Improving Ensemble Forecasts of Abnormally Deflecting Tropical Cyclones with Fused Atmosphere-Ocean-Terrain Data

Qixiang Li; Yuan Zhou; Shuwei Huo; Chong Wang; Xiaofeng Li

cs.LGMar 31, 2026

Improving Ensemble Forecasts of Abnormally Deflecting Tropical Cyclones with Fused Atmosphere-Ocean-Terrain Data

Authors: Qixiang Li, Yuan Zhou, Shuwei Huo, Chong Wang, Xiaofeng Li

Organizations: School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China · School of Computer Science and Engineering, Northeastern University, Shenyang 110819, China · Key Laboratory of Ocean Circulation and Waves, Institute of Oceanology, and the Center for Ocean Mega-Science, Chinese Academy of Sciences, Qingdao 266071, China

Abstract

Deep learning-based tropical cyclone (TC) forecasting methods have demonstrated significant potential and application advantages, as they feature much lower computational cost and faster operation speed than numerical weather prediction models. However, existing deep learning methods still have key limitations: they can only process a single type of sequential trajectory data or homogeneous meteorological variables, and fail to achieve accurate forecasting of abnormal deflected TCs. To address these challenges, we present two groundbreaking contributions. First, we have constructed a multimodal and multi-source dataset named AOT-TCs for TC forecasting in the Northwest Pacific basin. As the first dataset of its kind, it innovatively integrates heterogeneous variables from the atmosphere, ocean, and land, thus obtaining a comprehensive and information-rich meteorological dataset. Second, based on the AOT-TCs dataset, we propose a forecasting model that can handle both normal and abnormally deflected TCs. This is the first TC forecasting model to adopt an explicit atmosphere-ocean-terrain coupling architecture, enabling it to effectively capture complex interactions across physical domains. Extensive experiments on all TC cases in the Northwest Pacific from 2017 to 2024 show that our model achieves state-of-the-art performance in TC forecasting: it not only significantly improves the forecasting accuracy of normal TCs but also breaks through the technical bottleneck in forecasting abnormally deflected TCs.

Figures & tables

Fig. 1 : Abnormal deflection of TC GEIMI during landfall in Taiwan in 2024. The Central Mountain Range (peaking at nearly 4,000 m) interacted with GEIMI’s low-level circulation within the 600–700 hPa vertical layer, inducing a venturi effect that guided its low-level center southward. Satellite imagery revealed complete eye dissolution after GEIMI’s collision with the mountain range.

Fig. 2 : The architecture of AOT-TCNet, comprising Encoders and TMA-MoE. The environment data $D$ and observed trajectories $X$ are encoded and passed to the fusion module. The experts can predict different mode trajectory distributions for the given observation. The gating network estimates probabilities $\pi$ for the experts. The model samples or selects a expert from $\pi$ and predicts a trajectory $Y$ conditioned on the features $c$ .

Methods	6h	12h	18h	24h	6h	12h	18h	24h	6h	12h	18h	24h
	Distance (km)				Pressure (hpa)				Wind (m/s)
SGAN [ 36 ]	28.88	61.75	98.74	140.61	1.91	3.12	4.2	5.12	1.05	1.69	2.28	2.81
GBRNN [ 27 ]	29.93	65.06	105.74	152.06	-	-	-	-	1.16	1.89	2.52	3.1
MMSTN [ 9 ]	27.57	59.09	96.54	139.19	1.69	2.86	3.94	4.74	0.95	1.52	2.1	2.55
MGTCF [ 34 ]	23.14	43.37	67.09	93.08	1.37	2.04	2.66	3.29	0.73	1.17	1.55	1.86
TCN m [ 41 ]	22.98	43.83	66.41	93.76	-	-	-	-	0.7	1.09	1.43	1.75

TABLE I : Comparisons of average absolute error of TC prediction of different methods

Fig. 3 : Forecast results for selected TCs in 2024. Historical trajectories are shown in red, ground truth future paths in blue, and model-predicted trajectories in green.

Methods	6h	12h	18h	24h	6h	12h	18h	24h	6h	12h	18h	24h
	Distance (km)				Pressure (hpa)				Wind (m/s)
MMSTN	28.47	59.53	94.82	135.2	2.05	3.58	5.01	6.14	1.08	1.83	2.54	3.12
MGTCF	25.21	46.63	71.04	98.74	1.57	2.51	3.23	3.9	0.83	1.39	1.82	2.25
Our	27.16	45.97	64.76	87.36	1.54	2.25	2.65	3.01	0.85	1.23	1.41	1.65

TABLE II : Comparison of different methods for TC prediction from 2020 to 2024

	2020				2021				2022
	Distance (km)		Wind (m/s)		Distance (km)		Wind (m/s)		Distance (km)		Wind (m/s)
Methods	12h	24h	12h	24h	12h	24h	12h	24h	12h	24h	12h	24h
CMA	-	73.5	-	4.5	56.7	88.3	-	4.4	51.9	76.5	-	4.5
JMA	-	74.2	-	4.5	51.3	85.3	-	4.4	48.4	72.7	-	5.5
JTWC	-	74.9	-	4.9	56.6	89.3	-	4.8	55.6	78.2	-	5.4
KMA	-	88.3	-	4.7	61	97.9	-	4.9	59.8	83.8	-	5.1

TABLE III : Comparison with Global Operational Models

Components				Distance (km)				Pressure (hpa)				Wind (m/s)
A	M	O	T	6h	12h	18h	24h	6h	12h	18h	24h	6h	12h	18h	24h
				27.99	60.76	97.85	141.73	1.94	3.23	4.34	5.18	0.95	1.58	2.11	2.58
✓				25.98	49.67	74.42	105.72	1.31	2.01	2.73	3.23	0.73	1.18	1.54	1.85
✓	✓			25.45	47.28	71.16	98.03	1.39	2.11	2.81	3.44	0.77	1.21	1.58	1.86
✓	✓	✓		25.08	42.63	60.32	83.95	1.21	1.87	2.27	2.7	0.71	1.06	1.27	1.46
✓	✓		✓	26.94	47.34	70.41	100.76	1.24	1.78	2.22	2.61	0.69	1.04	1.32	1.52

TABLE IV : Comparative Results from the Ablation Study

Method	12h	24h	12h	24h	12h	24h
	Distance (km)		Pressure (hpa)		Wind (m/s)
MMSTN	73.61	173.72	3.04	5.26	1.52	2.64
MGTCF	49.72	107.80	2.27	4.05	1.23	2.06
Our	47.39	93.29	2.02	3.04	1.02	1.56

TABLE V : Comparison of different methods for abnormally deflected TC prediction

Fig. 4 : Ensemble forecasting results for abnormally deflected TCs in 2024. Historical trajectories are shown in red, actual future tracks in blue, predicted paths in green, and shaded regions denote the projected trajectory trends.

Panel of Figure 10 — Fig. 5 : Results showcase trajectory predictions using AOT-TCNet for the TC YAGI. The red dots are input, the green dots are prediction, and the blue dots are GT.

Tropical Cyclone Intensity Forecasting

See all 14 papers

Feb 28, 2026

Phys-Diff: A Physics-Inspired Latent Diffusion Model for Tropical Cyclone Forecasting

Tropical cyclone (TC) forecasting is critical for disaster warning and emergency response. Deep learning methods address computational challenges but often neglect physical relationships between TC attributes, resulting in predictions lacking physical consistency. To address this, we propose Phys-Diff, a physics-inspired latent diffusion model that disentangles latent features into task-specific components (trajectory, pressure, wind speed) and employs cross-task attention to introduce prior physics-inspired inductive biases, thereby embedding physically consistent dependencies among TC attributes. Phys-Diff integrates multimodal data including historical cyclone attributes, ERA5 reanalysis data, and FengWu forecast fields via a Transformer encoder-decoder architecture, further enhancing forecasting performance. Experiments demonstrate state-of-the-art performance on global and regional datasets.

Lei Liu, Xiaoning Yu, Kang Chen +4

University of Science and Technology of China, Hefei, China · Hefei University of Technology, Hefei, China

Tropical Cyclone Intensity Forecasting Weather Forecasting Physics-Informed Diffusion Models Latent Diffusion Models Spatiotemporal Forecasting Multivariate Time Series Forecasting

Ask PDF

Apr 14, 2026

CycloneMAE: A Scalable Multi-Task Learning Model for Global Tropical Cyclone Probabilistic Forecasting

Tropical cyclones (TCs) rank among the most destructive natural hazards, yet their forecasting faces fundamental trade-offs: numerical weather prediction (NWP) models are computationally prohibitive and struggle to leverage historical data, while existing deep learning (DL)-based intelligent models are variable-specific and deterministic, which fail to generalize across different forecasting variables. Here we present CycloneMAE, a scalable multi-task forecasting model that learns transferable TC representations from multi-modal data using a TC structure-aware masked autoencoder. By coupling a discrete probabilistic gridding mechanism with a pre-train/fine-tune paradigm, CycloneMAE simultaneously delivers deterministic forecasts and probability distributions. Evaluated across five global ocean basins, CycloneMAE outperforms leading NWP systems in pressure and wind forecasting up to 120 hours and in track forecasting up to 24 hours. Attribution analysis via integrated gradients reveals physically interpretable learning dynamics: short-term forecasts rely predominantly on the internal core convective structure from satellite imagery, whereas longer-term forecasts progressively shift attention to external environmental factors. Our framework establishes a scalable, probabilistic, and interpretable pathway for operational TC forecasting.

Renlong Hang, Zihao Xu, Jiuwei Zhao +3

School of Computer Science, School of Software, Nanjing University of Information Science and Technology, Nanjing, 210044, China. · School of Atmospheric Sciences, Nanjing University of Information Science and Technology, Nanjing, 210044, China. · Shanghai Typhoon Institute, China Meteorological Administration, Shanghai, 200030, China. +1

Tropical Cyclone Intensity Forecasting Weather Forecasting Probabilistic Forecasting Multi-Task Learning Masked Autoencoders Explainable Time Series Forecasting

Ask PDF

Aug 19, 2026

Tianmu-TC: Physics-constraints Generative Artificial Intelligence for Global Tropical Cyclone Forecasting

Tropical cyclones (TCs) pose severe risks from strong winds and heavy rainfall. However, forecasting their track and intensity remains challenging due to chaotic atmosphere and the rapid amplification of initial condition errors, leading to growing forecast uncertainty. While numerical weather prediction (NWP) and deep learning models have made progress, they remain computationally demanding and often fail under complex meteorological scenarios. Here, we present Tianmu-TC, a physics-constraints generative framework for global TC forecasting. Trained on Western North Pacific data, Tianmu-TC leverages physics-constraints to generate controllable outputs with reduced uncertainty thus improving forecast reliability. Experiments show Tianmu-TC outperforms deterministic and ensemble meteorological artificial intelligence models and authoritative NWP systems such as ECMWF in global ocean basins, with significantly lower computational cost. We further show Tianmu-TC performs well in challenging scenarios such as data sparsity, anomaly tracks, rapid intensification and weakening. These findings suggest physics-constraints generative AI offers a promising approach for reliable, efficient global TC forecasting.

Shiqi Zhang, Pan Mu, Cheng Huang +6

College of Computer Science and Technology, Zhejiang University of Technology, Hangzhou & 310023, China. · School of Control Science and Engineering, Shandong University, Jinan & 250061, China. · School of Computer Science and Engineering, Tianjin University of Technology, Tianjin & 300384, China. +1

Tropical Cyclone Intensity Forecasting Weather Forecasting Physics-Informed Generative Modeling Spatiotemporal Forecasting Probabilistic Forecasting

Ask PDF

Abstract

Tropical Cyclone Intensity Forecasting

Explore similar work

Phys-Diff: A Physics-Inspired Latent Diffusion Model for Tropical Cyclone Forecasting

CycloneMAE: A Scalable Multi-Task Learning Model for Global Tropical Cyclone Probabilistic Forecasting

Tianmu-TC: Physics-constraints Generative Artificial Intelligence for Global Tropical Cyclone Forecasting