediction results of the distribution transformer overload risk assessment, and some singular samples in the samples should be eliminated, including capacity 5kVA, just put into operation, special transformer, the sample selection principles are as follows: 1) Select the S9 model and the upper transformer; 2) Select a transformer with a capacity greater than or equal to 30kVA and less than or equal to 800kVA; 3) Only conventional transformers are considered, and high overload distribution transformers and on-load capacity regulating transformers are not considered; 4) Distribution transformer operation data for two years or more; 5) Select samples according to the proportion of overload distribution transformers with different capacities.

Images