The Design And Implementation Of Short-term Wind Power Prediction System Based On Deep Learning

The inherent randomness and volatility of the wind determine the uncertainty and volatility of wind power,which increases the difficulty of power grid scheduling.With the increasing development of wind power generation in the world,its impact on the stability and economy of the power market and power system is deeper and deeper.In order to relieve the pressure of peak regulation and frequency regulation of power systems,accurate wind power prediction is particularly important.Wind power prediction mainly relies on numerical weather prediction data and supervisory control data.The numerical weather prediction data contains the regional meteorological features,while the supervisory control data contains features such as current and voltage,which are different.At the same time,different wind turbines have their own features,such as geographical location,etc.These features are essential for accurate wind power prediction.In addition,there is a strong correlation between wind power and historical data,but the traditional serialization methods have defects when dealing with the long-distance dependence between historical data and wind power.Aiming at the challenges mentioned above,this thesis proposes two short-term wind power prediction methods based on deep learning from two perspectives,and designs and implements a short-term wind power prediction system based on deep learning.The main contents of the thesis are as follows:(1)Aiming at the differences between weather prediction data and supervisory control data,and considering the characteristics of each wind turbine,a method based on multi-source information fusion is proposed in this thesis.Firstly,the Gated Recurrent Unit(GRU)is used to extract the time series of meteorological information from the numerical weather forecasting data.Secondly,to explore the implicit periodic characteristics in the wind power sequence,the original wind power series is decomposed into a series of different sub-modes by Variational Mode Decomposition(VMD),and the sub-modes matrix is further constructed for the power feature extraction part performed by Convolutional Neural Network(CNN).Finally,the wind turbine embedding is obtained by modeling the characteristics of the wind turbine.Experiments show that the method proposed in this thesis can effectively improve the accuracy of short-term wind power prediction.(2)Aiming at the long-distance dependence between wind power and historical data,a method based on self-attention mechanism and End-To-End Memory Networks is proposed in this thesis.To mine the correlation between historical status,self-attention mechanism is used to encode the historical data,the encoded features are stored in the memory pool.The attention mechanism is then used to search the memory pool for similar states,so as to capture the longdistance dependency information between the historical data and wind power,and the weighted combination of the states is used for prediction.Experimental results show the accuracy of the proposed method in short-term wind power prediction.(3)A short-term wind power prediction system based on deep learning is designed and implemented in this thesis.From the perspective of software engineering,the requirements of the system are defined,the detailed design and implementation of the system are carried out.The system is divided into three modules:data entry,power prediction,and state monitoring.The power prediction module integrates the wind power prediction methods mentioned above,combines the prediction results through multi-model combination and Monte Carlo Dropout method to quantify the uncertainty of the prediction.Finally,each module of the system has been tested,the test results show that the system has met the system design requirements and goals.