Short-term Wind Speed Forecast Based On Wavelet Process Neural Network

With the increase scale of wind power, large power fluctuations were caused bywind power centralized access would bring a greater impact to grid. When powerfluctuations become serious, it would affect the safety and stable operation of grid.Accurate wind speed forecast helped to reduce such disadvantages, and attracted alot of attention. In this paper, short-term wind speed forecast model based onwavelet process neural network was given, after time correlation, non-stationary,non-linear and chaotic characteristics of short-term wind speed time series signalwas analyzed. Focus on short-term wind speed forecast model of wavelet processneural network, which has space weighted aggregation and time’s decompositioncumulative operation, and correspond algorithms were designed and simulated.Results of simulation were satisfied.The main works were described as follows:1. The time correlation, non-stationary, non-linear and chaotic characteristics ofshort-term wind speed time sequence is analyzed. Autocorrelation and partialautocorrelation functions were used to determine wind’s time correlation andnon-stationary. Then mean, variance and skewness functions were used todetermine wind’s non-linear. Power spectrum method was used to determine itschaotic properties. The analysis provided the design basis for wavelet processneural network wind speed forecast model.2.3-layer wavelet process neural network short-term wind speed forecastmodel and parameter selection method were used to deal with the wind speed timeseries prediction. Process neuron was used to achieve the spatial aggregationoperation, and wavelet analysis achieved time resolution accumulate operations.Phase space reconstruction embedding dimension of short-term wind speed wasused to determine the input layer neurons. Wavelet analysis theory and empiricalformula was of the hidden layer neurons of the wavelet neuron network were usedto determine the hidden layer neurons. Short-term wind speed forecasting value asthe output of the network.3. Training and prediction process of short-term wind speed wavelet processneural network were studied. Gradient descent function was used to train the telescopic parameters and translation parameters of hidden layer and the connectionweights of wavelet process neural network. The evaluation of training of the neuralnetwork model and forecasting process was achieved by the model error evaluation.4. Measured wind speed data of the wind field was used for matlab simulationexperiments. The simulation of wind speed prediction algorithm compared withautoregressive moving average model, BP neural network model and wavelet neuralnetwork model. The results showed that this method had a good predictive effectand the universality of environmental change.