Research On Short-Term Large-Scale Wind Power Forecasting Methods Based On Spatio-Temporal Correlation

Large-scale wind power integration has brought serious challenges to the security and stability of power system operation.Wind power forecasting(WPF)is an indispensable tool to cope with uncertainties and to facilitate the accommodation of wind power into systems.The accuracy improvement of traditional forecasting methods is limited due to their inadequate input information.Thus,it is urgent to develop new theories and approaches to improve WPF by exploiting and making full use of the spatio-temporal correlations among wind farms over a wide geographical area.In this dissertation,firstly the analysis and processing methods for abnormal wind farm operation data are proposed to provide high-quality data for forecasting,and the spatio-temporal characteristics of large-scale wind power are analyzed.Then,in the spatio-temporal framework,a series of novel short-term WPF methods are proposed from different aspects for separate wind farm,clustered wind farms and regional cluster.The core principle of these methods is to reasonably use rather than to abuse the spatio-temporal information for forecasting.The main innovations of this dissertation are as follows:To deal with the abnormal wind farm operation data caused by wind curtailments,their basic characteristics are analyzed in terms of time series,probability distribution and wind speed-power scatters.Two combined approaches are proposed to eliminate the abnormal operation data,i.e.,the combination of quartiles and k-means clustering and the combination of the quartiles and density-based clustering.The problem brought about by determining the number of abnormal data clusters when using k-means is solved by a novel"re-cluster" method.The parameter sensitivity of the density-based clustering is analyzed and the parameter setting rule is provided.Results show that both of the two combined approaches can effectively identify and eliminate outliers and significantly improve the data quality.They are efficient and can often be used for both wind turbines and wind farms,and consequently provide high-quality data for forecasting.To forecast the wind power of a separate wind farm,a new concept of backward temporal correlation of time series is proposed on the basis of traditional forward temporal correlation of time series.The statistical characteristics of the backward time series are derived and analyzed,according to which a backward wind power forecasting model is built.Then a bidirectional mechanism is established by considering the difference characteristics between the forecasting results of forward and backward temporal correlation.A case study shows that the WPF model based on bidirectional mechanism can reduce the probability of large forecasting errors at extreme points of wind power time series,and therefore can improve the WPF accuracy.The backward and bidirectional mechanisms have provided a new paradigm and framework for WPF,which could be used as a reference when developing other advanced forecasting methods.To forecast the wind power of wind farms in a large-scale regional cluster,the multivariate vector autoregressive(VAR)model is adopted as a basic forecasting model to characterize the spatio-temporal correlation among wind farms.Based on the sparse modeling theory,a sparsity-controlled VAR(SC-VAR)model that allows adjusting the sparse correlation structure of the forecasting model is established,by transforming the VAR model into a constrained mixed integer nonlinear programming.To solve the implementation problem caused by the complicated constraints and parameters of the SC-VAR,and to incorporate necessary a priori knowledge into the forecasting,a correlation-constrained SC-VAR(CCSC-VAR)model is proposed by constructing a sparsity-controlled matrix derived from spatial correlation information of wind farms.The CCSC-VAR can meet the special restrictions or requirements on the sparse correlation structure in practical use.It serves as a new solution for the controllability of a WPF method using spatio-temporal correlation.To forecast the wind power of wind farms in a large-scale regional cluster and to improve the adaptability and efficiency of the spatio-temporal forecasting model,an online large-scale spatio-temporal WPF model solved by sparse recursive estimation algorithm is proposed.The online forecasting model is adjusted and updated in real-time by using the latest wind power measurements.The relationship between the spatial correlation of wind farms and the sparse structure of online forecasting model is analyzed.The operating mechanism of the online sparse forecasting model is revealed.Results indicate that by using the proposed forecasting model,the most valuable spatio-temporal correlation information can be dynamically optimized for the target wind farm.The forecasting model has high adaptability since it can reflect the characteristics of the varying forecasting environment.Moreover,the forecasting efficiency is improved by the online training mode rather than offline batch training mode.To forecast the total wind power output of a large-scale regional cluster by upscaling,the relationships between fluctuation characteristics of cluster wind power and the spatial correlation of wind farms as well as the spatial distribution of wind farms are quantitatively analyzed.The impacts of input feature variables on the wind power output of a cluster is explained.Then,a numerical weather prediction(NWP)feature selection approach based on minimal-redundancy-maximal-relevance strategy is presented.The selected NWP features are directly used for upscaling WPF for the cluster,which reduces the uncertainties and complexity of traditional indirect upscaling that needs to forecast individual wind farm's output.Furthermore,once the featured wind farms are selected,the data from non-featured wind farms are no longer required,so it can both guarantee the forecasting accuracy and significantly reduce the costs on data resource and computation.