Optimised weighted mean temperature model based on generalised regression neural network

Abstract

ABSTRACTThe weighted mean temperature (Tm) is a key parameter to calculate the Global Navigation Satellite System (GNSS)-based precipitable water vapour (PWV). Data fusion provides a solution to depict the characteristics of Tm in detail. However, multi-source heterogeneity, unequal accuracies and even serious system deviation may lead to unreliable and inconsistent accuracies in the fusion results. We utilise generalised regression neural network (GRNN) to establish an optimised model for the Tm from the stratified numerical weather forecasting products of the mesoscale version of the Global/Regional Assimilation and PrEdiction System (GRAPES_MESO) of China and the Tm from the European Centre for Medium-Range Weather Forecasts (ECMWF) fifth-generation reanalysis (ERA5) data around China from 2016 to 2017. Then, an example fusion using the radiosonde (RS) Tm and the optimised Tm is carried out. The results confirm the systematic deviations between GRAPES/ERA5 Tm and RS Tm. After optimisation, the bias of GRAPES and ERA5 Tm is almost eliminated, and the root mean squared error (RMSE) decreased by 21.1% and 18.7%, respectively. Compared to RS Tm, the fusion results based on the optimised Tm have good consistencies and unbiased accuracies, and can merge more detailed spatial features than that of a single data source.

Keywords

• Weighted mean temperature
• multi-source data
• optimization
• fusion
• generalized regression neural network