1、 Abstract Risk visualization of power system under typhoon disaster has important scientific significance and engineering application value for power system disaster prevention and mitigation.In this paper,based on the multi-source hetero-generous information database,such as equipment operation inf
2、ormation,meteorological information and geographic information,the damage probability models of the main power tower network based on 6 machine learning algorithms including AdaBoost iteration algorithm,GBRT(Gradient Boost Regression Tree),RF(Random Forest),LR(Logistic Regression),SVR(Support Vector
3、 Regression)and CART(Classification and Regression Tree)are established by utilizing the historical damage data of the main power tower network under typhoon disaster in a coastal city,and the error and prediction accuracy of the models are compared.Then by combining with the historical data of typh
4、oon Mujigae,the predicted damage probability and risk value of each model are visualized with the geographic grid of 0.15 0.1.The predicted effectiveness of these models is compared,and the ideal model and display index is selected.Index Terms typhoon;power tower;risk assessment;machine learning;ris
5、k visualization I.INTRODUCTION S one of the extreme weather,the typhoon has a huge impact on the power system,which will not only cause damage to power equipment,but also cause large-scale power outages that will seriously affect the social production and life.Therefore,it is necessary to study the
6、reliability of power systems under typhoon disasters.At present,some researches at home and abroad have evaluated the reliability of the power system in extreme weather.Zhou analyzed the toughness of the distribution system under extreme weather.Based on the information entropy and multiple failure
7、models,the toughness index and assessment ideas of the distribution system were given1;Yin pointed out that most of the faults under typhoon are caused by the power tower and pole failure,thus a new evaluation algorithm based on Batts model was proposed2;Song analyzed the dynamic safety performance
8、of transmission lines by means of transmission line segmentation3;According to the complex structure of the distribution network,Huang proposed a health index based on the operating state of the system and an importance index based on user differences4;Liu regarded the reclosing success rate as a di
9、chotomous variable in risk assessment of the distribution network and use function to fit,thus improved the overall risk assessment speed5;Yang used the vulnerability curve to describe the relationship between wind speed and transmission line failure rate,but the 10-minute average wind speed is diff
10、icult to reflect the impact of the maximum gust6.The above research still has problems such as incomplete factors,strong subjectivity and complicated calculation.After assessing the power system under typhoon,it is necessary to visualize the risks in order to visually display the severity and geogra
11、phic location of the risk.At present,there are few applications of visualization technology in the early warning of power system risks in China.In 2008,Fujian Electric Power Company launched the comprehensive disaster prevention and mitigation system for power grid7,using the GIS(Global Information
12、System)to dynamically monitor and display the actual path,predicted path,impact range and wind level of the typhoon;Hunan Electric Power Company built a mountain fire prevention system for transmission lines,and realized the visualization and alarm of the mountain fire point through the prediction t
13、echnology and satellite infrared monitoring data8.In general,visualization technology has been applied in domestic engineering practice,but it is mainly used in emergency response after disasters,and it is still relatively rare in early warning research.A space GLMM(Generalized Linear Mixed Model)wa
14、s proposed by Liu to predict the number of power outages in the target area under hurricane and ice storm.Risk Visualization of Power Tower under Typhoon Disaster Based on Multi-source Heterogeneous Information Yong Huang1,Ruizeng Wei1,Zhuangling Zhang2,Tong Wang1,Shiwen Yu3,Hao Geng3,Hui Hou3 1.Gua
15、ngdong Power GRID Co.,LTD Electric Power Research Institute,Guangzhou 510000,China;2.Guangdong Power GRID Co.,LTD Center for Emergency Repair,Guangzhou 510000,China;3.School of Automation,Wuhan University of Technology,Wuhan 430070,China.A 2018 China International Conference on Electricity Distribut
16、ion Tianjin,17-19 Sep.2018CICED2018 Paper No.201805280000178 Page1/6 906Finally,the visual technology was used to show the number of predicted and actual power outages in each zip code area of the target area9;A.F.Mensah proposed a flexible evaluation framework,and used visualization technology to s
17、how the proportion of users power outages in each geographic grid in the target area10;Seung-Ryong Han dealt with weather conditions,environmental conditions and used statistical models to assess the risk within geographic grids11.It can be seen that foreign early warning research has applied visual
18、ization technology more generally,but there were also problems such as insufficient factors and strong subjectivity.This paper comprehensively considers the multi-dimensional factors such as equipment operation information,meteorological information and geographic information,and uses a variety of m
19、achine learning algorithms to compare and analyze them.A more objective damage probability and risk value prediction model of the power tower is established,and the visualization is realized.II.THE OVERALL STRUCTURE OF THE RESEARCH The overall structure of this research is shown in Fig.1.It is mainl
20、y divided into three levels:data level,artificial intelligence level,and visualization processing level.Typhoon informationPower tower informationGeographic informationSpacial multi-source informationData levelMachine learning algorithmsArtificial intelligence levelTraining set,test setPower tower d
21、amage risk modelActual dataPredict the Power tower damage riskStage 1 Stage 1Initially select the most ideal modelMulti-index comparisonVisualization processing levelPrediction resultForecast result imageActual damage dataActual damage imageCompare intuitivelyThe most ideal modelThe most ideal displ
22、ay indicatorTyphoon monitoring stationsPower departmentsMicrometeorological monitoring devices Geographical surveying and mapping departmentsSVRLRRF Data collectionAlgorithm surpportDisaster prevention,mitigation decisionsTheory basisPractical guidanceFig.1.The overall structure of this study The fi
23、rst level and the second level mainly use the machine learning algorithm to establish the power tower damage risk prediction model and compare them.The third level is mainly based on multi-source heterogeneous information data from the target area,visualizing and comparing the forecast results intui
24、tively.III.MODELING AND COMPARISON OF MACHINE LEARNING ALGORITHM A.Data level Based on the information collected by typhoon monitoring stations,micrometeorological monitoring devices,power departments,and geographical surveying and mapping departments,etc.,a spatial multi-source heterogeneous inform
25、ation database is established,consisting of attributes and labels.The attributes include the equipment operation information including DW(Design Wind)and T(Time),weather information MG(Maximum Gust),geographic information including AL(Altitude),AS(Aspect),SL(Slope),SP(Slope Position),US(Underlay Sur
26、face),R(Roughness),etc.The label is defined as a dichotomous variable y(y=1 indicates the damage situation,and y=0 indicates the undamaged situation).B.Artificial intelligence level As the core of artificial intelligence,machine learning algorithms enable computers to learn knowledge from existing d
27、ata and summarize laws,and use these knowledge and laws for prediction and decision making.The current machine learning algorithms mainly include AdaBoost12,GBRT13,RF14,LR15,SVR16,CART17 and so on.Based on historical damage data,this paper uses the above 6 machine learning algorithms to build a powe
28、r tower damage probability prediction model and compare them,the process is shown in Fig.2.Spacial multi-source hetero-geneous information databaseTraining setModeling with PythonTest set:m data in total,the category label is y kk=0Predict damage probability p k|p k-y k|0.5?Correct+=1Correct+=0k m-1
29、?Output 4 indicators:Correct,MSE,MAE and R2Comparing the four indicators of each model,initially select the most ideal modelk+=1 Fig.2.Modeling and comparison process of the power tower damage probability Where,k denotes the serial number of the test set data;m is 2018 China International Conference
30、 on Electricity Distribution Tianjin,17-19 Sep.2018CICED2018 Paper No.201805280000178 Page2/6 907 the total number of test set data;y k(k=0,1,2,.,m-1)denotes the label(y k=1 indicates the damage situation,and y k=0 indicates the undamaged situation);p k is the predicted damage probability for the da
31、ta k;Correct denotes the defined prediction accuracy.Finally,indicators such as Correct,MSE(Mean Squared Error),MAE(Mean Absolute Error),R2(R Square),etc.are obtained for each model.C.Visualization level Visualization of the damage probability or risk value will make the decision-making process in d
32、isaster prevention and mitigation of the power sector intuitive,simple and specific,which is of great significance for improving efficiency,saving costs,and reducing losses.The power tower damage probability and risk value visualization process proposed in this paper is shown in Fig.3.Typhoon gust l
33、ayer The power tower coordinate layer Geographic information layern grids of a certain sizeMaximum gust in grid i:MG i Power tower quantity in grid i:N i Average geographic information in grid i:AL i,AS i,SL i,SP i,US i,R i Average or median design wind in grid i:DW iAttribute table of grid iAverage
34、 or median operating time in grid iDamage probability of grid i:p i Risk value of grid iZ score rendering Z score renderingPredicted image of r Predicted image of p Actual damage imagen grids n gridsSimilarity comparisonThe most ideal model and display indexN iModels built by Python Fig.3.Visualizat
35、ion process of the damage probability and risk value As shown in Fig.3,using ArcGIS to extract the MG i,N i,and average geographic information AL i,AS i,SL i,SP i,U i,R i in grid i,where i(i=0,1,2,n)denotes the the serial number of a grid and n denotes the total number of grids.With average power to
36、wer design wind DW i and average operation time T i,the model can predict the damage probability and risk value of grid i.IV.CASE STUDY In recent years,the typhoon“Rammasun”,“Mujigae”,“Hato”and etc.caused great impact on the power grid of Chinas coastal areas18-20,and the power tower were badly dama
37、ged.Based on the historical data of the typhoon Mujigae,this paper predicts the power tower damage probability and risk value in 0.15 0.1 grid,and further visualizes the damage probability and risk value.The predicted results of the 6 models are displayed and compared to select the most ideal model
38、and display index.A.Modeling and Comparing According to the modeling and contrast process shown in Fig.1,this paper selects the spatial multi-source heterogeneous information of the damaged main network power tower under typhoon disasters in a coastal city grid of China.6 machine learning algorithms
39、 are selected and evaluated by cross validation method.TABLE I and Fig.4 show the MSE,MAE and R2 of these models and give the Correct on the test set.TABLE I THE EVALUATION RESULTS OF 6 MODELS Model MSE MAE R2 Correct AdaBoost 0.071 0.210 0.714 0.986 GBRT 0.014 0.031 0.945 0.986 RF 0.015 0.029 0.940
40、 0.986 LR 0.014 0.014 0.944 0.986 SVR 0.045 0.160 0.820 0.944 CART 0.052 0.116 0.792 0.937 Fig.4.The evaluation results of 6 models Combining these 4 indicators,it is preliminarily determined that LR and SVR are the best models.However,this is only the performance on the test set,and further researc
41、h on generalization performance is needed to determine the most ideal model.Therefore,this paper uses the meteorological information under the typhoon“Mujigae”in a city,and then combines the visualization technology to select the most ideal model.B.Damage and Risk Visualization The average value of
42、geographic information is extracted with 0.1 0.1 grid,and the median 35m/s of the design wind speed of the main network power tower is taken as the DW i of the power tower in each grid.The median 15 years of the operating time is taken as the T i of the power tower in each grid.2018 China Internatio
43、nal Conference on Electricity Distribution Tianjin,17-19 Sep.2018CICED2018 Paper No.201805280000178 Page3/6 908 Then the damage probability and risk can be visualized according to Fig.3.The monitoring stations are shown in Fig.5.The maximum gust map of the“Mujigae”is shown in Fig.6.and the distribut
44、ion of the power tower is shown in Fig.7.Fig.5.The monitoring stations Fig.6.The maximum gust map of the“Mujigae”Fig.7.The distribution of the power tower The predicted damage probability and risk value of AdaBoost,GBRT,RF,LR,SVR and CART are shown in Fig.8 Fig.19.Fig.20 shows the actual damage of t
45、he main network power tower under typhoon“Mujigae”.Fig.8.Predicted damage probability by AdaBoost Fig.9.Predicted risk by AdaBoost Fig.10.Predicted damage probability by GBRT Fig.11.Predicted risk by GBRT Fig.12.Predicted damage probability by RF 2018 China International Conference on Electricity Di
46、stribution Tianjin,17-19 Sep.2018CICED2018 Paper No.201805280000178 Page4/6 909 Fig.13.Predicted risk by RF Fig.14.Predicted damage probability by LR Fig.15.Predicted risk by LR Fig.16.Predicted damage probability by SVR Fig.17.Predicted risk by SVR Fig.18.Predicted damage probability by CART Fig.19
47、.Predicted risk by CART Fig.20.The actual damage of the main network power tower According to Fig.8 Fig.19 and Fig.20,we can see that:1)As shown in Fig.10,Fig.11,Fig.14,and Fig.15,the ideal model initially determined by TABLE I and Fig.4 is not good in visualization,so GBRT and LR are not ideal mode
48、ls.2)As shown in Fig.9,Fig.13,and Fig.17,it can be seen that AdaBoost,RF and SVR can identify the most dangerous areas,and thus the three models are the most ideal model.3)Comparing Fig.8,Fig.10,Fig.12,Fig.14,Fig.14,Fig.16,Fig.18 and Fig.20,we can see that the amount of the power tower in the in the
49、 grid with low damage probability may be large,thus the actual damage is more serious.Therefore,the damage probability P should not be used as a display index.4)Any method to predict the damage accurately must consider the amount of the power tower.V.CONCLUSION This paper proposes a method for risk
50、assessment of tower damage under typhoon disaster by using machine learning 2018 China International Conference on Electricity Distribution Tianjin,17-19 Sep.2018CICED2018 Paper No.201805280000178 Page5/6 910 algorithm combined with visualization technology,and divides it into three levels:data leve
