[Paper Review] Intensity Prediction of Tropical Cyclones using Long Short-Term Memory Network

Koushik Biswas, Sandeep Kumar+1 more|arXiv (Cornell University)|Jul 7, 2021

Tropical and Extratropical Cyclones Research12 citations

TL;DR

This paper proposes a stacked bidirectional Long Short-Term Memory (BiLSTM) network to predict tropical cyclone intensity in the North Indian Ocean up to 72 hours in advance, using historical meteorological data including wind speed, pressure, sea surface temperature, and storm position. The model achieves a mean absolute error (MAE) of 1.52 knots for 3-hour predictions and 11.92 knots for 72-hour predictions, demonstrating high accuracy and robustness in forecasting both gradual and sudden intensity changes.

ABSTRACT

Tropical cyclones can be of varied intensity and cause a huge loss of lives and property if the intensity is high enough. Therefore, the prediction of the intensity of tropical cyclones advance in time is of utmost importance. We propose a novel stacked bidirectional long short-term memory network (BiLSTM) based model architecture to predict the intensity of a tropical cyclone in terms of Maximum surface sustained wind speed (MSWS). The proposed model can predict MSWS well advance in time (up to 72 h) with very high accuracy. We have applied the model on tropical cyclones in the North Indian Ocean from 1982 to 2018 and checked its performance on two recent tropical cyclones, namely, Fani and Vayu. The model predicts MSWS (in knots) for the next 3, 12, 24, 36, 48, 60, and 72 hours with a mean absolute error of 1.52, 3.66, 5.88, 7.42, 8.96, 10.15, and 11.92, respectively.

Motivation & Objective

• To improve long-term tropical cyclone intensity forecasting in the North Indian Ocean, where cyclones cause significant societal and economic damage.
• To address the limitations of existing statistical and numerical models, which often lack accuracy and computational efficiency.
• To develop a deep learning model capable of predicting continuous changes in Maximum Sustained Surface Wind Speed (MSWS) over multiple future time points.
• To enable early and accurate warnings by forecasting MSWS up to 72 hours in advance with high temporal resolution.
• To demonstrate that BiLSTM networks can effectively capture complex, non-linear temporal dependencies in cyclone evolution from sequential meteorological features.

Proposed method

• A stacked bidirectional Long Short-Term Memory (BiLSTM) network is used to model sequential meteorological time series data from tropical cyclones.
• Input features include latitude, longitude, MSWS, estimated central pressure, distance, direction, and sea surface temperature (SST) at regular time intervals.
• The model is trained to predict future MSWS values for up to 72 hours ahead, using a sequence of past observations as input.
• The architecture leverages bidirectional processing to capture both past and future context in the time series, improving temporal representation learning.
• Model training uses mean absolute error (MAE) and root mean square error (RMSE) as optimization objectives, with 5-fold cross-validation for evaluation.
• The model is evaluated on two recent cyclones, Fani and Vayu, which were not included in the training data, to test generalization.

Experimental results

Research questions

• RQ1 Can a BiLSTM-based model predict tropical cyclone intensity with high accuracy over extended future time horizons (up to 72 hours)?
• RQ2 How does the model perform in capturing sudden changes in cyclone intensity, such as rapid intensification or stabilization?
• RQ3 Does increasing the number of past time points used for input significantly improve prediction accuracy for future MSWS?
• RQ4 Can the model generalize well to cyclones not seen during training, such as Fani and Vayu?
• RQ5 How does the model's performance compare to existing statistical and numerical methods in terms of MAE and RMSE across multiple future time points?

Key findings

• The model predicts MSWS with a mean absolute error (MAE) of 1.52 knots for 3-hour forecasts, increasing to 11.92 knots for 72-hour forecasts.
• For predictions up to 24 hours ahead, the MAE remains below 6 knots, which corresponds to a 15-knot range in MSWS classification, enabling reliable grade prediction (≥3) for cyclonic storms.
• The model maintains high accuracy even for longer horizons (e.g., 72 hours), with MAE increasing gradually rather than sharply, indicating stable long-term forecasting capability.
• The model effectively captures both increasing and stabilizing intensity trends, as demonstrated by accurate predictions on cyclones Fani and Vayu, even when intensity patterns in training data were inconsistent.
• Increasing the input sequence length (T₁) from 4 to 8 time points results in minimal improvement in MAE (1.54 vs. 1.52 for T₂=1), indicating that the model learns most from early data points, reducing computational overhead.
• The model outperforms single-point prediction methods by providing combined accuracy across multiple future time points, making it more suitable for operational early warning systems.