Short-term Wind Power Forecast Based On Mathematical Morphology And Local Predictor

During the decade, as a significant part of renewable energy, the application of the windpower has been increasing fast because of energy crisis and environmental issues. Obviously,with the rising penetration levels, the impact of wind power generation on the electric powersystem must be taken into account, mainly due to the intermittency and variability of wind.Large-scale wind power integration will increase the difficulties for power system planning, op-eration, and control, which makes wind power forecast a necessity. An accurate and reliableforecast of wind power is an effective and economically feasible solution to solving the opera-tional issues caused by the variability of wind, as well as adjusting short-term generation plansto match specific requirements. The forecasting accuracy of wind power makes a great differ-ence to the efficiency of unit commitment, the economic effects of dispatch, and the operationalcost.Efforts have been made to improve the accuracy of wind power forecast, and several state-of-the-art techniques have been identified. A brief review of literature is firstly presented in thispaper, and some of the popular forecasting models are introduced. This paper focuses on thestatistical models, which are data-driven and developed using historical data to provide timelypredictions. Therefore, some forecasting models are proposed based on the analysis and pre-processing of historical data.For the purpose of covering the absence of the selection of the training data for the forecastmodel, this paper proposes a pre-processing method to enhance the accuracy of wind powerforecast. Instead of using the whole dataset indifferently for training, the proposed methodonly uses the segments that share the same pattern(tendency) as the forecast segment, which iscomposed of the data collected from the forecast moment back to some previous moment. Suchselected segments are referred to as similar segments, which are to be used as the training setfor forecast. In order to reveal the tendency, components of high frequencies are filtered outby a k-OCCO filter, which involves a triangle structuring element(SE) specially designed. Toselect the similar segments according to the tendency, a weighted MMG operator with the sametriangle SE is used. Once the similar segments are obtained, they are used to train the LS-SVMmodel, which is the forecast engine used in this paper. In order to evaluate the impact of selectingsimilar segments as the training set to the accuracy, simulation studies have been carried out onwind power generation using a public database. As the results shown, using the similar segmentsas the training set of the forecast model improves the accuracy, and the forecast model performsmore stable than its original version. Moreover, although in this paper it is applied to improvethe accuracy of the LS-SVM model, it is noteworthy that the pre-processing method for selectinga proper training set can be applied to improve the accuracy of other forecast models.Since wind power is closely related to complicated meteorology and the situation of thesystem, and has the nature of high randomness and rapid fluctuation, it is relatively difficultto undertake accurate forecast directly on the original time series of wind power. Therefore,some researchers have proposed to decompose the time series into a set of constitutive seriesby wavelet transform(WT) or empirical mode decomposition(EMD). However, there are noguiding principles for how many components the time series should be decomposed into, andthe time series may be over-decomposed if the number of constitutive components is determinedimproperly. Furthermore, it is hard to interpret the physical meaning of each constitutive series.Therefore, this paper proposes a novel forecasting model based on mean trend detector(MTD) and mathematical morphology based local predictor(MMLP) for short-term wind pow-er forecast. Different from previous researches, the proposed MTD/MMLP model decomposesthe non-stationary time series of wind power generation, by MTD, into only two components—the mean trend and the stochastic component, and each of them is interpretable and has explicitphysical meaning. Subsequently, the p-step forecast is conducted for these two components sep-arately. According to the recent researches, local prediction methods would generally performbetter than global methods for time series prediction. The mean trend is forecasted on the basisof the least square support vector machine(LS-SVM) model; while the p-step forecast for thestochastic component is carried out by the MMLP, which involves performing morphologicaloperations employing a novel structuring element(SE) in the phase space. Finally, the forecastof wind power generation is achieved by combining the separate forecasts of two components.In order to evaluate the feasibility of the MTD/MMLP model, simulation studies have beencarried out on wind power generation using three different databases. The three databases notonly locate in different places, the data are sampled in different periods, as well. Moreover, thecomparison studies are conducted among five forecasting models, i.e., Per., RBFNN, LS-SVM,SVRLP, and the proposed MTD/MMLP model. As the results show, the MTD/MMLP modelworks better than Per., RBFNN, LS-SVM, and the SVRLP model, as it obtains the smallestNMAE and NRMSE values, which indicates the MTD/MMLP model provides more accurateforecasts and performs more stably.