Research On Complexity And Pattern Classification Of Natural Wind-Field Time Series

Time series refer to the sequence of the same statistical index values in the order of their occurrence time.It is the main external manifestation of the working or running status of many complex systems and contains rich information of system structure and dynamic evolution law.Many wind engineering applications,such as wind energy prediction,wind-resistant design and optimization of large buildings,and the diffusion of air pollutants,require accurate knowledge of the flow and evolution characteristics of the wind farm.However,in many practical applications,the time series of wind speed and wind direction recording natural wind field show complicated nonlinear and non-stationary features,which brings great difficulties and challenges to the wind engineering application.Due to the imperfect model and theory,there are some limitations in the conventional physics modeling.Therefore,this thesis intends to reveal the law of flow of complex gas flow field in depth from the data-driven point of view,focusing on the complexity and pattern classification of time series of natural wind farms and providing important theoretical and technical support for the application of wind engineering.The main work of this thesis is as follows:Firstly,in this thesis,aiming at the shortcomings of two kinds of mainstream methods(time domain method and frequency domain method)of time series,a new method of time series quantitative analysis of two-dimensional visual state diagram is proposed.The method maps a one-dimensional time series into a two-dimensional visual state diagram,and on the basis of which,defines a plurality of indicators for quantitative analysis to realize the quantitative analysis of complex time series.This method is applied to wind farm time series analysis in two different environments.The test results show that the complexity of indoor wind speed time series is less than that of outdoor wind speed time series in general.Visual rate,the degree of certainty and other quantitative indicators can better distinguish between indoor and outdoor wind speed signal.From the perspective of time series pattern,this method can effectively and accurately characterize the degree of difference between different time series,thus providing a new idea for the similarity measurement of complex signals.Secondly,aiming at a kind of time series signal that the amplitude shows jumping change law,taking into account the correlation degree and direction difference of signal sign pattern,this thesis represent a novel directed weighted complex network construction method based on time series symbolic pattern representation combined with sliding window technique.The proposed method firstly implements symbolic procession according to the equal probability segment division and then combines with the sliding window technique to determine the symbolic patterns at different time as nodes of the network.Next,the transition frequency and direction of symbolic patterns are set as the weights and direction of the network edges,thus establishing the directed weighted complex network of the analyzed time series.Test results using the Logistic system with different parameter settings show that the network topologies of our proposed method are more concise and intuitive than those of the classical visibility graph method.Finally,the proposed technique is applied to investigate the natural wind field signals collected from regular arrangement positions.The corresponding results of network characteristics can more accurately predict the spatial deployment relationship of nine 2D ultrasonic anemometers,while the visibility graph method has no useful cues.Thirdly,the time series classification method of near-surface wind field combining convolutional neural networks(CNN)and support vector machines(SVM)is proposed,which is compared with the traditional classification methods: search and find of density peaks clustering algorithm and C-means fuzzy clustering.The traditional method is effective when dealing with small data sets,but it relies too much on the initial value setting and it is difficult to meet the needs of big data mining for large data sets and samples with high data dimensions.The unsupervised classification results of the CNN + SVM model used in this thesis are obviously superior to the traditional classification methods,and the error rate is reduced to 12.6%.