Research And Application Of Short-Term Wind Speed Hybrid Prediction Model Based On Deep Learning

With the increasing capacity of the large-scale wind power grid,the intermittence and randomness of the wind speed will directly lead to fluctuations of wind power so as to create challenges for the power system to maintain dynamic balance.To solve these problems,it is an effective way to accurately predict the wind speed,which will also improve the stability of the system’s operation.Accurate wind speed prediction is a complex time series prediction problem.Compared with traditional methods,deep learning models have the ability to learn complex feature representations and deal with long-term dependencies in time series,and it is necessary to apply deep learning to complex time series problems.In this paper,from the perspective of feature extraction,we propose two novel hybrid prediction models for multi-step forward short-term wind speed prediction by utilizing the advantages of different deep network models in the modeling process of wind speed time series prediction.For univariate wind speed data,the VMD-LA-TCN model is proposed in this paper.The model extracted features from the original wind speed data with Variational Mode Decomposition(VMD),designed an LSTM Autoencoder(LA)to learn the feature transformation from the original sequence to the feature extraction sequence,and finally input the obtained feature values into a Temporal Convolutional Network(TCN)for multi-step forward wind speed prediction.To test the model’s performance,two sets of comparison experiments were conducted based on the selected wind speed datasets of three different heights of wind.As the results of the experiment show,compared with other eight models,the VMD-LA-TCN model has effectively improved the accuracy of univariate wind speed series prediction.For multivariate wind speed data,the SDAE-Seq2Seq model is designed and explained in this paper.The model first preprocessed the data for multiple variables,compared the correlation between different variables with the Pearson correlation test,and screened the variables with strong correlation.After data normalization,the robust features in the data were automatically mined and extracted from the large number of correlated variables with a Stacked Denoising Auto Encoder(SDAE).The extracted features were input into the Seq2Seq network structure for wind speed prediction.Eventually,it produced multi-step wind speed prediction values.To evaluate the performance of the model,three sets of experiments for the model’s validity were conducted based on two wind speed datasets with two periods of different seasons.The results of the experiment show that SDAE-Seq2Seq effectively mixes the two modules.The model outperforms other single and hybrid models in terms of prediction performance on both datasets.Based on the above two models,this paper designs and develops a short-term wind speed prediction system,which is mainly built with several techniques such as Python,Bootstrap,Javascript,datart and MYSQL.The system has a simple interface with functions that can predict short-term wind speed,which provides guiding data for wind speed prediction in smart grid.