Abstract
It is very important for power grid development research and related technical improvement to obtain the disaster situation of fine-scale distribution network, such as the transportation condition evaluation of distribution network and the wind waterlogging disaster prediction of distribution network. Among them, the wind waterlogging disaster prediction of distribution network is the main one, and the prediction of its disaster degree often determines whether the distribution network can be prevented before and rescued after the disaster. Therefore, in view of the above problems, combined with the actual transmission situation of the distribution network, after collecting the measured disaster data of the distribution network in relevant areas, combined with the multi-source data fusion technology and neural network modeling technology, this paper analyzes the disaster degree indicators of different distribution networks and constructs the relevant fuzzy matrix through the fuzzy theory to evaluate the disaster degree, which is verified by the measured data. This distribution network disaster loss prediction model can effectively implement the disaster loss prediction of distribution network and compare its prediction results with the other two different common models. The comparison results show that the prediction accuracy of the multi-source data fusion prediction model constructed in this paper is more than 0.95 compared with the other two models, while the prediction accuracy of the other two models is not more than 0.9, which proves that the model constructed in this paper has smaller errors. It has the advantages of higher accuracy and faster convergence speed.
1. Introduction
At present, the global climate is changing abnormally, and severe weather such as typhoons and rainstorms is increasing year by year. At the same time, in order to improve the national happiness index and living standards, the power grid facilities put into use and newly built in various places are also increasing year by year. As a result, under the double encounter of the two, the power grid is suffering from more and more disasters caused by severe weather such as typhoons and rainstorms. In places with harsh regional environments and poor geological conditions, the degree of disaster also increases exponentially [1, 2]. In the power grid, due to the fact that the distribution network line grid often requires a special overhead structure, coupled with the complex design of the distribution network lines, the demand is large, and it involves a wide range of factors. The early design was limited due to economic and other factors. The damage level of its wind resistance and waterlogging has also caused the distribution network to become one of the most damaged facilities during the invasion of wind and waterlogging [3–5]. For this reason, with the increasing demand for electricity by the people in our country year by year, the requirements for the distribution network are gradually increasing toward the demand and reserves. How to improve the disaster resistance of the distribution network is one aspect of its development and research. How to improving the ability to predict disaster damage in distribution network settings is also an aspect of its development research [6–9].
However, at present, the research on disaster prediction of distribution network facilities under the influence of wind, flood, and other bad weather at home and abroad is in its infancy. Due to the countless economic and property losses caused by inaccurate distribution network disaster loss prediction, relevant research scholars have to pay attention to it [10–14]. Among them, Shu and Wu [15, 16] conducted a powerful analysis on the disaster resistance and prediction of disasters for the existing power grid and carried out a mechanical model study on the transmission tower line, which greatly promoted the development of the power grid infrastructure. The research on mechanical properties has pointed out the direction for the research on disaster resistance and disaster prediction of distribution network [17]. Li et al. [18, 19], based on previous studies on the mechanics of transmission line towers, focused on simulating the stress of power grid lines under real typhoon conditions through numerical simulation software. On the basis of previous studies, this paper studies the situation of power network disaster under different influence factors, and establishes a targeted distribution network disaster impact assessment model. In addition to the research on the mechanical properties of the power grid, Huang et al. [20] also studied the damage mechanism of the distribution network in the face of typhoons, with the help of damage experiments, simulated the damage model of the distribution network under different typhoon conditions, and obtained a lot of experimental data; the relevant experiments have laid a solid theoretical foundation for the disaster loss prediction of the distribution network. At the same time, in order to obtain a clear distribution network disaster assessment index model, Wei et al. [11], on the basis of the above research, the operation state of electric poles under different typhoon weather conditions is simulated through experiments, and a model is established on this basis. A set of impact indicators and fault correction models are used to evaluate the steady state of power poles in distribution networks under different typhoon wind speeds.
With the progress and introduction of neural network and artificial intelligence technology, the research on the disaster damage prediction model of distribution network has been put on the agenda again. At present, many researchers at home and abroad have studied the application of neural network technology in the disaster prediction model of distribution network, including [21]. Using BP neural network to build the disaster prediction model of distribution network, some also use deep learning [22–27] and other neural network models. However, despite the research of the above researchers, it is found that using neural network technology to predict the disaster of distribution network has good results, and its prediction accuracy and prediction error are greatly improved compared with previous traditional research methods. However, due to insufficient theoretical depth and limited data, the prediction accuracy of distribution network disaster is still not as high as that of other fields. The disaster loss prediction of distribution network is still in its infancy.
Therefore, in view of the above problems, based on the previous research, this paper collects the disaster data of some distribution network lines in Guangdong Province, and improves and optimizes the related neural network model again. Source data fusion technology studied the disaster mechanism of distribution network with different neural network models under different activation function conditions, constructed a fuzzy evaluation function based on wind and waterlogging disasters, and realized the analysis of distribution network under the conditions of wind and waterlogging disasters and disaster forecast.
2. Research Methods and Basic Theories
2.1. Convolutional Neural Network
CNN is a deep learning algorithm for processing time series and has been widely used in power system. It is composed of convolution layer, down-sampling layer, full connection layer, and output layer, as shown in Figure 1.
The input layer of CNN is a set of two-dimensional data matrix, which is convoluted by convolution kernel and sample matrix, and a characteristic surface is obtained through activation function. The convolution calculation formula iswhere L is the number of network layers; is the output characteristic surface of convolution calculation; is the characteristic surface of L-1 layer; is the weight and offset of convolution kernel, respectively; and σ is the activation function.
The down-sampling layer after convolution is to filter the output of the convolution layer, remove nonimportant feature parameters, and extract secondary features through the sampling kernel k. The operation formula of the mean sampling method iswhere ⊗ is the sliding convolution process.
The full connection layer is connected with the output layer, which is a traditional multi-layer perceptron. Its function is to use features for classification. CNN’s supervised learning training process includes feedforward operation and feedback operation.
Feedforward operation: assuming that the network has been trained, that is, the parameter has converged to the optimal solution, the network can be used to predict the fault category. The process of prediction is a feedforward operation.
In formula (3), is input; is output; is convolution operation; and is the down-sampling layer.
Feedback operation: compare the result of the output layer with the data label, calculate the error between them, back propagate to each layer, and update the weight parameters.
2.2. Fuzzy Evaluation Theory
Fuzzy comprehensive evaluation is based on fuzzy mathematics theory and uses the principle of fuzzy linear transformation to analyze the relationship between different states and evaluation levels. On the basis of single-level fuzzy comprehensive evaluation, the multi-level fuzzy comprehensive evaluation method is more suitable for objects with higher complexity. For these objects with higher complexity, using this type of evaluation method can greatly improve the overall evaluation effect.
When solving practical evaluation problems, we take the commonly used secondary evaluation as an example, and the use of the above evaluation system is generally divided into the following six steps: Step 1: before constructing the evaluation system, we should first analyze the related problems, that is, analyze the damage of distribution network transmission circuit caused by wind and flood disasters, and preliminarily identify the main influencing factors, so as to build an index system that meets the problems. Step 2: then, the main risk factors of the above distribution network wind and flood disasters are studied in depth, and the above evaluation system is analyzed by using principal component analysis and other methods, so as to build a comment set. Step 3; Using analytic hierarchy process and other methods to analyze the proportion weight of relevant indicators. Step 4: build a fuzzy evaluation matrix for the relevant evaluation factors. Step 5: select the appropriate fuzzy operator to comprehensively evaluate the above evaluation factors. Step 6: get the evaluation results.
2.3. Interval Analytic Hierarchy Process
This method is a comprehensive analysis method that integrates interval algorithms and different AHP methods and can perform quantitative and AHP [28]. It has been widely used in various engineering fields. The main implementation steps of using this method to determine the indicator weight are as follows:(1)Establish interval number judgment matrix. The reciprocal 1–9 scaling rules shown in Table 1 are used to compare the indexes at the same level, and an interval number is used to represent the result of the comparison, where and . The interval judgment matrix can be formed according to the interval number, and Table 1 shows relevant industry specifications.(2)The weight vector is calculated from the interval judgment matrix. The interval judgment matrix is decomposed into and , and the maximum eigenvalues of the two judgment matrices and and the corresponding normalized eigenvectors and are calculated, respectively. Then, the weight vector is obtained as follows:where and .(3)Calculate the final weight of the index:(4)According to the matrix consistency test method of analytic hierarchy process, the consistency of interval matrix is measured.
3. Disaster Damage Assessment and Model Construction of Distribution Lines Based on Multi-Source Data Fusion
3.1. Disaster Analysis of Distribution Lines
For the distribution network facilities in the eastern coastal areas, there are many failures caused by severe weather such as typhoons and rainstorms. For a long time, the failure of transmission lines caused by natural disasters such as typhoons and rainstorms has not only caused great inconvenience to people’s lives, but also caused great inconvenience to people’s lives. Problems such as circuit tripping of power grid lines, collapse, and inclination of distribution tower lines also pose a huge threat to people’s lives and property safety. The typhoon disasters that cause distribution network lines are nothing more than the following three points: first, it is due to uncontrollable meteorological elements, which is also one of the main reasons for the failure of distribution network lines and one of the important reasons for damage. The second is the terrain factor. Because the layout of the distribution network is often located at the edge of the city, the environment is also mountainous and watery. Therefore, the terrain conditions are extremely complex, which often causes problems for the erection of the distribution network lines. It is extremely difficult. Under bad weather conditions, it is easy to increase the probability of causing a series of accidents. The third is the stiffness and strength of distribution network infrastructure. The design intensity of relevant infrastructure is often difficult to match the actual environment. When the design strength of the first-hand infrastructure is too high, it is easy to cause waste of materials, and when the design strength is too low, when encountering harsh conditions, this is more likely to cause various security accidents to occur.
3.2. Selection of Typhoon Disaster Assessment Indicators
The risk assessment of typhoon disasters is mainly to measure the degree of risk of disaster-causing factors to the disaster-bearing body, which is an important part of risk assessment. In daily life, we use the method of typhoon grade to display the intensity of typhoons, and the classification of typhoon grades is based on wind speed. It can be seen that wind speed is an important indicator for assessing the risk of typhoon disasters; in addition, typhoons can also cause heavy rain, and the rainstorm caused by typhoon has rainfall. The characteristics of high intensity and ferocious force may cause secondary disasters such as floods and landslides, which have brought great adverse effects on people’s lives. Therefore, the torrential rain and precipitation caused by typhoons are also important indicators for assessing the risk of typhoon disasters. To sum up the above analysis, in order to better provide an evaluation index system for the model constructed in this paper, this paper selects the daily maximum wind speed, maximum daily rainfall, and recorded storm surges in Guangdong Province in recent years as experimental data support, and based on this, a risk assessment index system for typhoon-causing factors in Guangdong Province was established.
3.3. Construction of Disaster Loss Prediction Model for Distribution Lines
Under the influence of severe weather such as typhoons and rainstorms, distribution network lines are often seriously damaged, and the resulting economic and property losses are also immeasurable. This paper collects the disaster data of distribution network lines under the actual typhoon weather conditions, and uses these data as the training data of the prediction model for the loss of distribution network lines under the wind and waterlogging weather conditions. From this theoretical basis, it can be seen that based on multivariate data the disaster loss of distribution network predicted by fusion technology and neural network model is a typical regression problem. Therefore, when building relevant models and analyzing problems, we also mainly consider the elements of regression analysis. Taking the fuzzy evaluation used in this paper as an example, it is suggested to collect the disaster data of relevant distribution network lines and preprocess the data, and then take the fuzzy evaluation method used in this paper as an example. These data are analyzed, the relevant fuzzy evaluation set is constructed, and then the appropriate membership function is selected to determine the output objects such as injection tripping and broken pole probability. The schematic diagram of the distribution network line disaster damage prediction model constructed is shown in Figure 2.
3.4. Construction of Fuzzy Evaluation Model for Wind and Flood Disasters in Distribution Network
3.4.1. Determination of Judgment Set
(1) Set of Evaluation Factors. Assuming that the number of primary indicators is m, the primary evaluation factor set is established, and the secondary evaluation factor set ui contains ni factors, expressed as . The factor set of wind and flood disaster assessment index system of distribution network is . These subsets are set as primary evaluation indicators and then subdivided to obtain secondary indicators: , .
(2) A Collection of Critical Comments. The state comment set corresponding to different factors can be established as . According to the daily maintenance needs of the distribution network, combined with the successful experience of experts and front-line operation and maintenance personnel, the operation status of the distribution network is divided into four different levels: “intact,” “good,” “poor,” and “damaged,” corresponding to , , , and in the comment set.
(3) Weight Set. The weight value can represent the relative importance of different factors in the same level. The fuzzy weight set on the factor set is expressed as , where .
3.4.2. Determination of Fuzzy Evaluation Matrix
(1) Relative Degree of Deterioration. If the distribution network line is abnormal, the actual value of one or more evaluation indicators must deviate from the normal range. In order to eliminate the differential influence caused by the index type, physical meaning, and quantitative unit, the deviation degree between the current evaluation index and the expected value can be expressed by the relative deterioration degree, and its value range is [0, 1].
The higher the value, the better the index, and the relative deterioration degree is
For the index whose value is smaller and better, its relative deterioration degree is
For intermediate index factors, the relative deterioration degree is
In the above formula, η is the relative deterioration degree of the evaluation index; xmax and xmin are limit values; [ xa, xb] is the best range.
(2) Membership Function. Membership function is used to describe fuzzy theory. In specific engineering examples, the commonly used distribution forms of membership function include ridge, triangle, trapezoid, and normal distribution. For ridge distribution, its characteristics are that the principal value interval is wide and excessively flat, and the shape is relatively simple. Therefore, the paper selects the ridge distribution membership function to reflect the differences of different evaluation objects on the same evaluation factor and constructs the membership function required for evaluation. The ridge membership function of indicators corresponding to each state level is shown in Figure 3.
When the relative deterioration degree of the index is η, at this time, the membership functions corresponding to the comment states ∼ are as follows:
3.5. Analysis of Correlation Degree between Wind and Flood Disasters Based on Fuzzy Evaluation
(1)Fuzzy set and fuzzy membership. Fuzzy is a method used to describe objective things that are difficult to define boundaries. Fuzzy set includes the set of all fuzzy objects. Assuming that U is a nonempty set, it is called universe. Any given element x in U, there is a fuzzy set A. Fuzzy set X on U is a mapping from u to [0, 1], μA is called the membership function of a, and μA(x) is called the membership degree of x to fuzzy set a, which is written as Among them, the point x0 with membership degree of 0.5 is called the transition point of fuzzy set a, which is the most fuzzy.(2)According to the disaster damage evaluation standard of the distribution network, we can divide the severity of the distribution network due to wind and waterlogging disasters into 4 situations as shown in Table 2: one is the slight impact that can be used without repair, the second is the moderate impact that needs to be repaired, the third is the serious impact that is difficult to repair, and the fourth is the extreme impact that cannot be repaired.
4. Experimental Results and Analysis
4.1. Selection of Experimental Data
In the prediction of wind and waterlogging disasters in the distribution network in this paper, the destructive power of wind and waterlogging disasters is basically the same, and wind disasters are more serious, especially typhoons in coastal areas. Therefore, when analyzing, it is necessary to ensure that the analysis results are reasonable. Based on the nature and representativeness, in order to simplify the analysis methods and data, only the simulation analysis of wind disasters is carried out.
Considering that the various types of data collected will be more or less missing, duplication, and other errors due to various human and other factors during the collection, this paper firstly analyzes the relevant data before conducting data analysis and model training. Perform min-max normalization processing, and the relevant disposal measures are shown in the following formula. Using this method can not only eliminate the influence of the dimension value of the data itself on model training, but also greatly shorten the model training time, increase the accuracy of model training rate, and improve the robustness of the model.
After the sample data are standardized, it can be used as the input of the model constructed in this paper, and the disaster characteristic data of the selected distribution network can be used as the model output.
In order to study the influence of model performance parameters and the comparison with the prediction results of other methods, the root mean square error (RMSE) and the mean absolute error (MAE) are selected as the evaluation indicators of model accuracy. The calculation formula iswhere xi is the predicted value and yi is the actual value.
4.2. Prediction Results of Distribution Network Wind and Flood Disaster Prediction Model Based on Multi-Source Data Fusion
As the activation function is one of the important components of all neural networks, the activation function injects nonlinear features into the neural network. If the activation function is not added, the whole neural network is just a combination of many linear changes, which can be reduced to each other in the continuous iterative process and cannot be trained and learned. Therefore, it is necessary to introduce the activation function [63]. Common activation functions are as follows:(1)Sigmoid function. This function is the most frequently used nonlinear function in machine learning, and its mathematical expression is as follows: The geometric image of sigmoid function and its derivatives is shown in Figure 4. It can be seen from the figure that the function range 0∼1 is symmetrical about the center of the point (0, 0.5), and the closer it is to the y-axis, the greater the slope. If the value of X is large, the result will be close to 1, and if it is small, it will be close to 0.(2)Tanh activation function. The expression of this function is The Tanh activation function and its derivatives are shown in Figure 5. It can be seen from the image that this function will compress the input value to between -1 and 1, and the image is centrosymmetric.(3)ReLU activation function. The expression of this function is
An image of the ReLU activation function and its derivatives is shown in Figure 6.
The ReLU (the rectified linear unit) function consists of two parts. When the input value is negative, zero is output, and when the input value is positive, the independent variable itself is output. Through continuous engineering practice, it can be seen that compared with Tanh and sigmoid, choosing the appropriate gradient descent algorithm, the convergence of ReLU is greatly improved, and the derivative of the activation function is easy to solve. In engineering practice, it is found that when training multi-layer neural network model, Tanh and sigmoid often have the problem of gradient disappearance, and because of the linear and unsaturated form of ReLU, it can effectively resolve the phenomenon of gradient disappearance. In addition, the function does not need to participate in exponential operation, but only has a linear relationship. In both forward and back propagation algorithms, the convergence speed is very high. Because when the value of ReLU activation function is negative, the function output value is zero, which can greatly reduce the dependence between variables and prevent over-fitting. The disadvantages of ReLU function cannot be ignored: first, its output is not zero centered. Second, if the input weight is negative, the neuron will die irreversibly. Usually, we set a small learning rate to avoid this.
In order to analyze the influence of different activation functions on the experimental accuracy, the RMSE is used as the evaluation standard of the above model, and the sigmoid function, the Tanh function, and the ReLU function are used as the excitation functions to conduct experiments. The resulting model prediction errors are shown in Figure 7.
From Figure 7, it can be found that the model RMSE changes with the number of iterations under different activation functions, when the model uses sigmoid as the activation function, the initial error value of the model is relatively large, close to 10, and with the increase of the number of iterations, the model convergence speed is also slow. It takes nearly 300 iterations for the model to approach convergence; when using the Tanh function as the activation function, due to the advantages of the Tanh function itself, the initial error value is small, and the overall convergence speed is also the fastest, but it is easy to fall into the local optimum value, resulting in a large final error value of the model, which is not conducive to the overall prediction accuracy of the model; when using ReLU as the activation function, because the function has less dependence on the model variables, its overall effect is better than the former two. In terms of model initial error and model convergence speed, it is better than other functions, the final convergence error value is much smaller than other functions, and the minimum value is close to 10-3. Therefore, combining the above different model activation functions for this model from the perspective of the test error, using the ReLU activation function can not only greatly reduce the model error, but also make the model converge faster. Therefore, in this paper, the ReLU function is selected as the activation function of the function.
4.3. Practical Application of the Model
Considering the relationship between the training error of the model and the number of neurons in the hidden layer, the ReLU function is selected as the excitation function of the experiment, and the number of nodes in the hidden layer is 60 for the experiment. The model constructed in this paper is used to simulate the relevant disaster data, and the final detection result is shown in Figure 8.
Figure 8 shows the prediction results of the number of trips, the number of broken wires, the number of downed poles, and the number of broken poles under the influence of typhoon, and the overall prediction results are basically consistent. This verifies the accuracy of the model constructed in this paper.
4.4. Influence of Hidden Layer Activation Function on Prediction Model Accuracy
At the same time, in order to reflect the superiority of this model over other models, after the model itself is tested, the proposed fuzzy evaluation-based convolutional neural network (FE-CNN) algorithm is compared with the conventional convolutional neural network algorithm and the commonly used convolutional neural network algorithm, conventional BP neural network model. And predict the number of trips of utility poles, which is the most serious and typical disaster-affected accident under typhoon weather. Under the same experimental conditions, the same data were collected for the experiments of the above three different models, and the model prediction accuracy was selected as the prediction evaluation index. The prediction results obtained are shown in Figure 9.
From the prediction accuracy of the three different models in Figure 9 for the number of trips under typhoon weather conditions, it can be found that for different models, the prediction accuracy is not only the same. Among them, for CNN model, because of its strong feature extraction ability, it is more impressed with the discovery of model trips during prediction, so it can quickly identify the trip problem, but at the same time, because it is too sensitive. As a result, its prediction accuracy is not high, and its highest accuracy is only 0.836, For BP neural network model, its feature extraction ability is poor, but compared with CNN model, it has better prediction accuracy, with the highest accuracy of 0.865, and its prediction accuracy is 0.039 higher than CNN model, exceeding 3.5% of CNN model. The FE-CNN model constructed by this paper obviously combines the advantages of the two and shows good prediction accuracy. Although the number of prediction iterations required is large, which affects the prediction time, the overall number of prediction iterations is controlled within 400, and its maximum accuracy is as high as 0.964, which is far higher than the other two traditional prediction models. Its prediction accuracy exceeds CNN model by 15.31% and BP prediction model by 11.45%. Therefore, based on the above performance, it can be found that although the FE-CNN prediction model constructed in this paper requires longer prediction time than the other two models, the overall prediction time is within the controllable range, and the prediction accuracy is higher, showing good prediction ability for wind and flood disasters.
5. Conclusion
In view of the serious damage of distribution network lines under severe weather conditions such as typhoon and rainstorm, based on the existing research and combined with the actual transmission situation of distribution network, this paper collected the measured disaster data of distribution network in relevant areas, improved the traditional disaster prediction model, and combined the multi-source data fusion technology and neural network modeling technology to build a new combined prediction model. With the help of fuzzy theory, the relevant fuzzy matrix is constructed, and the disaster degree indexes of different distribution networks are analyzed with this model to evaluate the disaster degree. The measured data verify that the disaster loss prediction model of distribution network constructed with this model can effectively implement the disaster loss prediction of distribution network, and the prediction results are compared with BP and ordinary CNN models. The comparison results show that the multi-source data fusion prediction model constructed in this paper has the advantages of smaller error, higher accuracy, and faster convergence speed than BP and ordinary CNN models. This example proves that the FE-CNN constructed in this paper has good prediction ability for wind and flood disasters. The relevant results can not only provide theoretical and practical support for the disaster prediction of distribution network, but also provide theoretical guidance for the relevant prediction models.
Data Availability
The dataset used in this paper is available from the corresponding author upon request.
Conflicts of Interest
The authors declared that they have no conflicts of interest regarding this work.
Acknowledgments
This work was supported by the Guangxi Key Laboratory of Intelligent Control and Operation and Maintenance of Power Equipment of Electric Power Research Institute of Guangxi Power Grid Co., Ltd., and has practical application in the project “Research and Application of Key Technologies for Disaster Prevention Early warning and Production Decision-Making in Guangxi Distribution Network” (GXKJXM20200012).