A Signal Decomposition Technique Based Hybrid Model For Wind Power Prediction

With the excessive consumption of fossil fuels,energy crisis and environmental pollution problems are becoming increasing severe.Looking for renewable and clean energy resources is attracting attention of people all over the world.As an important kind of renewable energy,wind energy has promising application prospects for its enormous reserved amount,pollution-free and low development costs.However,wind power is influenced by many factors,such as wind direction,air pressure and temperature.As a result,wind power has strong intermittency and instability.These characteristics affect the connection of electricity generated from wind energy into power grid and wind turbines operation,which restrict further development of wind energy.Accurate wind power prediction is an effective way to solve these problems.Traditional approaches are point prediction models,which provide a certain value of wind power.However,due to the strong randomness and volatility of wind energy,errors of point prediction always exist,which lead to point prediction results can hardly match real wind power values totally.Furthermore,this method cannot calculate the probability of these certain values.Luckily,interval prediction models,which are built on the basis of point prediction models,can give the fluctuation range of wind power under different confidences.This kind of prediction results can describe the fluctuation trend of wind power better and enable power grid staff make better scheduling decision.Therefore,this paper proposes a hybrid wind power interval prediction model based on signal decomposition technique.First,considering that real wind power often contains many intrinsic mode functions(IMFs)with different characteristics,which leads to strong volatility,this model applies Ensemble Empirical Mode Decomposition(EEMD)to process wind power so as to extract different intrinsic mode components.This procedure transforms unstable raw data into a series of relatively stable sub-series.Second,the Least Square Support Vector Machine(LSSVM)model,which has strong generalization ability and high training efficiency,is used to predict all the sub-series.To improve prediction accuracy of LSSVM,there are two tasks,namely finding out input variables with strong correlation connection to output variable,and optimizing model parameters to improve training effect.Unlike traditional methods which confirm input number of model via personal experience,this model applies Partial Autocorrelation Function(PACF)to calculate correlation among different data points of time series,and determine number of inputs as the lag number.To optimize model parameters,this hybrid model first applies the Moth-flame Optimization(MFO)algorithm with strong global optimization ability to carry out a certain algebraic training on the model.And the Grid Search(GS)algorithm with strong local optimization ability is then used to optimize model parameters on the basis of optimization results of MFO to exclude influence of MFO's slow searching speed at the area near the global finest solution.After finishing point prediction,Weibull Distribution function is used to estimate probability density distribution of point prediction errors,and then calculate fluctuation range of these errors under different confidence levels.Then certain value of point prediction is added to obtain interval prediction result of sub-series.Finally,the interval prediction results of all the sub-series are superimposed as the final wind power prediction result.To test the prediction performance of the proposed model,EEMD-LSSVM and three contrast models are used to predict wind power data.Simulation results demonstrate that EEMD-LSSVM has higher prediction accuracy and more reliable prediction interval than contrast models.