What question did this study set out to answer?

The aim is to improve temperature forecasting accuracy and resolution by integrating deep learning with numerical weather prediction.

April 1, 2026Open Access

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

Key Points

The aim is to improve temperature forecasting accuracy and resolution by integrating deep learning with numerical weather prediction.
Developed a deep learning model incorporating temperature bias correction and downscaling techniques.
Partitioned the target area into subregions to reduce model complexity and augment the training dataset.
Used convolutional neural networks to analyze various meteorological parameters and terrain attributes.
Evaluated model performance against predictions from the Integrated Forecasting System (IFS).
Achieved a 35.7% reduction in mean absolute error (MAE) compared to previous models.
High-resolution forecasts of 0.05°×0.05° were produced, with significant stability in hourly predictions.
Observed a notable 42.1% MAE reduction in mountainous areas and 28.1% in eastern plains.
Hourly MAE improved from 0.9°C to a range of 0.5–0.7°C due to effective temporal downscaling.

Abstract

To enhance the accuracy and spatiotemporal resolution of temperature forecasting, this paper proposes a deep learning model that integrates temperature bias correction and spatiotemporal downscaling (DLSD) based on numerical weather forecasting (NWP). The target area is partitioned into multiple subregions, 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 at different atmospheric levels, terrain attributes, and geographic coordinates as input features. High-fidelity and high-resolution surface temperature datasets from the China Meteorological Administration Land Surface Data Assimilation System (CLDAS) serve as training targets. Leveraging a convolutional neural network (CNN)-based DLSD framework, the model conducts bias correction and spatiotemporal downscaling on NWP temperature forecasts. When implemented in Hubei Province and benchmarked against IFS predictions, DLSD offers 1-hour, 0.05°×0.05-resolution temperature forecasts with significantly improved evaluation metrics, including the mean absolute error (MAE), root mean square error (RMSE), accuracy (Acc), and structural similarity index measure (SSIM). Notably, the MAE is reduced by 35.7%, with DLSD demonstrating proficiency in reproducing fine spatial details and excelling in regions with substantial elevation variations; the western mountainous area yields a 42.1% MAE reduction, contrasting with a 28.1% decrease observed in the eastern plains. The model also demonstrates effective temporal downscaling, narrowing the hourly MAE range from 0.9°C to 0.5–0.7°C, thereby enhancing the stability of hourly forecasts. High-temperature daytime corrections are particularly impactful, with the MAE decreasing by 0.7–0.9°C. 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 attributed to the limited training data and observational sites available for the western mountainous terrain.

Bookmark

View Full Paper