Research On Wind Power Prediction Based On Long And Short Time Memory Neural Network

With the increasing demand for clean energy,the use of wind energy to convert into electricity has been vigorously developed.With the introduction of China’s "carbon peak" and "carbon neutral" national policies,wind power generation is even more effective as an alternative to traditional thermal power generation.Wind power generation also has its own shortcomings.For example,the source of power is wind energy,which is not as persistent and continuous as thermal power generation.It has volatility and interruption,and the output of electric energy is uneven and intermittent.Therefore,during the process of grid connection,it will increase the burden on the grid and affect the safety of users’ electricity consumption.The current treatment method is to predict the power of wind power generation,Reduce the impact of wind power generation on the grid by controlling power output.Aiming at the problem of wind power generation power prediction in the ultra-short term,this paper forecasts the power of wind power generation by optimizing the data cleaning process and the long and short term memory neural network calculation method.Firstly,based on the real-time wind power output and real-time wind speed data of a wind turbine,the data cleaning method used is divided into three steps.1.Classify data and label zero power and artificially limited power data.By analyzing the correlation of data,a fitting curve between wind power generation power and wind speed is obtained as a benchmark,using which zero power and artificially limited power data are eliminated,and data points for normal operation of wind turbines are retained.2.Use the Romanov benchmark to eliminate scattered data in the data,eliminate large data of numerical deviation caused by mechanical or sensor reasons,and retain data with relatively stable wind speed and power mapping.3.The retained data is resampled based on time and combined into a time series.Multiple sets of time series are used to predict the next moment.Two methods,Pearson correlation coefficient method and minimum absolute contraction and selection algorithm,are used for data correlation analysis to obtain an optimal result with a number of time series of five-series.Then,using a long and short-term memory neural network,using different numbers of time series as prediction data,the above results based on data correlation analysis were verified through calculation.The data of the wind turbine for 12 months throughout the year are verified to have good root mean square error,and the numerical value is stable,indicating the robustness.The universality of this method is illustrated by analyzing the data of two wind turbines adjacent to the wind turbine in the same month.Finally,field tests were conducted on the data of three wind turbines at a remote wind power station.Using the spatiotemporal correlation of neighboring wind turbines,the prediction model was optimized.The prediction results did not deviate from the measured power curve,and met the requirements of national industry standards in terms of root mean square error.Develop a set of wind power generation power prediction software that can connect to the wind power station database to read,display,and predict wind power generation information.