Research On The Forecast Correction Of Xinjiang Surface Temperature And Wind Based On U-net Neural Networ

Numerical weather prediction(NWP)is the current mainstay of operational forecasting.However,it is difficult for NWP models to accurately reflect the development of weather as a result of the chaotic characteristics of atmospheric dynamics,uncertainties in the initial conditions and limitation of parameterization schemes.There are systematic errors from control or ensemble forecasts of a single model,and its accuracy needs to be further improved.Statistical postprocessing methods can effectively improve raw forecasting skills and are widely used in scientific researches and operational forecast due to their low cost and high efficiency.Both calibrations of single-model forecasts and multi-model ensemble forecasts can reduce systematic biases of models and improve forecasting skills.Xinjiang,located in the arid region of northwest China,has complex terrain and frequent strong winds,facing serious meteorological situation such as intensified drought and increased wind-sand disasters in the past 60 years.Meanwhile,weather forecasts with lead times of 1–7days play important roles in issuing early warnings and assisting governmental decision making.Therefore,in this study,several statistical postprocessing methods such as unary linear regression(ULR),the decaying averaging method(DAM),quantile mapping(QM)and U-net neural network are utilized to calibrate the gridded forecast of surface air temperature and wind in Xinjiang from the Global Ensemble Forecasting System(GEFS)of the National Centers for Environmental Prediction(NCEP).Four prediction test methods are adopted to make a quantitative evaluation of the correction effect after statistical postprocessing,namely,the mean absolute error(MAE),the mean absolute error skill score(MAESS),the hit rate of 2℃(HR2)and the pattern correlation coefficient(PCC).The temporal and spatial distribution characteristics of errors are also illustrated.Furthermore,in order to distinguish the error sources of each forecasting scheme and to reveal their capabilities of calibrating errors of different sources,the error decomposition analysis is carried out based on the mean square errors.The main conclusions are as follows:(1)In terms of the raw temperature forecasts from GEFS,the largest MAEs mainly occur over the Altai Mountains,the Junggar Basin,the northern Tianshan Mountain,the Tarim Basin and the Kunlun Mountains.The forecast skills in northern Xinjiang are better than in the south.The error dispersion of GEFS for surface air temperature is asymmetrical and there are more warm biases than cold biases.The four calibration methods are characterized by different magnitudes of ameliorations and all have better improvement in southern Xinjiang than in the north,especially in the Tarim Basin where the skill improvement magnitude is the largest.Among three conventional linear methods,DAM has best calibration performance.However,compared with DAM,U-net can improve temperature forecast skills over almost the whole area of Xinjiang and effectively eliminate the warm biases of the raw GEFS forecasts.(2)The largest MAEs of the wind speed forecasts from GEFS mainly occur in high altitude areas such as the Tianshan Mountain and the Kunlun Mountains,while the direction forecasts have limited skills over the Junggar Basin and the Tianshan Mountain.The error dispersion of GEFS for wind speed is asymmetrical and there are more positive biases than negative biases.All four calibrations methods can improve the wind speed and direction forecast skills,the most obvious magnitudes of ameliorations are located in high altitude areas(the Tianshan Mountain and the Kunlun Mountains).Among three conventional linear methods,QM and ULR have best calibration performance for wind speed and direction respectively.However,compared with best conventional methods,U-net neural network can further improve wind speed and direction forecast skills over almost the whole area of Xinjiang and effectively eliminate the positive biases of the raw wind speed forecasts from GEFS.(3)After error decomposition,for temperature forecasts,the bias term is the leading source of error in the raw GEFS forecasts,and barely changes as the lead time increases.After four statistical postprocessing methods,the sequence term becomes the leading source of error and increases significantly with the lead time.All four forecast calibrations effectively reduce the bias and distribution error of the raw forecasts,among which DAM has the best performance.Only U-net significantly reduces the sequence term.For wind speed and direction forecasts,the sequence term is the leading source of error of all forecast schemes and increases significantly with the lead time.The growth rate of the sequence term of the wind direction after U-net decelerates during lead times of 4-7days.All calibration methods reduce the bias term and the distribution term of wind speed and direction,among which QM has the best performance,but it has negative improvements for the sequence term of wind direction.U-net has greatest improvement magnitudes for the sequence term of wind speed and direction.During long lead times,the advantages of U-net over conventional methods become more obvious in improving wind direction skills.This study aims to use various postprocessing methods such as U-net neural network to effectively improve the temperature and wind forecasting skills of 1-7 lead days,and to diagnose and analyze the forecast errors of different schemes.In the current background that numerical grid forecasting has become the mainstream trend of meteorological business forecasting,it provides an important reference for the use of numerical model prediction results in business departments and scientific research,and has important scientific significance and application value.