Study And Application Of Artificial Intelligence Based Multi-step Time Series Forecast

Single-valued time series forecast always contains unavoidable uncertainties,which results in unstable model performance.Especially the multi-step time series forecast with complex variation characteristics,the uncertainty-caused forecasting error will be rapidly enlarged along with the forecasting step increases.This is a quite difficult and urgent problem in actual applications.Combining different single-valued methods can reduce the forecasting uncertainty and improve the multi-step forecasting accuracy,to a certain degree.Constructing the multi-step forecasting model for the datasets which have complex variation characteristics,this thesis applies a de-noising process,based on empirical mode decomposition(EMD)filtering,to the original data series.Two general forms of EMD-based multi-step forecasting model are given: decomposition-filteringforecasting-reconstruction and decomposition-filtering-reconstruction-forecasting are adopted.Combining with outlier detection method,artificial neural networks,and different multi-step forecasting strategies,this thesis develops a sort of filtering-based multi-step time series forecasting methods.Considering the sub-question forecast,the random optimal hidden parameters are generalized as global optimal hidden parameters of enhanced increased extreme learning machine;based on it,this thesis proposes a parameter estimation method of single-layer feed-forward network,named as CS-IELM,and theoretically proves the convergence of this method.It is from the study of these that solves the intelligent selection of the optimal hidden parameters of a singlehidden-layer feed-forward neural network.Considering the multi-step forecast with a condition of delay data acquisition(DDA),this thesis introduces the fuzzy method into combination forecast to select the single-valued models,in order to solves the problems of single-valued model selection and weighted parameters optimization.All the single-valued forecasts are gathered as a domain,then a standard model on this domain is defined and the single-valued models are filtered by a generalized maximum subordination principle.After that,a maximum k subordination principle induced ordered weighted averaging operator is developed,denoted as k-IOWA operator,and the basic properties of k-IOWA operator are also proved.Combining with Cuckoo Search optimization algorithm,this thesis constructs a multi-step combination model based on self-adaptive optimized fuzzy system,denoted as CS-FS-F-E model.Considering the multi-step forecast with a condition of on-line data acquisition(OLDA),this thesis designs an optimized combination model,which hybridizes grey correlation degree and Markov state transition.It is from the study of these that solves the transition of observed information,and the related single-valued model selection and weighted parameters optimization.Specifically,the grey correlation degree is regarded as the characteristic index of combination forecast,and the membership matrix is defined of which each single-valued model transits to different system states by a Markov probability transition process.The single-valued models are filtered by a developed maximum k subordination principle.After that,a maximum k subordination principle induced bio-factor ordered weighted averaging operator is proposed,denoted as k-BIOWA operator,and the basic properties of k-BIOWA operator are also proved.Then,a k-BIOWA operator based combination model is proposed,denoted as CSMarkov-F-E model.Applying the proposed multi-step models in this thesis to real ultra-short-term and short-term wind forecasts,all the data sets were collected from anemometer towers of different wind farms in China.Simulation results indicate that the proposed multi-step models outperform the traditional methods,especially the forecasting accuracy and model stability are significantly improved;this implies that the proposed multi-step models have good applicability.The simulation of CS-FS-WRF-E model,which applies to the operational wind forecast of real wind farms,considers the real data acquisition ways,forecasting timeliness,and operational precision requirement.This method is currently planned to apply to a real operational wind forecasting system.