Wind Speed Uncertainty Modeling In Wind Farm Based On Computational Intelligence

The development of wind power plays an important role in global efforts to address the fossil energy supply dilemma,climate warming and to maintain the sustainability of economic construction and human development.However,wind speed uncertainty fundamentally cause the randomness and volatility of wind power,which in turn causes the problems in planning and operation of wind farms and power grid safety,affecting the sustainable development of wind power generation.At present,the important role of the wind speed uncertainty models,including wind speed probability distribution fitting models and wind speed forecasting models used to cope with wind speed uncertainty,to wind power development has been widely recognized,and related modeling research has been widely emphasized.Based on the rapid development of computational intelligence,the advantages of computational intelligence methods over traditional methods in solving complex modeling problems have become increasingly apparent,and it is highly promising and urgent to improve the capability of wind speed probability distribution fitting and wind speed prediction models using advanced computational intelligence methods.Although the exploration of using computational intelligence methods in the above-mentioned two modeling studies has been carried out for some time,there are still many problems,such as:the differences in the mechanisms of different computational intelligence methods are not taken into account in the modeling,and the selection is not based on the actual situation;the application of computational intelligence methods is relatively single and stays in the replacement of traditional basic models,while applying it in some new modeling ideas such as cooperation modeling and systematic modeling is insufficiently studied;In addition,with the expanding of wind power development,the planning and operation of wind farms also put forward some new demands on wind speed uncertainty modeling,such as the fitting of bimodal wind speed data and the forecasting of wind speed intervals,etc.,and the current application of computational intelligence in the modeling related to these new demands is insufficient.Considering the limitations in existing research,this paper explores more effective ways to apply computational intelligence to improve the performance of wind speed probability distribution fitting models and wind speed prediction models by using two key techniques of computational intelligence,artificial neural networks and intelligente optimization algorithms,as the basis of modeling,and combining a variety of distribution functions and signal processing techniques,etc.,and also paying attention to new modeling demands to fill the existing research deficiencies.This paper consists of seven parts:Chapter 1 introduces the research background and purpose of this paper,the significance of the research,the current status of domestic and international research,the research ideas and main research contents,and the main innovations and shortcomings;Chapter 2 discusses the main modeling foundations of this study,which defines some key concepts to further clarify the research objectives and research contents of this paper and introduces two important categories of methods of computational intelligence to provide methodological support for the relevant models proposed in the subsequent chapters.Chapters 3 and 4 focus on the wind speed probability distribution fitting modeling problem,studying the intelligent approximation of Weibull distribution parameters and the intelligent mixture distribution in fitting the distribution of common unimodal wind speed data and bimodal wind speed data,respectively,to improve the accuracy of wind speed probability distribution fitting,which will benifit the wind farm planing efficiency.Chapters 5 and 6 focus on the modeling problem of wind speed forecasting,and propose the wind speed point forecasting model based on the intelligent multi-dimensional cooperation strategy and the wind speed interval forecasting model based on the intelligent modified scaling approach and two-output deep learning algorithm(gated recurrent unit)to improve the direction consistency,accuracy and stability of wind speed point forecasting values,and obtain better quality of wind speed forecasting intervals,which will benifit the wind farm operation.Chapter 7 is a summary of the whole paper and an outlook for future research.The main work and findings of this paper include.(1)Intelligent approximation of the Weibull parameters based on different intelligent optimization logics is proposed to better model the probability distribution fitting of unimodal wind speed data.First,the loss function of the intelligent approximation of Weibull parameters based on the minimum error is constructed;then,three intelligent approximation logics based on different global optimization search ways are introduced by embedding simulated annealing algorithm,genetic algorithm and grasshopper optimization algorithm;and three intelligent Weibull parameter approximation algorithms are finally proposed.Compared with similar existing studies,the introducing of multiple intelligent optimization logics in this paper can more fully exploit the advantages of computational intelligence algorithms and improve the fitting capability of the intelligent Weibull distribution in fitting unimodal wind speed data.Empirical analysis shows all three Weibull parameter intelligent approximation methods can make the Weibull distribution obtain a better fitting result than the Weibull distribution obtained by traditional methods,especially the GOA-[c,k](Weibull)algorithm can make the Weibull distribution obtain the optimal fitting ability for unimodal wind data.(2)The intelligent construction of the mixture distribution is studied to better fitting the probability distribution of bimodal wind speed data.Four probability distribution functions are used as component distributions,including three traditional distribution functions,namely Weibull distribution,Gamma distribution,and Lognormal distribution,and the Burr distribution,which is relatively newly introduced in the fitting of wind speed probability distribution.Ten specific mixture distributions are proposed:including the Weibull-Weibull distribution,the Gamma-Gamma distribution,the Lognormal-Lognormal distribution,the Burr-Burr distribution,and the Weibull-Gamma distribution,the Weibull-Lognormal distribution,the Weibull-Burr distribution,the Gamma-Lognormal distribution,the Gamma-Burr distribution and the Lornonnal-Burr distribution.Compared with the existing studies,this study incorporates more distribution functions as component distributions,including not only the traditional wind speed distribution functions but also the newly introduced wind speed distribution functions in recent years,and involves both the homogeneous mixture distributions and heterogeneous mixture distributions.More importantly,this paper explores the use of intelligent optimization algorithms to estimate the parameters of the mixture distribution.The results of the empirical analysis show that the intelligent mixture distributions can obtain better fitting result on bimodal wind speed data than the baseline Weibull distribution and the mixuture distributions constructed based on traditional parameter estimation methods;among them,the intelligent Weibull-Weibull distribution,intelligent Weibull-Gamma distribution,and intelligent Weibull-Burr distribution are proved to have the optimal result.(3)A wind speed point forecasting model based on the intelligent multi-dimensional cooperation strategy is proposed to provide accurate wind speed point prediction information.Since any single model always has limitations in wind speed point forecasting,cooperation strategies have been widely used for improvements.However,the traditionally used cooperaton strategies have their own drawbacks.This chapter proposes a more effective cooperaton strategy,the intelligent multi-dimensional cooperation strategy and applies it in wind speed forecasting modeling,which is an novel exploration of the use of computational intelligence algorithms.The proposed model first uses signal processing techniques to transform the original wind speed series to obtain several subseries,and uses both cubic exponential smoothing and backpropagation feedforward neural networks as predictors on all subseries for prediction,and then optimizes the best contribution weights of each predictor to the forecasting on the subseries by intelligent optimization algorithms.When integrating the forecasting results of subsequences,not only the results of each predictor are integrated,but also the respective contribution weights are integrated,and the final piont forecasting results are obtained in the multi-dimensional cooperation of different predictors.The empirical results show that the constructed model has better forecasting ability than various single forecasting models as well as forecasting models based on different traditional cooperation strategies.(4)A wind speed interval forecasting models based on the intelligent modified scaling approach and the two-output gated recurrent unit is proposed to provide more accurate and reliable wind speed interval predictions.Specifically,first,an interval construction approach that does not rely on strict assumptions,is with relaxed application conditions,and has a relatively low computational burden is developed,which also improves the shortcomings of the approach in its traditional framework where parameters are not easy to set and lack of adaptiveness to data.At the same time,a deep learning method,the gate recurrent unit,is used in the final forecasting step of wind speed interval forecasting,which can effectively exploit the long-time dependence in wind speed time series and exploit its good fault tolerance and generalization ability in sequence modeling to improve the effectiveness of wind speed interval forecasting.The results of the empirical study demonstrate the superiority of the proposed intelligent modified scaling approach compared with the traditional fixed-parameter scaling approaches and the single-parameter optimization-based scaling approach.It can be an efficient candidate solution to address the difficult wind speed interval constructing problem,and is highly competitive with the traditionally used error distribution approach and the paired Bootstrap approach.Introducing in the gate recurrent unit in the final forecasting step can overcome the limitations of shallow neural networks in learning long-time dependencies in wind speed time series and improve the overall quality of the prediction intervals.Through proper and innovative use of computational intelligence methods,this study improves the defects of traditional wind speed uncertainty modeling,improves the deficiencies of existing computational intelligence-based wind speed uncertainty modeling studies,and establishes several more effective,stable,and robust wind speed probability distribution fitting models and wind speed forecasting models,which can provide some effective management tools to the scientific planning and operation of wind power;It also enriches and develops some theoretical methods in probability distribution fitting,prediction and computational intelligence.Thus,it has certain practical and theoretical significance.The main innovations of this paper include the follows:(1)This paper reexamines the wind speed uncertainty modeling problem in wind farms.Based on the relevant theories and various specific techniques of computational intelligence,it provides innovative thinking on the existing modeling issues of wind speed probability distribution fitting and wind speed forecasting and proposes the unimodal wind speed probability distribution fitting model based on the intelligent approximation of Weibull parameters,the bimodal wind speed probability distribution fitting model based on the intelligent mixture distribution,the wind speed point forecasting model based on the intelligent multi-dimensional cooperation strategy and the wind speed interval forecasting model based on the intelligent modified scaling approach and two-output deep learning algorithm(gated recurrent unit),expanding the existing model reserve that provides decision support for wind farm planning and management.(2)This paper also enriches and develops the relevant theories of computational intelligence to a certain extent and expands its application scope.By innovatively reforming the original algorithm logic,principle,and mechanism of the original intelligent optimization algorithm,the specific optimal solution iteration rule is designed to modify the original single-objective artificial optimization algorithm for avoiding the invalid iterations,improving the search efficiency of optimal solutions,which ensures the realization of the full function of the intelligent optimization algorithm when it is used for multi-objective optimization problems with constraints.(3)This paper innovatively develops a series of specific techniques and strategies for wind speed probability distribution modeling and wind speed modeling.These include:the intelligent approximation methods for distribution parameters based on various optimization search logics,the intelligent modified scaling approach,etc.The shortcomings of this paper include:(1)This paper requires the use of a series of specific techniques,such as probability distribution functions,predictors,signal processing techniques,and intelligent optimization algorithms,in constructing various models.In fact,there exist a large number of relevant techniques to choose from,and based on the limitation of space,this paper does not consider incorporating all of them into the modeling,but only selects some representative ones that are considered to be more effective in a certain range or at the current stage and introduces them into the modeling,but the reasonableness of such selection is time-sensitive and subjective.(2)In this paper,when evaluating the constructed models,the focus is on the accuracy,stability,and reliability of the models in terms of fitting or prediction,and not enough attention is paid to the complexity of the models.(3)Some parameters in the constructed models still need to be set based on experience or by means of trial-and-error,and manual intervention is still required.