Temperature Forecasting Correction Based on Operational GRAPES-3km Model Using Machine Learning Methods

Abstract

Postprocess correction is essential to improving the model forecasting result, in which machine learning methods play more and more important roles. In this study, three machine learning (ML) methods of Linear Regression, LSTM-FCN and LightGBM were used to carry out the correction of temperature forecasting of an operational high-resolution model GRAPES-3km. The input parameters include 2 m temperature, relative humidity, local pressure and wind speed forecasting and observation data in Shaanxi province of China from 1 January 2019 to 31 December 2020. The dataset from September 2018 was used for model evaluation using the metrics of root mean square error (RMSE), average absolute error (MAE) and coefficient of determination (R2). All three machine learning methods perform very well in correcting the temperature forecast of GRAPES-3km model. The RMSE decreased by 33%, 32% and 40%, respectively, the MAE decreased by 33%, 34% and 41%, respectively, the R2 increased by 21.4%, 21.5% and 25.2%, respectively. Among the three methods, LightGBM performed the best with the forecast accuracy rate reaching above 84%.

Machine learning based post‐processing of model‐derived near‐surface air temperature – A multimodel approach

“…(2016), and Zhang et al . (2022) use the Continuous Rank Probability Score or RMSE as verification scores, so comparisons with bias must be made with caution.…”

“…As the forecast length of AROME is much shorter compared to the forecast lengths of the other two models, only a period all three models have in common was used (from three to 30 hours). For each station, forecasts are taken from the closest model grid point without any interpolation (Rasp & Lerch, 2018; Zhang et al ., 2022). Most of the studies on the subject of post‐processing do apply interpolation (Rasp & Lerch, 2018; Zhang et al ., 2022), which is dictated either by the sparse grid of a model or by complex terrain.…”

“…For each station, forecasts are taken from the closest model grid point without any interpolation (Rasp & Lerch, 2018; Zhang et al ., 2022). Most of the studies on the subject of post‐processing do apply interpolation (Rasp & Lerch, 2018; Zhang et al ., 2022), which is dictated either by the sparse grid of a model or by complex terrain. In this work, however, the majority of stations are located in areas of little altitude diversity, where the greatest elevation difference between a station and its nearest grid point is 54 m. Additionally, as this study involves high‐resolution NWP models, the greatest distance between a station and the closest model grid point for the coarsest model is less than 5 km.…”

“…Since no MOS is operationally applied in IMGW‐PIB, predictions of RF were evaluated against single NWP models, a multiple linear regression model (MLR) and a basic ANN model. Linear regression is a benchmark reference to ML‐based methods within the area of post‐processing (Bouallègue et al ., 2023; Cho et al ., 2022; Zhang et al ., 2022). The MLR model was fitted using a core R package, stats (R Core Team, 2019).…”

“…T2m Model-derived air temperature at 2 m above ground level (AGL) Note: Predictors transformed to sine and cosine are marked with an asterisk. Zhang et al, 2022). The MLR model was fitted using a core R package, stats (R Core Team, 2019).…”