# Short-Term Forecast of Tropospheric Zenith Wet Delay Based on TimesNet

## TL;DR

TimesNet improves short-term weather predictions by accurately forecasting atmospheric water vapor delays, outperforming existing models in various challenging conditions.

## Contribution

TimesNet introduces a novel method for ZWD prediction using dynamic temporal decomposition, achieving higher accuracy and robustness in diverse environments.

## Key findings

- TimesNet achieves an average seasonal RMSE of 5.73 mm, surpassing Informer and CNN-ATT by 27.4% and 42.8%, respectively.
- It maintains RMSE ≤ 7.8 mm year-round in high-altitude mountainous regions, where other models fail.
- In stable continental zones, TimesNet achieves sub-5 mm precision consistently.

## Abstract

What are the main findings?
TimesNet achieves an average seasonal RMSE of 5.73 mm across all 80 stations season samples, surpassing Informer (7.89 mm) by 27.4% and CNN-ATT (10.03 mm) by 42.8%.During summer severe convective events, TimesNet yields a mean RMSE of 5.91 mm. In a complex high-altitude mountainous region, it is the only model maintaining RMSE ≤ 7.8 mm year-round across all such stations; in a stable continental zone, it consistently achieves RMSE between 4.2 mm and 5.0 mm.

TimesNet achieves an average seasonal RMSE of 5.73 mm across all 80 stations season samples, surpassing Informer (7.89 mm) by 27.4% and CNN-ATT (10.03 mm) by 42.8%.

During summer severe convective events, TimesNet yields a mean RMSE of 5.91 mm. In a complex high-altitude mountainous region, it is the only model maintaining RMSE ≤ 7.8 mm year-round across all such stations; in a stable continental zone, it consistently achieves RMSE between 4.2 mm and 5.0 mm.

What are the implications of the main findings?
This provides a reliable algorithm selection basis for short-term precipitation forecasting in GNSS real-time meteorology.Subsequent work will integrate numerical weather prediction data to construct a precipitation forecasting neural operator architecture coupled with physical information, further enhancing predictive performance.

This provides a reliable algorithm selection basis for short-term precipitation forecasting in GNSS real-time meteorology.

Subsequent work will integrate numerical weather prediction data to construct a precipitation forecasting neural operator architecture coupled with physical information, further enhancing predictive performance.

The tropospheric zenith wet delay (ZWD) serves as a pivotal parameter for atmospheric water vapour inversion. By converting it into precipitable water vapour, high-temporal-resolution atmospheric humidity monitoring becomes feasible, providing crucial support for enhancing short-term rainfall forecast accuracy. However, ZWD exhibits significant non-stationarity due to complex influencing factors, and traditional models struggle to achieve precise predictions across all scenarios owing to limitations in local feature extraction. This article employs a ZWD prediction method based on the dynamic temporal decomposition module of TimesNet, re-constructing one-dimensional high-frequency ZWD time series into two-dimensional tensors to overcome the technical limitations of conventional models. Comprehensively considering topographical characteristics, climatic features, and seasonal factors, experiments were conducted using 30 s ZWD data from 20 IGS stations. This dataset comprised four consecutive days of PPP solutions for each season in 2023. Through comparative experiments with CNN-ATT and Informer models, the global prediction accuracy, seasonal adaptability, and topographical robustness of TimesNet were systematically evaluated. Results demonstrate that under the input-prediction window configuration where each can achieve the optimal accuracy, TimesNet achieves an average seasonal Root Mean Square Error (RMSE) of 5.73 mm across all seasonal station samples, outperforming Informer (7.89 mm) and CNN-ATT (10.02 mm) by 27.4% and 42.8%, respectively. It maintains robust performance under the most challenging conditions—including summer severe convection, high-altitude terrain, and climatically variable maritime zones—while achieving sub-5 mm precision in stable environments. This provides a reliable algorithmic foundation for short-term precipitation forecasting in Global Navigation Satellite System (GNSS) real-time meteorology.