Intelligent prediction of positive polarity switching impulse discharge voltage of a rod-plate long gap

Author： GOZ Electric Time：2024-10-21 09:40:29 Read：20

Abstract: The discharge characteristics of long air gaps are an important basis for the design of external insulation of UHV transmission and transformation projects, and the discharge characteristics of typical rod-plate air gaps have always been a basic issue of concern to researchers. At present, there are many methods for calculating the discharge voltage of rod-plate gaps, but most of them cannot adapt well to a wide range of temperature and humidity changes. In order to achieve accurate calculation of the discharge voltage under extreme temperature and humidity conditions, this paper proposes a method for predicting the discharge voltage of rod-plate long gaps based on regularized invariant risk minimization-neural network (IRM-NN). The influencing factors of rod-plate gap discharge are systematically analyzed, and key features are extracted as input to train the model. The mean absolute percentage error of the model on the test set is only 1.6%, which verifies that the model can be effectively extrapolated to application scenarios outside the test conditions. Then, the prediction effect of the model is compared with that of three commonly used machine learning models. The results show that the calculation accuracy of the model on samples outside the test range of the training samples is significantly higher than that of other models. The proposed calculation method of 50% discharge voltage of rod-plate gap can adapt to a wide range of temperature and humidity and certain electrode size changes, and can provide a reference for the study of discharge characteristics of long air gaps.

Keywords: long air gap; discharge voltage; temperature and humidity changes; invariant risk minimization; neural network

Long air gap is the main form of external insulation of UHV transmission lines. Its insulation characteristics are affected by many factors such as electrode structure and atmospheric environment. The characteristics are very complex and are one of the basic issues that have long been concerned in the field of high voltage engineering. As a typical gap, rod-plate gap is an extremely non-uniform gap, and its discharge voltage is much lower than other gap types. Therefore, studying the discharge characteristics of rod-plate gap is representative and of engineering value.

Due to the lack of a solid and complete theoretical foundation, the discharge characteristics of air gaps are generally obtained by conducting a large number of air gap discharge tests, but the real test has the disadvantages of high test cost and long cycle. To this end, many researchers are committed to the study of gap discharge voltage calculation methods, such as empirical formulas, semi-empirical formulas and physical models. Empirical formulas are obtained by fitting discharge test data, such as Gallet formula and CRIEPI formula. Semi-empirical formulas use analytical formulas to calculate discharge voltage, such as the Mosch method, the Carrara method, and the Rizk method. Although empirical and semi-empirical formulas are simple and convenient, some parameters in the formulas are inferred from test data, which are only applicable to certain specific conditions and have poor generalization. Physical models simulate the entire discharge process and obtain the discharge voltage according to the final jump moment, such as the Huztler model, the Goelian-Gallimberti model, the Becerra-Cooray model, and the Fofana equivalent path model. Since the current observation methods and mechanism research of gap discharge are not perfect, the calculation results of the physical model usually have obvious errors with the experimental values.

With the rapid development of artificial intelligence technology in recent years, the excellent ability of machine learning algorithms in solving multidimensional nonlinear problems has gradually become known to the public, and the use of intelligent algorithm modeling to achieve control and decision-making of complex systems has become a common idea for solving engineering problems. In terms of the calculation of the discharge voltage of air gaps, some researchers have successively proposed methods for constructing discharge voltage calculation models based on various machine learning algorithms. The relevant literatures respectively apply neural network algorithms such as BP neural network, RBF neural network and self-organizing map neural network, and use gap distance and meteorological parameters as input to build models, achieving high prediction accuracy within the test conditions. However, the samples are all short gap test data, and the number of samples is small. The model is very easy to overfit and difficult to generalize to the calculation of long gap discharge voltage. The relevant literature obtains the electric field distribution of the shortest path of the air gap through simulation, extracts the electric field characteristic quantity as input, and constructs a support vector machine model to realize the calculation of discharge voltage. Compared with the neural network model based on the principle of empirical risk minimization, the support vector machine updates parameters based on the principle of structural risk minimization, is not easy to overfit, and is more suitable for small sample scenarios. However, these studies are still only for short gap discharge test data, and rod-plate and DC tower gap discharge voltage prediction models based on support vector machines and extreme random trees are respectively constructed. The model uses gap structure parameters and meteorological parameters as input, does not require complicated electric field simulation process, and forms an air gap discharge voltage calculation method that is easy to operate and suitable for actual UHV projects. The gap structure types covered by this method are relatively comprehensive and can basically meet application requirements. However, the samples are all taken from the test results of the Beijing UHV DC test base in China, and there are fewer test data in winter and summer, covering a limited range of meteorological conditions. Such methods are usually only applicable to target prediction within the training sample parameter range. When the target's meteorological parameters exceed the training sample range, the generalization performance of the method cannot be guaranteed.

When the meteorological parameters change, the probability distribution of the characteristic variables and the discharge voltage will also shift, making it difficult for the classic machine learning algorithm to predict targets that are different from the training sample distribution. As the prediction failure caused by distribution shift becomes more and more obvious, the out-of-distribution generalization learning method for this problem has become one of the hot spots in artificial intelligence research in recent years. Researchers have successively proposed new machine learning frameworks such as domain generalization, invariant learning, and stable learning, and the means to solve the distribution shift problem have become more mature. In order to extend the model trained by limited test samples to application scenarios outside the test conditions, this paper constructs a rod-plate air gap discharge voltage prediction model based on regularized IRM-NN based on the idea of invariant learning. The model uses 76 sets of test data as the training set, selects gap distance, air pressure, dry temperature, and absolute humidity as input, and 50% discharge voltage as output. The test set contains 50 groups of test data, of which 41 groups of samples have temperatures outside the training sample temperature range. When the test samples are input into the model, the maximum absolute percentage error of the predicted value is 5.1%, and the average absolute percentage error is only 1.6%, which verifies the effectiveness and generalizability of the model.