Research On Short-term Wind Field Forecast And Correction Based On Machine Learning

A fine forecasting system of wind field is an important part of modern refined weather forecasting.Wind field changes will directly or indirectly affect human life,especially the most obvious impact on flight action.It will not only affect the flight efficiency,but even affect the safety of flight that under a varying airport wind speed and direction.Therefore,accurate airport weather forecast is pretty necessary.Recently,in the research on precision of weather forecast are achieved by scholars in China and other countries,and have made great progress,but the effect of short-term wind feild forecast is still unable to meet the needs of aeronautical meteorological forecasting.This study was carried out in Gansu Province.Based on the statistics of wind field data in Gansu's station over the years,the temporal and spatial distribution characteristics of wind fields are briefly analyzed,and the characteristics of wind field changes in Gansu Province are explored.The historical ground observation data was used to test the wind field data forecasted by ECMWF.Based on this and combined with the special requirements for airport's wind forecasting,propose a new calculation(A_Int)to test accuracy of wind field forecasting that integrates wind direction and wind speed.On this basis,using the random forest(RF)and long-short-term memory based on random forest feature selection(RF-LSTM)in machine learning,carried out a short-term forecast and forecast-correction method of wind field.By comparing and analyzing the accuracy of different methods and different forecast times,the short-term wind field forecast model of the airport was established.We can have some conclusions.(1)Through statistical analysis of wind speed data at various stations in Gansu Province in the past 30 years,the average large wind speed in the province occurred from March to May,and the average wind speed was the largest in April;the average wind speed in Hexi Corridor and its north,southwestern and eastern Gansu is large,and the average wind speed in the central and southeastern Gansu is relatively small.Through statistical analysis of the wind direction data of each stations in Gansu Province in the past 10 years,the Hexi Corridor and its north,as well as most areas in the central and southwest of Gansu are northerly winds prevailly,while easterly or southerly winds reign in the eastern part of Gansu.(2)In the wind field forecast,the calculation results of the importance by the RF for each meteorological element are different under different forecast time.For example,in the 3h wind speed forecast,the most important meteorological element is 0cm ground temperature,but it is 3 hours air pressure change in the 6h forecast.(3)RF and RF-LSTM has strong forecasting capabilities for the station wind.The forecast results of the representative stations were tested.The wind speed MAE by RF predicted of 1-6h is 1.0m/s,the MAE of wind direction was 59°,and wind speed MAE by RF-LSTM predicted of 1-6h is 0.86m/s,the MAE of the wind direction is 55°.When studying the prediction of the wind field by machine learning based on historical observation data,the prediction effect is better within 1h-6h.The wind field forecasting models constructed by the two algorithms have a reduced ability to predict the higher wind speeds and the direction at higher speeds.Whether for wind speed forecast or direction forecast,the result of RF-LSTM is better than that of RF model.(4)Using the surface observations data of each station in Gansu Province,the average MAE of the wind speed forecast by ECMWF is 1.5m/s-1.9m/s,and the MAE of the wind direction forecast is 96°-107°,and found that there is a connection between the error of ECMWF forecast wind feild and the size of observed wind speed.Generally speaking,the observed wind speed is large,the forecast wind speed error is large,but the forecast wind direction error is small;the observed wind speed is small,the forecast wind speed error is small,but the error of the forecast wind direction is large;the A_Int method is used to test the accuracy of the 3-12 h wind field forecast by ECMWF,and it is found that the average annual accuracy rate is between 20% and 38%.(5)The RF-LSTM wind field forecast-correction model has an obvious correction effect on the wind field forecast results of the ECMWF.For the correction of wind direction,the MAE of the wind direction can be reduced by 36°-41° under different forecast time,and the MAE of the corrected wind direction is 61°-65°;for the correction of wind speed,the average MAE of each forecasting time is 0.59m/s-1.14m/s,the corrected MAE of wind speed is 0.76m/s-0.9m/s.The forecast-correction model of RF-LSTM has a good adaptability in Gansu province.The average forecast accuracy rate of each station can be increased by about 32%,and the average wind forecast accuracy rate can be increased to 49%-69%.(6)In practical applications,the RF-LSTM 1h-6h wind field forecast results have a high accuracy rate and a time resolution of 1h,which can meet the requirements of airport forecasting for wind field forecasting.For wind field forecasting greater than 6h,the ability of RF-LSTM gradually weakens,and the accuracy rate gradually decreases.The results of the forecast correction using RF-LSTM are more in line with reality.Combining the two methods is the short-term wind field forecast model of the airport.