Research Of Spatiotemporal Distribution Prediction Method Of Power Transmission Fault Events Based On Big Data Analysis

To reduce the impacts of the potential fault events on power supplies,the successful monitoring and precaution of fault events are indispensable.Owing to the limited manpower and resources available for the monitoring and precaution of fault events in real situations,it is hard to realize the whole time observation for every part of the power transmission system.Thus,the limited manpower and resources need to be applied in the locations or within the periods with a higher outage risk.There are overt temporal and spatial regularities of the distribution of fault events in systems,hence,the prediction of the spatiotemporal distribution of fault events can be generated based on these regularities.Once achieved,more reaction and preparation time can be available for handling fault events,and the more effective planning and arrangement of countermeasures for the different fault event types can also be conducted.To sum up,the prediction of the spatiotemporal distribution of fault events will benefit the security and reliability of power transmission systems.In order to realize the long-term predictions whose lead time can be one hour,one day or even one week,this thesis builds a prediction model for the spatiotemporal distribution of fault events.The records of fault events including the environmental condition information are utilized as inputs,and the association rules between these environmental condition information and fault events are explored via the proposed prediction model.In light of these rules,this model can predict the locations and periods with a higher risk in the system through the forecasted environmental condition information,which is the so-called spatiotemporal distribution of fault events.The main contributions of this thesis can be concluded as follows:1)To solve the problem that the rarely occurred environmental elements and the periods with fewer fault events are commonly excluded from the analyses in the previous association rule mining models,this thesis develops the association rule mining with rare elements and time series(ARMret)model.This model can extract the high impact low probability(HILP)variables precisely from the input data so that the predictions can be enhanced.The threshold selection methods are designed for five significance measurements according to the distribution of fault events within different periods.Thus,those periods with fewer fault events can be evaluated when the rare variables are mined from the input data.Then,the score calculation methods in five significance measurements are revised to conditional forms according to the distribution of rare environmental elements in each feature,and therefore,these rare environmental elements can be diagnosed when the HILP variables are further mined from the rare variables.Lastly,the advantages of this model are verified via an empirical case study based on the real fault records in a central province.2)There are three difficulties in the previous relative weights calculation methods of the input data: the measurement approaches they choose are crude,the imbalanced distribution of fault events within different periods is not considered,and their parameters keep unchanged during the whole prediction.To solve them,this thesis designs the dynamic association rule mining with rare elements and time series(DARMret)model,which can measure the relative weight of input data more accurately,and can make the predictions more precise.Firstly,the two-fold risk index(TFRI)model is established to measure the weights.In the TFRI model,the different correlations between environmental elements and fault events are deployed to calculate the weights,the dissimilar risks of fault events within different periods are also incorporated.Ergo,the TFRI model can assess the relative weights of input data more comprehensively via these two perspectives.Secondly,the proposed self-adaption dynamic adjustment model can automatically modify the parameters of the TFRI model according to the comparison between the previous prediction and the real records,and thus the relative weight calculations,as well as the next prediction,can be improved.Finally,the escalation of the prediction performance by the DARMret model can be validated through an empirical case based on the real fault records from a southern city.3)The previous prediction methods generally handle the different types of input data via the same prediction model,and hence,the properties of each type of input data are not considered.Generally speaking,the input data can be classified into discrete and continuous features.To incorporate the different properties and to further ameliorate the predictions,this thesis proposes the fuzzy dynamic conditional association rule mining(FDCARM)model.Firstly,this model combines the DARMret model and the fuzzy inference system based on ensemble learning,then the discrete and continuous features can be handled by different models to play their roles.Next,in order to reduce the computational complexity and to assist the mining of continuous features,the utilized FIs is revised in two aspects: one is the deployment of the hierarchical structure,and the other is the deployment of the probability fuzzy risks when establishing the membership functions.Ultimately,the better predictions achieved by the FDCARM model are proved through an empirical case based on the real fault records from a southern province,and the characteristics and applicable scenarios of this model are discussed and concluded by comparing with other prediction models.4)For the purpose of enhancing the availability and feasibility of the proposed models when applied in real scenarios,the uncertainties during real applications and the impacts of the prediction uncertainties of each fault cause on the overall prediction uncertainties are both assessed in this thesis.On the one hand,the impacts of three common uncertainties on the proposed model during real applications are assessed.An empirical study based on the real fault records from a central province demonstrates that when these uncertainties are considered,predictions made by this model during real applications will be more credible.Also,the design of strategies to reduce these uncertainties can be supported.On the other hand,besides the evaluation of the prediction uncertainties of each type of fault events,the impact of these prediction uncertainties on the overall prediction uncertainty is also analyzed in this thesis.Through an empirical case based on the real fault records from a southern province,the two-perspective diagnosis of the prediction performance of each type of fault events is more pinpoint,therefore,it can be more convenient to enhance the data collection and evaluation for the fault causes with a worse prediction performance.