climate-weather/ai produced the result/PLoS ONE 2024 · v2
Neural network trained on a decade of records forecasts typhoon strength and position
Researchers gathered 13,495 six-hourly typhoon observations from 2009 to 2019 and trained a recurrent neural network to predict a storm's next wind speed, central pressure and centre coordinates from the three readings before it.
spectrum · one line per step, placed by what the step does · bright lines used AI
Typhoon disaster emergency forecasting method based on big data
PLoS ONE, 2024
doi:10.1371/journal.pone.0299530 · record aix-00282 v2 · checked 2026-11-05
- AI was for
- Forecasting
- Model family
- Recurrent neural network, Multilayer perceptron
- Checked by
- Held-out
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
A typhoon is a tropical cyclone in the western Pacific, a spiralling storm that can sit for days over open ocean before reaching land. Forecasters want to know two things about it: where its centre will go, and how strong it will be when it gets there. Strength is usually tracked through wind speed near the centre and through central pressure, the air pressure at the storm's core, which falls as the storm intensifies. Both are hard to predict because a typhoon's behaviour depends on conditions across a wide stretch of ocean and atmosphere, and small differences early on can move a landfall by hundreds of kilometres.
This work takes a different route from physics-based weather models, which simulate the atmosphere directly. Instead the authors looked only at the record of past storms, asking whether the recent history of a typhoon's own measurements is enough to say what the next reading will be. They collected six-hourly observations published by China's National Meteorological Observatory typhoon network, covering 2009 to 2019, with each entry listing the storm's name, the latitude and longitude of its centre, its wind speed, its direction of movement and its intensity.
Where AI came in
The forecasting itself was done by the AI. The authors used an LSTM, a type of recurrent neural network built to carry information forward along a sequence, deciding at each step what to keep and what to discard. Each network was given three consecutive readings and trained to produce the next one, with separate networks fitted for wind speed, central pressure, centre latitude and centre longitude. The data was scraped with Python, cleaned in Excel and rescaled, then split so 85% trained the networks and 15% was held back for testing. Settings such as the number of hidden units, the learning rate and the number of training rounds were chosen by hand by trying values and comparing errors.
What the network stood in for is a statistical forecasting rule: rather than someone specifying how past readings should be combined into a prediction, the relationship was learned from the records. The trained networks were then run on the held-back data and on the track of Typhoon Chan-hom. Checking the output was done without AI. The predictions were compared with the recorded observations, giving mean absolute errors of about 2.4 metres per second for wind speed, about 2.48 hectopascals for central pressure, about 1.12 degrees for longitude and about 0.51 degrees for latitude. An ordinary feed-forward neural network was trained on the same data for comparison.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
The authors crawled 13,495 six-hourly typhoon records for 2009–2019 from China's National Meteorological Observatory typhoon network and trained an LSTM recurrent network to predict the next value of wind speed, central pressure and the latitude and longitude of the typhoon centre from the three preceding values. 85% of the data was used for training and 15% for testing, and mean absolute error was compared with that of an ordinary neural network, which the paper reports favoured the LSTM for latitude and longitude. For Typhoon Chan-hom the reported mean absolute errors were about 2.4 m/s for wind speed, about 2.48 hPa for central pressure, about 1.12 degrees for longitude and about 0.51 degrees for latitude, and the paper reports the predicted track as basically consistent with the observed one. The paper also reports that the fit was best for strong tropical storms and that forecasts for tropical depressions, typhoons and strong typhoons need improvement.
How AI was used
Typhoon records were scraped from a public meteorological website with Python, cleaned in Excel, and rescaled with min–max normalisation. The resulting time series was split 85/15 into training and test sets, and a three-layer LSTM recurrent network with forget, update and output gates was fitted from scratch by back-propagation through time, using root mean square error as the loss and gradient descent with the adam solver; separate networks were fitted for wind speed, central pressure, centre latitude and centre longitude, each mapping three consecutive time steps to the following one. Hyperparameters including the number of input and hidden neurons, the gradient threshold, the iteration cap and the learning rate were selected by hand by comparing root mean square error across trial settings. The trained networks were then run over the held-out data and over the track record of one named typhoon, and an ordinary feed-forward neural network was trained on the same data as a comparison. Mean absolute error and root mean square error against the recorded observations were computed outside the model, as was a separate, non-learned descriptive analysis of typhoon counts and landfall characteristics.
The shape of the work
Structural · the record, drawn
Crawl historical typhoon records
Obtaining raw data, whether by measurement, download or retrieval.
Data crawling adopts Python technologywhere the paper describes this · verbatim
Clean and normalise data
Cleaning, filtering, normalising or labelling data already obtained.
In this paper, we choose the common normalization method: maximum and minimum standardization to process the data.where the paper describes this · verbatim
Train LSTM forecasting network
Fitting model parameters, including fine-tuning an existing model. The AI stood in for statistical model.
The training set data is input into the cyclic neural networkwhere the paper describes this · verbatim
Forecast next-step typhoon state
Running a trained model over new data to predict, classify or score. The AI stood in for statistical model.
The test set is input into the trained model to get the result data.where the paper describes this · verbatim
Compare predictions with observations
Testing outputs against ground truth.
the predicted value and the real value are compared, and the results are analyzedwhere the paper describes this · verbatim
Summarise typhoon disaster trends
Extracting understanding from model behaviour.
According to the typhoon landing data set, the characteristics of typhoons landing in Chinese provinces and cities are analyzedwhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The paper's central result is the forecast of wind speed, central pressure and centre coordinates produced by the LSTM network; the error analysis and case study are evaluations of that model's output
Therefore, this paper chooses LSTM to build a typhoon disaster prediction model.where the paper describes this · verbatim
This article uses 85% data for training and 15% data for testing.where the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- DataWhether the data are available is not stated.
- ComputeThe hardware or time used is not stated.
- How many were testedThe paper gives no count of what was tested.
- Version of LSTM recurrent neural networkWhich version of the model was used is not stated.
- Version of ordinary (BP) neural networkWhich version of the model was used is not stated.
About this article
Record aix-00282, version 2, checked by a person on 2026-11-05. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error
Cite this record: AIxSci (2026). Neural network trained on a decade of records forecasts typhoon strength and position. Record aix-00282, version 2, checked 2026-11-05. https://aixsci.org/articles/aix-00282/