How Effective Are Time-Series Models for Precipitation Nowcasting? A Comprehensive Benchmark for GNSS-based Precipitation Nowcasting

TL;DR

RainfallBench is the first comprehensive GNSS-PWV–based benchmark for precipitation nowcasting, explicitly addressing zero inflation, temporal decay, and non-stationarity by integrating PWV with ERA5-Land and IMERG data across 140 stations. It evaluates 17 time-series models from six architectures using multi-scale, multi-resolution, and extreme-rainfall tasks, and introduces Bi-Focus Precipitation Forecaster (BFPF) as a plug-in module that enhances Transformer-based forecasts by emphasizing non-zero events and recent temporal context. Across experiments, Transformer-based Informer with BFPF achieves the best overall accuracy, while RNNs like xLSTM excel in extreme-event forecasting, highlighting the value of model diversification for rainfall nowcasting. The framework and findings have practical impact for disaster mitigation and real-time decision-making, enabling more reliable, PWV-informed precipitation nowcasting at global scales and guiding future research on domain-adaptive time-series forecasting.

Abstract

Precipitation Nowcasting, which aims to predict precipitation within the next 0 to 6 hours, is critical for disaster mitigation and real-time response planning. However, most time series forecasting benchmarks in meteorology are evaluated on variables with strong periodicity, such as temperature and humidity, which fail to reflect model capabilities in more complex and practically meteorology scenarios like precipitation nowcasting. To address this gap, we propose RainfallBench, a benchmark designed for precipitation nowcasting, a highly challenging and practically relevant task characterized by zero inflation, temporal decay, and non-stationarity, focusing on predicting precipitation within the next 0 to 6 hours. The dataset is derived from five years of meteorological observations, recorded at hourly intervals across six essential variables, and collected from more than 140 Global Navigation Satellite System (GNSS) stations globally. In particular, it incorporates precipitable water vapor (PWV), a crucial indicator of rainfall that is absent in other datasets. We further design specialized evaluation protocols to assess model performance on key meteorological challenges, including multi-scale prediction, multi-resolution forecasting, and extreme rainfall events, benchmarking 17 state-of-the-art models across six major architectures on RainfallBench. Additionally, to address the zero-inflation and temporal decay issues overlooked by existing models, we introduce Bi-Focus Precipitation Forecaster (BFPF), a plug-and-play module that incorporates domain-specific priors to enhance rainfall time series forecasting. Statistical analysis and ablation studies validate the comprehensiveness of our dataset as well as the superiority of our methodology.

Figures (12)

• Figure 1: Overview of the RainfallBench framework. The benchmark is organized into three main components: the data layer, the model layer, and the evaluation layer. The data layer integrates three sources: GNSS-PWV, ERA5, and GPM. The model layer includes 17 models across six major deep learning architectures, while the evaluation layer encompasses multi-scale prediction, multi-resolution forecasting, and extreme rainfall assessment.
• Figure 2: Global distribution of 140 selected GNSS stations from the proposed RainfallBench dataset across seven continents, ensuring balanced spatial coverage for evaluating precipitation forecasting models.
• Figure 3: Distribution of meteorological variables at the HKST station from 2018 to 2024. The dataset contains six variables per hourly observation.
• Figure 4: Pairwise correlation matrices among meteorological variables and precipitation in the RainfallBench dataset, computed using (a) Pearson, (b) Kendall, and (c) Spearman coefficients.
• Figure 5: Analysis of data properties in RainfallBench. The benchmark exhibits three key characteristics distinguishing it from standard time series: (i) zero inflation, (ii) temporal dependency decay, and (iii) non-stationarity, along with their implications for modeling.