Research On Short-term Wind Power Output Prediction Based On Deep Learning

As an important type of renewable energy,wind energy is suitable for large-scale development and utilization.However,due to its randomness and volatility,the output power of wind power generation is unstable.Large-scale wind power integration poses a serious challenge to the normal operation,dispatching and safe operation of electrical power systems.Wind power output prediction technology provides a possible solution to this problem.It can predict the output power of wind power plants in the future,thus providing a basis for reasonable scheduling and maintenance plan.It also is of high significance for improving the utilization rate of wind power.To improve the prediction accuracy of wind power output,this dissertation studies short-term wind power output prediction,integrates multivariate data based on deep learning technology.It uses machine learning and deep learning technology for data cleaning,designs an algorithm to construct an input data set of the prediction model.Moreover,TLW-LSTM and LW-CLSTM deep learning models are built,and the wind power output prediction experiment is conducted.Their prediction results are compared with those of traditional machine learning models such as the decision tree,random forest and support vector machine.In this way,the effectiveness of the wind power output prediction model based on deep learning is verified,and the prediction accuracy is improved.The main research contents of this dissertation are as follows:(1)The data cleaning of wind power's unsteady time series based on machine learning is realized.Firstly,multivariate data fusion of historical wind power data,historical meteorological data and wind turbine state data is carried out.The outliers are detected by an isolated forest algorithm and marked as missing values.The missing values are filled by a WGAN network model constructed by the GRUI network structure.The data cleaning of wind power data is completed and the characteristics of the original data are preserved to the maximum extent.Better data than the traditional mean filling before and after value filling are obtained.After that,through a series of procedures,such as dimension reduction,discretization,normalization and one-hot coding,data preprocessing meets the needs of deep learning;(2)An input data set of deep learning construction algorithm based on time sliding window is proposed.The output power of wind power has a certain periodicity.this dissertation proposes to use a time sliding window to construct time series data set of wind power,while expanding the original data set.It improves data utilization to fully mine data features.The time cycle characteristics of wind power output are effectively extracted,which creates conditions for the use of a deep neural network to realize the multivariable nonlinear fitting of wind power output;(3)The TLW-LSTM wind power output prediction model based on deep learning is constructed.The time sliding window algorithm is used to construct a data set as the input data set of deep learning,and the LSTM long-term and short-term memory network is used to construct a TLW-LSTM deep learning model.The model uses two-layer full connection layers as input and output layers,a three-layer multi-node LSTM layer as the hidden layer.Nadam optimizer,Droutout and regularization technology are used to improve the performance of the model.The prediction effect is higher and the accuracy of d?MAE reaches 92.7%;(4)The LW-CLSTM deep learning prediction model based on the CNN network is constructed.To overcome the disadvantages of the TLW-LSTM wind power output prediction model,such as long calculation time and slow network convergence,this dissertation uses the CNN network structure to optimize the TLW-LSTM model and construct an LW-CLSTM deep learning model.The model takes advantages of CNN's excellent feature extraction and LSTM's time series processing to effectively improve calculation efficiency without reducing the prediction accuracy.It significantly reduces the calculation time by 66% on the same data set,which lays a basis for the model's production;(5)The prediction accuracy and errors of traditional machine learning models and TLW-LSTM and LW-CLSTM deep learning models are compared and analyzed.To meet the prediction evaluation of wind power industry application,two kinds of accuracy evaluation standards,the maximum relative error statistical distribution method and the MAE average difference method,are designed and applied to the comparative experiment.Three traditional machine learning prediction models,the decision tree,random forest and support vector machine,are used to predict wind power output with the same input data set.The prediction results are compared with those of the TLW-LSTM and LW-CLSTM deep learning models constructed in this dissertation,and their prediction errors and prediction accuracy are analyzed.It is confirmed that the deep learning models constructed in this dissertation have high prediction accuracy.