Research on power dynamic data sample generation technology based on brain-like computation and its efficient computation methods

Extra-high voltage DC converter station refers to the station established in the high-voltage DC transmission system in order to complete the mutual conversion of alternating current (AC) and direct current (DC) and to meet the requirements of the power system for safety, stability and power quality. An extra-high voltage DC converter station is a high-voltage DC system facility that functions as a rectifier, inverter, or both rectifier and inverter stations. It consists of one or more converters, corresponding transformers, reactors, filters, reactive power compensation equipment, control, vision, protection, measurement equipment, and auxiliary equipment.

2.1.1

Site distribution and equipment

Taking an extra-high voltage DC converter station in China as an example, the internal field distribution of this converter station is shown in Fig. 1. It mainly includes the converter area, AC field, DC field and AC filter area, and each area contains the corresponding main equipment, which is introduced as follows [21].

Figure 1.

Internal field distribution of UHV DC converter station

1)

Converter area

The converter area contains two main devices: the converter valve and the converter variable. The converter valve is responsible for converting AC to DC (rectifier) or DC to AC (inverter). The converter is responsible for power transmission, converting AC voltage to the phase-change voltage required by the converter valve, and isolating the DC system from the AC system. The converter station contains 32 sets of converter transformers and 5 sets of standby converter transformers. Each converter transformer has a capacity of 395MVA, and the total capacity of the converter is 108,340.5MVA.

2)

DC field area

The DC field area is mainly responsible for handling DC current. It contains dozens of DC equipment, including a flat wave reactor, which is used to smooth DC current ripples and reduce transient current. DC voltage dividers divide the current flowing through the equipment to achieve DC voltage measurement. Current transformers are used to convert large and small currents, high and low voltages, and for electrical isolation. The converter station contains 5 DC voltage dividers, configured on the pole bus and neutral of each pole. There are 25 photocurrent transformers and 10 zero-flux transformers. There are 5 DC circuit breakers, one NBS DC switch configured at the station for the grounding pole and two NBGS configured at the neutral line of the two poles in the station.

3)

AC field area

There are a large number of GIS equipment, i.e., gas-insulated metal-enclosed switchgear, in the AC field area, which is a kind of electrical equipment using sulphur hexafluoride or other gases as insulating medium and sealing many high-voltage components in a grounded metal casing, including circuit breakers, disconnecting switches, voltage transformers, current transformers, surge arresters, busbars, bushes, casings and so on.

4)

AC filter area

The converter station will generate harmonics in the process of converting AC and DC power, and the existence of harmonics will affect the whole power production process. The role of the AC filter is to suppress harmonics and, at the same time, to perform reactive power balance and reactive power compensation for the converter. The AC filters of this converter station are divided into 5 major groups and 24 subgroups, with a single group capacity of 295Mvar and a total reactive power capacity of 5050Mvar.

Power equipment is distributed in different fields according to their functions in the converter station, and equipment of the same type is gathered together to provide specific services for the entire system. This feature of aggregation and distribution according to the field area helps the management and maintenance of the equipment, and the upper computer distributed in the field area can be especially responsible for the individual equipment managed under it, which creates the conditions for the design and deployment of the monitoring system and the early warning system.

2.1.2

Converter station equipment operation and inspection programme

Due to the problems of a few fault samples of UHV converter equipment and the inability to effectively support intelligent operation inspection of equipment, this paper proposes an intelligent operation inspection scheme for UHV DC converter station equipment based on multimodal brain-like learning. The scheme is designed based on the principles of reliability, safety, economy, practicality, operability, and maintainability, while taking into account technological advancements, scalability, and compatibility. While leading the future development of DC converter station intelligent operation and inspection technology, it improves the technical standards and work efficiency of DC converter station equipment operation and inspection, and the specific design principles of the operation and inspection programme are [22]: 1)

On the basis of meeting the technical requirements of the new generation of intelligent substations, all high-voltage electrical equipment in the UHV DC converter station is subjected to all-weather, all-around, panoramic data intelligent operation inspection, as shown in Figure 2.

2)

A special maintenance file and overhaul programme are customized for each piece of equipment, and a comprehensive assessment of the operating status is carried out by combining the inherent structural characteristics, electrical characteristics, operating data and other parameters of each piece of equipment. A scientific and effective fault assessment result is formed, which makes the transition from “regular maintenance” to “state maintenance” possible for the equipment maintenance of DC converter stations.

3)

Taking the nearest-neighbour generation segment technology, time-space fusion and other advanced computing tools in the digital era as the core, the fault database and expert knowledge base of high-voltage electrical equipment in DC converter stations are established, and maintained and updated by self-learning algorithms.

Figure 2.

Technical requirements of a new generation of intelligent substations

Aiming at the problem that the fault samples of UHV converter equipment are few and cannot effectively support the intelligent operation and inspection of equipment. In this paper, samples are generated based on the nearest-neighbour generative fragmentation technique (GPNN), fusing the temporal evolution law and the similarity of adjacent samples. A spatio-temporal fuser is constructed to fuse the samples based on nearest-neighbour temporal splicing with multi-level multi-dimensional space to construct a multimodal brain-like learning sample spatio-temporal correlation generation technology for the operation and inspection of UHV converter equipment.

2.2.1

The PNN model

When the fault data matrix exhibits extreme sparsity, traditional models may limit the ability to mine shallow patterns, i.e., low-order feature combinations, from the data [23]. On the other hand, deep models such as deep neural networks cannot be directly applied to high-dimensional inputs due to their huge feature space. Instead, a neural network (PNN) with embedding layers can solve such problems. The PNN has an embedding layer that can be used to learn a distributed representation of the categorical data, a product layer for capturing relational patterns between intermediate features, and a fully connected layer for exploring higher-order feature interactions.

The network architecture of the PNN model is shown in Figure 3. From the top down, the output of the PNN y is the predicted click-through rate, which is a real number between 0 and 1: 1y=σ(W3l2+b3) where W₃ ∈ R^1×D2 and b₃ ∈ R are the parameters of the output layer, l₂ ∈ R^D₂ is the output of the second hidden layer, and σ(x) is the sigmoid activation function: 2σ(x)=1/(1+e−x) and use D_i to denote the dimension of the i nd hidden layer.

Figure 3.

Network architecture of PNN model

The output l₂ of the second hidden layer can be represented as follows: 3l2=relu(W2l1+b2) Where l₁ ∈ R^D₁ is the output of the first hidden layer. The modified linear unit (relu), defined as: 4relu(x)=max(0,x)

Since relu has very good performance and is easy to compute, it is chosen as the activation function for the hidden layer. The first hidden layer is fully connected using multiplication. The input contains a linear input l_z and a quadratic input l_p. These two inputs can be represented as l₁: 5l1=relu(lz+lp+b1) which are all l_z, l_p and the bias vector b₁ ∈ R^D₁.

Next define the tensor inner product operation: 6 A□B=∑i,jAi,jBi,j

Firstly, the vectors A, B are multiplied bitwise, and how the results of the multiplication are added up to get a scalar. Finally, l_z, l_p is computed by multiplying in z, p : 7 lz=(lz1,lz2,……,lzn,……lzD1),lzn=Wzn□z 8 lp=(lp1,lp2,……,lpn,……..lpD1),lpn=Wzn□p Where Wzn , Wzn are the weight parameters of the product layer and their dimensions are determined by z and p.

By introducing a constant input of “1”, the product layer not only obtains the quadratic signal p, but also maintains the linear input z. Specifically: 9z=(z1,z2,……,zN)=(f1,f2,……,fN) 10p={ pi,j },i=1…N,j=1…N Where f_i ∈ R^M is the coding vector of the i nd element. p_i,j = g(f_i, f_j) is defined as paired features. The PNN model in this paper can be implemented differently depending on the unused g.

The coding vector f_i is the output of the coding layer: 11fi=W0ix[ starti: endi ] Where x is an input feature vector including multiple dimensions, x [start_i : end_i] represents a one-bit valid encoding of i. W₀ represents the parameters of the encoding layer, and w ∈ RM ×(end – starti + 1) and i are fully connected.

Finally, supervised training is applied to minimise the log loss, which is widely used to measure the variability of two probability distributions: 12L(y,y′)=−ylogy′−(1−y)log(1−y′) Where y is the true value (1 if clicked and 0 if not clicked) and y′ is the click-through rate predicted by the model in this paper.

2.2.2

GPNN model

On the basis of the PNN model, the concept of “group” is introduced to construct the nearest-neighbour generative fragmentation technique (GPNN). Different from the concept of feature domain in the native PNN model, the consistency of feature expression in each feature group is fully considered, and the GPNN model divides the features with strong correlation into one or two groups and enters the Product layer in the form of a group to carry out the analysis operation.

The network structure of the GPNN model is shown in Fig. 4, which is based on the PNN model, and embeds an inner-product structure in the Product Layer to realise that the generated fault samples are continuously adjusted internally.

Figure 4.

Network structure of GPNN model

Based on the PNN model, GPNN introduces the grouped Embedding Layer module before Embedding Layer, i.e., summing all feature vectors within each feature group. Taking D series of features as an example, the output d_i (i ∈ [0,I)) of the ith feature group in a series of features is: 13di=d[ starti: endi ]

Among them: 14d[ starti: endi ]=∑j=1Gap(i−1)<j≤Gap(i)d[ startij: endij ] Where the Gap(·) function defines the offset position of a feature group in the whole feature expression, and d_ij is the j th feature in feature group d_i. Similarly, the output of each feature group G, S is formally defined as g_k (k ∈ [0, K)), s_m (m ∈ [0,M)) respectively.

Typical fault samples are intercepted from the time series process of defect-fault development, and the samples are generated based on the nearest neighbour generating segment technique (GPNN) by fusing the time series evolution law and the similarity of the adjacent samples. The specific implementation process of the nearest neighbour generating fragment technique (GPNN) is shown in Fig. 5, which mainly includes the following: 1)

For any target, construct a sample generator to obtain the initial generation sample based on the device defect evolution law.

Figure 5.

GPNN sequential sample generation process

Spatio-temporal fusion is an effective method proposed to understand the “spatio-temporal contradiction” in the fault data of UHV converter equipment, which can comprehensively utilise the advantages of multiple single data to produce continuous and effective time series data for quantitative inversion, and satisfy the demand for high spatio-temporal data. In this paper, by constructing a spatio-temporal fuser, the samples based on nearest-neighbour time-series splicing are fused with a multilevel and multidimensional space, and the correlation relations in the spatial dimension are extended in the temporal dimension. The correlation in the spatial dimension is used to constrain the degree of change between neighboring samples in the temporal dimension to achieve joint spatiotemporal analysis.

In addition, a sample confrontation learner is constructed to further confront and fuse the generated samples obtained from the spatio-temporal joint analysis against the target samples, as well as the neighbouring samples. The differences between the generated samples and the existing samples in time and space are examined, and the differences are reduced by the adversarial learning method, so that complete and accurate generated samples are finally obtained.

In order to verify the validity of the proposed method, this paper uses the following three-dimensional standard function with a high degree of nonlinearity under consideration of the general case, and the function definition is shown below: 15f(x1,x2)=1.3541(1.8(1−x1))+e(2x1−1)sin(3π(x1−0.5)2)+e3(2x2−0.5)sin(4π(x2−1)2) Where x₁ and x₂ are the independent variables of the function and satisfy the conditions of x₁, x₂ ∈ (0,1).

Fig. 6 shows the 2D and 3D visualisation of the original dataset, including the training set and test set. 400 samples were randomly obtained from the 3D standard function as the original dataset. 70% of the sample set is used as a training set and 30% as a test set. The training set divided under this quantity shows an overall uneven distribution and constitutes a soft model with low accuracy.

Figure 6.

2D and 3D visualization of the original sample

Firstly, the points where the sparse regions are located are detected, and the sparse regions in the feature space are calculated. As the G group in the GPNN model is different, it has almost no effect on the change of the ultimately calculated hyper-rectangular sparse region, and in this paper, the number of nearest neighbours is set to 20 according to the rule of thumb. According to the results of the operation of the GPNN and the upper and lower bounds of the sparse range calculated for each dimension, the sparse intervals of the various dimensions make up the hyper-rectangular sparse region.

Build the GPNN model, use the training set to continuously adjust the hyperparameters, and complete the training of the GPNN model. The hyperparameters of the GPNN are shown in Table 1. Its generator and recogniser network uses a single layer perceptron containing 30 neurons to receive noise and passes through 6 layers, each containing 60 neurons with a sigmoid activation function, and the output is linear.

In the case of ensuring that the distribution of the generated samples is as similar as possible to the distribution of the original data, the group with the smallest difference in the no-fixed L2 in the pre-selected expanded dataset is selected as the new sample. Fig. 7 shows the effect of the virtual input samples after they have been generated based on 280 (400*0.7) samples as the original data, Fig. 7(a) shows the generation of 300 virtual input samples, and Fig. 7(b) shows the generation of 600 virtual input samples. The uniformity change of the sample set before and after adding virtual input samples is shown in Table 2. It can be seen that different numbers of new samples generated by the GPNN model significantly improve the uniformity as well as the coverage of the original samples after being added to the sample set. The minimum unfixed L2 difference of 1.29E-02 is obtained when the original sample + virtual generated sample input is 280+400, which effectively fills the blank area of the data sample.

Figure 7.

2D distribution of original input samples and virtual input samples

EMD is used to measure the similarity between two multidimensional distributions under a certain feature space, also known as Wasserstein distance. In this study, a smaller EMD value represents a greater similarity between the generated sample and the original sample. Four traditional data generation algorithms, MTD, TTD, SMOTE, and CVT, are compared, and the input space EMD values of the virtual data samples generated by different methods and the original data samples are obtained through numerical simulation as shown in Fig. 8. From the experimental results, it can be seen that the new samples generated by the GPNN model have the smallest EMD value of 0.0626 with the distribution of the original samples. That is to say, the generator G of the GPNN model is able to output high-quality samples similar to the original distribution after learning the distribution and information details of the original samples, which is because GPNN can evaluate the performance of the generator through a suitable perspective after improvement. Therefore, it is easier to produce high-quality data samples.

Figure 8.

Contrast the virtual generated by different methods with the original sample EMD

Figure 9 shows the comparison between the Pearson correlation coefficients of the original input samples and the virtual input samples, where Figure 9(a) shows the correlation between the variables of the original input samples, Figure 9(b) shows the correlation between the variables of the virtual input samples, and Figure 9(c) shows the heat map obtained by making the difference between the first two matrices. From the figure, it can be seen that there is almost no correlation between variables x₁ and x₁ (P=-0.132), and the virtual input samples generated by the GPNN model also maintain this property, so the two maintain a very small difference in the values of the matrices (P=-0.017). The validation of EMD indicators and correlations shows that the samples generated by GPNN can well maintain the characteristics of the distribution of the original input samples and the characteristics of the correlations between the variables, so the virtual samples on the feature space are somewhat reasonable.

Figure 9.

Pearson correlation coefficient heat maps of original and virtual samples

In the standard function of two-dimensional input and one-dimensional output, the proposed GPNN method can effectively and efficiently increase the coverage and uniformity of the samples. In addition, the distribution gap between the generated virtual input samples and the original input samples is smaller, which can be closer to the original distribution than other methods. In terms of correlation, the correlation coefficients between variables are also more similar to those of the original input samples. Therefore, GPNN can generate high-quality input attributes of virtual samples in numerical simulation, which provides a good condition for the subsequent output generation and makes the complete virtual samples obtained in the end bring a greater improvement in model accuracy.

Considering the logical, electrical, and physical association relationship between the target equipment and the associated equipment of the UHV DC converter station, the multi-level association relationship of the equipment is obtained. At the same time, the multidimensional correlation relationship between multimodal states is constructed by considering the correlation relationship between structured state quantities, the correlation relationship between state quantities of this paper class and structured state quantities, and the correlation relationship between data of the mapping class and structured state quantities. Among them, the correlation relationship between structured state quantities includes the relationship between seven kinds of gases dissolved in oil, the relationship between partial discharges and grounding currents, and so on. The correlations between inspection record text and state quantities include the relationship between corrosion degree and temperature and humidity, and the relationship between silicone discoloration degree and micro water in oil. Based on the above relationships, spatial generation samples that incorporate these relationships are formed using the spatial arbitrary point derivation method.

In order to verify the effectiveness of the multimodal brain-like learning sample spatio-temporal correlation generation technology in the generation of fault samples of UHV converter equipment, this paper embeds the local discharge of converter transformer, converter valve IGBT micromotor wear fault evolution law to generate a total of 539 columns of samples used for brain-like learning, of which 376 cases are for the converter transformer, and 163 cases are for the converter valve. Based on the equivalent physical model of the converter and the converter valve, the fault development law of partial discharge, high temperature overheating and micro-motion wear of the UHV DC converter equipment is obtained, as shown in Figure 10. Among them, Fig. 10 (a)~(d) shows the trends of discharge pulse repetition rate, maximum value of discharge, high-frequency signal amplitude, and ultra-high-frequency signal amplitude with the operating time of the UHV converter equipment (0~70 hours), respectively. The red line in the figure is the fitted curve, and the folded line is the actual value of the test. Through the time fusion apparatus, the equipment operation stage can be divided into five stages: no discharge, tiny discharge, low-energy discharge, high-energy discharge, and near breakdown. It can be found that the fitted values R² of the four adjusted characteristics of the UHV DC converter equipment are 0.9562, 0.98967, 0.98819, and 0.98066, respectively, which are all greater than 0.95, and the fitting effect is better. The temporal fault evolution mechanism of the UHV DC converter equipment and the correlation relationship between the spatial multimodal state quantities can be well determined.

Figure 10.

Correlation between temporal fault evolution and spatial multimodal state variables

Fusing the time series fault evolution mechanism and the spatial multimodal state quantity correlation relationship, the multimodal fault sample generation model with embedded fault mechanism established based on small sample active learning and nearest neighbour splicing technology generates samples and compares the generated samples with the actual samples to obtain the quality of the generated samples as shown in Table 3. It can be seen that the completeness, coverage, overlap, and recognition quality of the generated samples are 0.97, 0.90, 0.95, and 1.0 respectively. The consistency between the generated samples and the original samples is more than 90%.

The experimental results proved that after adding virtual samples to the original fault samples of the equipment operation inspection of the actual UHV DC converter station, the errors of the models obtained are not significantly different compared to the errors of only the original samples. All of the established models have the best performance with 100% quality of equipment fault identification when virtual samples are added. It shows that the generated samples can support the training of brain-like models for equipment health assessment, fault diagnosis, and trend prediction in UHV converter stations. The virtual samples generated by the GPNN method are basically compatible with the original sample distribution, which proves the effectiveness of the method presented in this paper.