Temperature Bias Correction and Downscaling Forecasting Based on Numerical Weather Prediction and Deep Learning

This paper proposes deep learning spatiotemporal downscaling (DLSD): a deep learning model that integrates temperature bias correction and spatiotemporal downscaling based on numerical weather prediction (NWP) to enhance the accuracy and spatiotemporal resolution of temperature forecasting. The target area is partitioned into multiple subregions, thereby reducing the complexity of the model and augmenting the training dataset. By employing embedding techniques, the subregions and temporal variables are encoded in combination with diverse meteorological parameters derived from NWP for different atmospheric levels, terrain attributes, and geographic coordinates as input features. High-fidelity and high-resolution surface-temperature datasets serve as training targets. DLSD is implemented in Hubei Province to produce hourly temperature forecasts at a resolution of 0.05°×0.05°. Upon implementation and benchmarking against Integrated Forecasting System predictions, it substantially improves evaluation metrics, including the mean absolute error (MAE), root mean square error, accuracy, and structural similarity index measure. DLSD efficiently reproduces fine spatial details and excels on regions with substantial elevation variations: the western mountainous area yields a 42.1% MAE reduction, contrasting with a 28.1% reduction observed on the eastern plains. The model also demonstrates effective temporal downscaling, narrowing the size of the hourly MAE range from 0.9 ℃ to between 0.5 and 0.7 ℃, thereby enhancing the stability of hourly forecasts. Site-specific validations affirm the efficacy of DLSD in capturing the variability of daily temperatures, albeit with less prominent correction enhancements at stations than at grid points; this is partly attributable to the limited training data and observational sites available for the western mountainous terrain.