AdaNAS: Adaptively Post-processing with Self-supervised Neural Architecture Search for Ensemble Rainfall Forecasts

TL;DR

AdaNAS tackles the challenge of post-processing ensemble rainfall forecasts by marrying self-supervised neural architecture search with a rainfall-aware search space and a rainfall-level regularization. The method automatically designs architectures tailored to precipitation data, achieving state-of-the-art performance on TIGGE records for both precipitation amounts and intensity classification, especially in coastal heavy-rain zones. It demonstrates significant gains in MAE, RMSE, NSE, ACC, and HSS over manually designed baselines and other NAS approaches, while also reducing inference time. This approach reduces manual design costs and offers practical gains for operational ensemble rainfall prediction.

Abstract

Previous post-processing studies on rainfall forecasts using numerical weather prediction (NWP) mainly focus on statistics-based aspects, while learning-based aspects are rarely investigated. Although some manually-designed models are proposed to raise accuracy, they are customized networks, which need to be repeatedly tried and verified, at a huge cost in time and labor. Therefore, a self-supervised neural architecture search (NAS) method without significant manual efforts called AdaNAS is proposed in this study to perform rainfall forecast post-processing and predict rainfall with high accuracy. In addition, we design a rainfall-aware search space to significantly improve forecasts for high-rainfall areas. Furthermore, we propose a rainfall-level regularization function to eliminate the effect of noise data during the training. Validation experiments have been performed under the cases of \emph{None}, \emph{Light}, \emph{Moderate}, \emph{Heavy} and \emph{Violent} on a large-scale precipitation benchmark named TIGGE. Finally, the average mean-absolute error (MAE) and average root-mean-square error (RMSE) of the proposed AdaNAS model are 0.98 and 2.04 mm/day, respectively. Additionally, the proposed AdaNAS model is compared with other neural architecture search methods and previous studies. Compared results reveal the satisfactory performance and superiority of the proposed AdaNAS model in terms of precipitation amount prediction and intensity classification. Concretely, the proposed AdaNAS model outperformed previous best-performing manual methods with MAE and RMSE improving by 80.5\% and 80.3\%, respectively.

Figures (8)

• Figure 1: (a) The study area $[21.0^{o}N \sim 29.0^{o}N, 109.5^{o}E \sim 117.5^{o}E]$ includes part of the coastal and inland regions of southern China. (b) The data characteristic of rainfall forecast are abstracted from the raw observational data by NWP models. (c) Overview of previous methods applied to rainfall forecast and our AdaNAS, which can be summarized by two aspects: statistic-based method and learning-based method.
• Figure 2: The overview of our AdaNas by using contrastive learning to adaptively search suitable architecture for ensemble rainfall forecast post-processing.
• Figure 3: Channel-aware block. Channel-aware block focuses on high rainfall pixels and assigns large weights to highlight the features of these high rainfall pixels
• Figure 4: The details of our channel-aware block (a) and space-aware block (b).
• Figure 5: Proportion of precipitation events in multi-model dataset. The proportion of rainfall at different levels is extremely uneven, with Light accounting for nearly 80%.