Research On Deep Learning Algorithm For Short-term Wind Speed And Wind Power Prediction Of Wind Farm

Wind energy is abundant in nature and is one of the widely used sustainable clean energy.However,the randomness and volatility of wind power itself lead to the instability of power grid in the grid connection of wind power,which affects the effective utilization of wind energy.Therefore,accurate prediction of short-term wind speed and wind power of wind farm is an important way to improve wind energy utilization,strengthen power grid stability,and enhance wind power market competitiveness.Based on the summary of the existing methods,this paper studies the related problems in the prediction of short-term wind speed and wind power of wind farm by using deep learning networks and signal processing algorithms.Aiming at the randomness and fluctuation of wind speed,a short-term wind speed prediction model integrating time series decomposition and deep learning is studied.Firstly,the complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN)algorithm is used to decompose the wind speed for the first time to obtain a set of components,and then the wavelet packet decomposition(WPD)algorithm is applied to decompose the first component for the second time to further extract the relevant trend of the wind speed series.Finally,the gate recurrent unit(GRU)deep learning network is established for each component to predict,and the final prediction values are obtained by synthesizing the prediction results of each component.Experimental results indicate that the proposed model can predict the change of wind speed more accurately than other methods.Aiming at the uncertainty of wind speed prediction,a short-term wind speed interval prediction model based on adaptive interval depth learning is studied.The model uses GRU network to directly generate the upper and lower bounds of the prediction interval.Firstly,two interval width adjustment variables are introduced to adaptively construct high-quality interval upper and lower bound labels for GRU network.Then a two stage search strategy is designed to optimize the model parameters: two interval width adjustment variables are optimized by dynamic inertia weight particle swarm optimization algorithm,and the structural parameters of GRU network are optimized by RMSProp algorithm.The two stages are carried out alternately until a satisfactory prediction interval is obtained.Experimental results indicate that the model obtains a prediction interval with narrower width and higher coverage probability,compared with other methods.Aiming at the application of supervisory control and data acquisition(SCADA)data of wind farm in wind power prediction,a short-term wind power prediction model based on SCADA data deep learning is studied.Firstly,the original SCADA data is cleaned and resampled,and the density-based spatial clustering of applications with noise(DBSCAN)algorithm is used to remove the outlier data to obtain a higher quality data set.Then the WPD algorithm decomposes and reconstructs each feature data of SCADA data except wind power data to remove the noise information.The maximum information coefficient theory is used to select the features with high correlation with wind power that are sent to GRU network together with historical wind power data for training and prediction.Experimental results indicate that this method can learn the effective information in SCADA data and obtain higher wind power prediction accuracy than other methods.