TLDR: A new benchmark called RainfallBench has been introduced for rainfall nowcasting, focusing on predicting precipitation within 0-3 hours. It uses a unique dataset from over 12,000 GNSS stations, incorporating precipitable water vapor (PWV) and addressing challenges like zero inflation, temporal decay, and non-stationarity in rainfall data. The research also proposes the Bi-Focus Precipitation Forecaster (BFPF), a plug-and-play module that significantly improves Transformer-based models’ performance by focusing on non-zero rainfall events and recent temporal information. Evaluations show that RNN-based models perform strongly, and the BFPF enhances extreme rainfall prediction.
Rainfall nowcasting, the prediction of precipitation within the next 0 to 3 hours, is a critical task for managing natural disasters and planning real-time responses. However, traditional time-series forecasting benchmarks in meteorology often focus on variables with predictable patterns, like temperature, which don’t fully capture the complexities of rainfall prediction.
To bridge this gap, researchers have introduced RainfallBench, a new benchmark specifically designed for rainfall nowcasting. This benchmark addresses the unique challenges of rainfall data, such as its sparse nature (many periods with no rain, known as zero inflation), the decreasing relevance of older data (temporal decay), and its constantly changing statistical properties (non-stationarity).
What makes RainfallBench stand out is its comprehensive dataset. It compiles five years of meteorological observations, recorded every 15 minutes from over 12,000 Global Navigation Satellite System (GNSS) stations worldwide, spanning from 2018 to 2022. Crucially, it incorporates precipitable water vapor (PWV), a vital indicator of rainfall that is often missing from other datasets. The dataset includes six key variables: temperature at 2 meters (t2m), surface pressure (sp), relative humidity (rh), wind speed, PWV, and total precipitation (tp) as the target variable.
Analysis of the data within RainfallBench confirms the significance of PWV. It shows the strongest positive correlation with total precipitation, making it an essential feature for accurate short-term rainfall predictions. This aligns with meteorological observations that suggest a higher probability of precipitation when PWV exceeds certain thresholds.
The benchmark also features specialized evaluation strategies. These assess model performance across different time scales, evaluate their ability to predict extreme rainfall events (defined as precipitation exceeding 2mm in a 15-minute period), and measure overall accuracy and reliability. Over 20 state-of-the-art models from various architectural families, including MLP-based, CNN/TCN-based, RNN-based, GNN-based, KAN-based, and Transformer-based designs, were evaluated on RainfallBench.
Through this extensive evaluation, the researchers identified that existing models often struggle with the zero inflation and temporal decay characteristics of rainfall data. To tackle these limitations, they developed the Bi-Focus Precipitation Forecaster (BFPF). This is a plug-and-play module designed to enhance rainfall time-series forecasting, particularly for Transformer-based models. The BFPF has two main components: a Non-Zero Focus module, which helps the model pay more attention to actual rainfall events rather than dry periods, and a Temporal Focus module, which emphasizes more recent temporal information, reflecting the temporal decay property of rainfall.
Experimental results showed that RNN-based models, such as P-sLSTM, performed consistently well, likely due to their strength in handling short-term memory and sparse, irregular sequences. Furthermore, the Informer model, when augmented with the BFPF module, significantly outperformed other Transformer-based models, especially in forecasting extreme rainfall events. The ablation studies also confirmed that incorporating PWV data substantially reduces nowcasting errors and that the BFPF module significantly improves the performance of Transformer models by addressing rainfall-specific challenges.
Also Read:
- Forecasting Ground Movement: A Multi-Modal Transformer Achieves Unprecedented Accuracy
- Time Series Models Struggle with Real-World Video Data, Study Finds
This research introduces a robust benchmark and a novel module that together advance the field of rainfall nowcasting, offering a more realistic and challenging environment for evaluating and developing predictive models. For more details, you can read the full research paper here.
