TCP-Diffusion: A Multi-modal Diffusion Model for Global Tropical Cyclone Precipitation Forecasting with Change Awareness

TL;DR

A multi-model framework with specialized encoders to extract richer information from environmental variables and results provided by NWP models is proposed, which outperforms other DL methods and the NWP method from the European Centre for Medium-Range Weather Forecasts (ECMWF).

Abstract

Precipitation from tropical cyclones (TCs) can cause disasters such as flooding, mudslides, and landslides. Predicting such precipitation in advance is crucial, giving people time to prepare and defend against these precipitation-induced disasters. Developing deep learning (DL) rainfall prediction methods offers a new way to predict potential disasters. However, one problem is that most existing methods suffer from cumulative errors and lack physical consistency. Second, these methods overlook the importance of meteorological factors in TC rainfall and their integration with the numerical weather prediction (NWP) model. Therefore, we propose Tropical Cyclone Precipitation Diffusion (TCP-Diffusion), a multi-modal model for global tropical cyclone precipitation forecasting. It forecasts TC rainfall around the TC center for the next 12 hours at 3 hourly resolution based on past rainfall observations and multi-modal environmental variables. Adjacent residual prediction (ARP) changes the training target from the absolute rainfall value to the rainfall trend and gives our model the ability of rainfall change awareness, reducing cumulative errors and ensuring physical consistency. Considering the influence of TC-related meteorological factors and the useful information from NWP model forecasts, we propose a multi-model framework with specialized encoders to extract richer information from environmental variables and results provided by NWP models. The results of extensive experiments show that our method outperforms other DL methods and the NWP method from the European Centre for Medium-Range Weather Forecasts (ECMWF).

Figures (13)

• Figure 1: The difference between the regular precipitation forecasting (a) and TC (b) precipitation forecasting. Regular precipitation forecasting (a) usually focuses on a specific fixed region, where the predicted area (blue box) does not move with the rainfall area (black dashed boxes). In TC precipitation forecasting (b), the primary region of interest (blue boxes) moves with the TC. Sub-figure (c) represents the main prediction process of our model. The light blue box represents historical data, which is considered to be the input of the forecasting model. The light orange box represents the TC rainfall that should be predicted. This paper just focuses on the prediction of TC rainfall.
• Figure 2: The framework of TCP-Diffusion. Sub-figure (a) shows the inference stage of TCP-Diffusion (black+red arrows). Our model will use $N$ loop denoising processes to obtain the future rainfall change $\Delta Rainfall_{Future}$ from a noise input. The sub-figure (b) shows the model structure of the denoising process and also shows the training stage of TCP-Diffusion (black+green arrows). In the training stage, the Output will be used to calculate the loss and update the parameters of TCP-Diffusion (green arrow). In the inference stage, the Output will be used to get the future rainfall change $\Delta Rainfall_{Future}$ with noise at Step $s-1$ (red arrow). The sub-figure (c) shows the process (Equation \ref{['accumulation']}) of getting the final predicted TC rainfall $Rainfall_{Future}$ with the latest rainfall $Rainfall_{Current}$ and the future rainfall change $\Delta Rainfall_{Future}$.
• Figure 3: The prediction results of different DL methods on TC Dumazile in the North Indian Ocean at 03/03/2018 06:00. The first column is the previous 4 timesteps of rainfall data used as input. The second column is the future TC rainfall we want to predict. Each subsequent column from the left shows the predictions from a different DL method. "TCP-Diff #1" and "TCP-Diff #2" are two samples from our TCP-Diffusion model with different initial random noise $r$.
• Figure 4: TC rainfall frequency distributions of different DL forecasting methods. The grey histogram is the distribution of MSWEP observations. The coloured lines show histograms of rainfall rates from different DL forecasting methods. The circular magnified region more clearly shows differences between these methods in the indicated span of rainfall intensity. Note the logarithmic vertical axis.
• Figure 5: Radially Averaged Power Spectral Density (RAPSD) analysis across datasets. The plots compare the power spectral density of predicted rainfall using different models over a range of wavenumber. In the low wavenumber region, higher power spectral density represents that the displayed data belongs to large-scale weather systems, such as fronts and tropical cyclones. The power spectral density of TCP-Diffusion in the low wavenumber region is closer to the ground truth rainfall (MSWEP), indicating that our method achieves better capture of rainfall characteristics at the TC scale. Note the logarithmic vertical axis.