Research On Very-short-term Wind Speed Prediction Based On Merged Long-short Term Memory Networks

The output of wind power is uncertain.Large-scale wind power integration brings severe challenges to the safety,stability and economic operation of the power system.Accurate wind speed prediction is an important technical means to meet the above challenges.Accurate prediction of wind speed can not only optimize the unit combination,reduce the rotational reserve capacity of the system,and reduce the power limitation due to wind power plant abandonment,but also provide important reference for wind power enterprises to reasonably arrange power generation trading plans and improve the competitiveness of wind power enterprises in market competition,which is of great significance to the power system.How to improve the accuracy of ultra-short-term prediction of wind speed is the focus and difficulty in current research.Considering the relationship of time and space,the wind speed predicting modeling method simultaneously inputs the wind speed information of several adjacent points,which can improve the overall wind speed predicting level of the region and is one of the effective methods to improve the wind speed predicting accuracy at present.In order to further improve the accuracy of wind speed prediction,the Long-Short Term Memory network(LSTM)in deep learning algorithm is introduced into wind speed prediction modeling,and an ultra-short term wind speed prediction model integrating long-short term memory network is constructed.The research focus of this model is to extract features from multi-dimensional time series among multiple wind farms and learn the spatio-temporal correlation among wind farms,thus improving the accuracy of wind speed prediction.In this paper,LSTM is used to learn the wind speed sequence features of each wind farm,the feature fusion module is used to fuse the wind speed sequence features of multiple wind farms,and finally the predicted wind speeds of multiple wind farms are uniformly output through the fully connected output layer.The innovation of this model is to make full use of the time series feature extraction capability of long-short time memory network and the feature fusion capability of fully connected network,and intuitively consider the spatio-temporal correlation of wind speeds in multiple wind farms.In addition,the model has strong adaptability.The design of the model time series feature extraction module fully takes into account the scale of the predicted wind farm group.The length of input data of the time series feature extraction module can be increased to ensure that the model is suitable for large-scale wind farm communities.The model is verified by the data of three adjacent wind farms in a province in the south for two years.The results of example analysis show that the average absolute error and root mean square error of the model proposed in this paper are reduced by 5.6%-39.3% and 5.2%-44.9% respectively in the four seasons of No.1 wind farm.The average absolute error and root mean square error of No.2 wind farm in four seasons are reduced by 3.3%-15.7% and 3.7%-22.6% respectively.The average absolute error and root mean square error of No.3 wind farm are reduced by 5.9%-15.9% and 2.7%-22.6% respectively.The proposed fusion short-term and long-term memory network model achieves a better prediction effect than that of a single wind farm.It not only can realize ultra-short-term wind speed prediction for multiple wind farms,but also can fuse wind speed information of neighboring wind farms,thus improving the overall prediction level.