Dissolved Gas Analysis of Transformer Oil Based on Bayesian Regularized BP Neural Network

The hidden fault diagnosis of the transformer will play an important role in the future smart grid, as the safe operation of the transformer is an important guarantee of power system stability. The attention value method which is being applied to actual substations is introduced, according to a real case that took place in 2010, after detailed experiments, the method proposed in this paper has been verified, that is, a model based on Bayesian regularized BP neural network has been successfully established. The hidden fault in the transformer can be found earlier by DGA based on Bayesian regularized BP network model.


INTRODUCTION
The future development of the smart grid must require highly reliable monitoring technology of electrical equipment.More intelligent monitoring technology will be bound to bring revolutionary progress to human life and industrial production based on the smart grid in the future.Currently, it is of utmost importance to diagnose the faults in electrical equipment, especially before the accident [1].On March 21, 2018, a severe power outage occurred in Belomonte, Brazil.This incident once again proves that the operational safety of electrical equipment is of paramount importance.In the future, the stability and reliability of power supply equipment is a key concern for electrical equipment.As a piece of important equipment to transmit electric energy, the safe operation of the transformer can ensure the smooth transmission of the power grid and avoid the bad social impact caused by the power outage.
The dissolved gas in transformer oil can directly reflect the operating state of the transformer, much work so far has focused on the monitoring technology of dissolved gas in transformer oil. Figure 1 shows the entire process of dissolved gas in transformer oil, obtained by online real-time monitoring.In 1952, researcher Martin proposed gas chromatography [2], and in 1961, Pew used gas direction chromatography to detect early transformer faults and problems [3].In the following decades, transformer oil dissolved gas analysis (DGA) has been intensively investigated.In 1972, the UK Power Authority proposed a four-ratio method for DGA in transformer oil [4].Syprotec Inc. of Montreal, QC, Canada developed a dissolved gas on-line monitoring device to detect early morning faults of transformers by detecting gases such as CO and H 2 in 1975 [5].Based on a large number of field experiences, it was shown that the ratio of C 2 H 6 to CH 4 in the four-ratio method has little effect on the result of the diagnosis.Therefore, the four-ratio method was improved to a three-ratio method by International Electrotechnical Commission (IEC) [6], and then further studied by the Japan Electric Association (JEAC) to excavate the inherent laws.The fluctuation range of fault characteristic quantity of three ratios were given, and the prediction accuracy was improved [7].The analysis technology of dissolved gas in transformer oil based on the BP neural network model was enhanced with the continuous development of neural networks and was officially proposed in 1993 [8].Subsequently, over the course of two decades, the BP neural network model was extensively improved and studied.
At present, great progress has been made in the research of DGA combined artificial intelligence.Although the improved three-ratio method improved the accuracy of DGA [9], the fault ratio characteristics of dissolved gas were not easy to discriminate so some researchers used intelligent algorithms to diagnose the hidden fault in the transformer [10].The fault diagnosis model of the transformer based on the back-propagation (BP) neural network was studied by using fault gas as input quantity [11].The fuzzy analysis method was also applied to the DGA of transformer oil, a fault diagnosis model was established based on the analysis of fault characteristics [12].In addition, the expert system of fault diagnosis had also been established by designing the expert database, the findings provided an effective guarantee of online fault diagnosis in transformers [13][14][15].Although many studies mentioned above have conducted in-depth research and discussion on DGA in transformer oil, the model still has great problems in accurately judging faults.Since the Bayesian regularized BP neural network model significantly increases the generalization ability used by the network, the accuracy of the test is improved [16][17][18].Therefore, this paper hopes to further study the process and change of dissolved gas in transformer oil based on the Bayesian regularized BP neural network model.

Figure 1. Flow chart of reviews of on-line monitoring of dissolved gas in transformer oil
This article is organized as follows.Section 2 shows the dissolved gas attention value method which is being applied to actual substations and the fault diagnosis of transformer oil.Then, Section 3 is mainly a systematic overview of the BP neural network model and the theory of Bayesian regularization, on this basis, a transformer fault diagnosis model is proposed.Section 4 verifies the method proposed in this paper based on the measured data of the No. 2 transformer in Dongjiao Substation, Jilin Province.In Section 6, we provide some important conclusions that are drawn from this investigation.

ATTENTION VALUE AND FAULT DIAGNOSIS
It is important evidence of transformer fault alarm that whether the concentration of dissolved gas reaches the attention value.Attention values for each gas are shown in Table 1.If the concentration of dissolved gas exceeds the attention value, the transformer is at risk of potential failure.However, a real case that took place in 2010 made people think about the method of diagnosis.The results of the transformer oil dissolved gas test showed that all the dissolved gas indexes in the transformer did not reach the attention value on April 30, but it was found that hydrogen was above the attention value on May 20.Therefore, the micro-water test and relatively simplified test of the transformer were carried out, but the data showed normal.Since the transformer was not detected for any fault, the transformer continued to operate for four months accompanied by regular oil-dissolved gas tests after May 20.The results of the monitoring on 16 July can be found that the CH 4 concentration exceeded the attention value.H 2 concentration was 479.76 / L L  , which was far more than that 100 / L L  .Until September, the transformer was shut down and the cause of the failure was found.The analysis shows that the transformer is wet during the rainy season, most of the water exists in the insulating material according to the strong absorbability of the insulating material, and the anomalies of the transformer are not easy to detect at room temperature unless the temperature is greater than 80 degrees Celsius that water was dissolved into the oil.
From the above analysis, the hidden fault of the transformer existed before May 20.However, the causes of failure were not found in time through a comparison of attention values.Therefore, it is necessary to study more effective methods to detect anomalies in time to ensure the stable operation of transformers.

The Basic Principle Of BP Neural Network
BP neural network algorithm is a simulation of the human brain, the mechanism is established through the accumulation of the neurons stimulation.
The information is transferred to the hidden layer neuron by each neuron in the input layer, and the information is weighted to the output layer by the hidden layer neuron, and the weight of each layer is improved by the feedback result of the output layer.When the model can meet the criteria of self-calibration of prediction samples, the model can be used to predict the results of other samples.
If the neural network has n inputs, q outputs and p hidden layers, there are p hidden weight and q output weight in the neural network.The hidden layer can be represented: Where ij w refers to the relationship and weight between the hidden layer of neurons NO. j and the input layer NO. i; i x refers to the simulated neuron NO. i in the input layer; j  is the threshold of NO. j momentum.Function logsig can be used for each layer of neuron output function: Output layer neurons can be shown as: where kj w is the connection weight between the output layer NO. k neuron and hidden layer NO. j neuron; k  is the threshold of the NO.k momentum.

The actual results k y and expected results k
o can be obtained through the above calculations so that the weights and the thresholds can be revised: )] ( 1)  is the revised coefficient.
The revised coefficient of weight and threshold between the output layer and hidden layer can be shown as: where k  is the revised coefficient.

Bayesian Regularization
Although the BP neural network model has a very powerful mapping role and ability, the speed and accuracy of its operation need to be further improved.Therefore, Bayesian Regularization can be used in the BP neural network to optimize network structure so that the generalization ability of the network can be raised.
Conventional BP neural network object function is: The n in the formula represents the total number of samples; i a refers to the actual network output; i t indicates the desired output.
The calculated sum of squares of network weights is: where m in the formula is the weight ratio; i w represents the actual weight of each NO.i.
Objective function of the BP Neural Network based on Bayesian regularization is: where  and  are parameters.
According to Formula (10), the complexity of the network can be controlled effectively by the correction of  and  , and the network generalization ability can be raised.

A transformer analysis fault model is established by using Bayesian regularized BP neural network model
Figure 3 shows the process of diagnosing faults and problems with a transformer by using the BP neural network model.There are 5 input nodes, 9 hidden layer neurons and 5 output nodes in the designed BP neural network model.

VERIFICATION OF THE METHOD
According to the dissolved gas conditions under different time conditions, combined with the BP neural network model, the actual operating status of the current transformer can be obtained.Table 3 shows the data output shown in Figure 2. The hidden fault that was not found by the attention value method on April 30 was found on May 20, and it worsened on June 16.The low energy discharge on April 30 was found by the method in this paper.Compared to the attention value method which is being applied to actual substations, Bayesian regularized BP neural network model can diagnose the hidden fault earlier.To explain the importance of Bayesian regularization, the normal BP neural network is also used in DGA, the result is shown in Table 4.According to the result in Table 4, the normal BP neural network cannot find the hidden fault on April 30 because the estimated status shows normal.However, the Bayesian regularized BP neural network can find hidden faults earlier.
To prove that there was a hidden fault in the transformer on April 30, this fault event was studied intensively by laboratory assistants, the conclusion is that the discharge in the transformer was caused by the dampness in the transformer which made low energy discharge change to high energy discharge, and the dampness in transformer existed on April 30.

CONCLUSION
The study demonstrates that the Bayesian regularized BP neural network can diagnose the hidden fault in the transformer earlier.Therefore, the method in this paper that analyzes the dissolved oil effectively can guarantee the safety of electric power equipment so that the safe operation of the power grid is ensured.The new method proposed in this paper is expected to make a significant contribution to the use of future smart grid systems to detect electrical equipment faults and problems.

Figure 2 .
Figure 2. Transformer oil dissolved gas of the No.2 transformer in Dongjiao Substation

6 Figure 3 .
Figure 3. BP neural network models diagnose faults and problems

Table 1 .
Attention values for each gas The changes in the No. 2 transformer in Dongjiao Substation are shown in Figure 2, based on experimental data from State Grid Jilin Maintenance Company.

Table 3 .
Actual output of BP neural network model

Table 4 .
Output information of normal BP neural network