17:15 〜 19:15
[AAS02-P03] Application of AI Techniques to Develop a Probabilistic and Region-Specific Week-2 Typhoon Formation Index Based on Large-Scale Environmental Factors
キーワード：tropical cyclone, formation, probabilistic forecasting
This study investigates the relationships between large-scale environmental factors and the formation of typhoons over a timescale of 1 to 2 weeks. Additionally, artificial intelligence (AI) methods are applied to develop a probabilistic model for predicting typhoon formation.
The large-scale environmental factors considered in this study include the Western North Pacific Monsoon Index (WNPMI), sea surface temperature (SST), Madden-Julian Oscillation (MJO), and Boreal Summer Intraseasonal Oscillation (BSISO). These factors are analyzed to examine their relationships with typhoon formation over a 1-2 week period. Subsequently, an AI-based probabilistic typhoon formation prediction model is developed using the Encoder-Decoder Long Short-Term Memory (LSTM) framework.
This study utilizes observations from 2002 to 2021. Given the relatively low frequency of typhoons, the Binary Focal Cross-Entropy (BFCE; Lin et al., 2018) loss function is employed to mitigate the impact of data imbalance on model training.
First, a baseline model for the entire Western North Pacific (WNP) basin is established. Then the entire WNP basin is divided into five sub-regions, with probabilistic forecasts provided for each sub-region, aiming to capture the influence of large-scale environmental indicators on typhoon formation in terms of timing and location over the next two weeks.
Preliminary test results indicate that using BFCE as the loss function increases the frequency of typhoon formation predictions, resulting in more high-probability forecast values. Additionally, the Reliability Diagram aligns more closely with the diagonal line. Test results for the five sub-regions indicate that typhoon formation primarily occurs to the east of the Philippines, where the forecast skill is higher. Regardless of whether the WNP basin is divided into sub-regions during model development, the typhoon formation forecast skill for this area is the highest, with an Area Under the ROC Curve (AUC) of approximately 0.84, indicating good discriminatory ability. Future work will involve a more detailed examination of the impact of each indicator on typhoon formation. Additional findings will be presented at the conference.
The large-scale environmental factors considered in this study include the Western North Pacific Monsoon Index (WNPMI), sea surface temperature (SST), Madden-Julian Oscillation (MJO), and Boreal Summer Intraseasonal Oscillation (BSISO). These factors are analyzed to examine their relationships with typhoon formation over a 1-2 week period. Subsequently, an AI-based probabilistic typhoon formation prediction model is developed using the Encoder-Decoder Long Short-Term Memory (LSTM) framework.
This study utilizes observations from 2002 to 2021. Given the relatively low frequency of typhoons, the Binary Focal Cross-Entropy (BFCE; Lin et al., 2018) loss function is employed to mitigate the impact of data imbalance on model training.
First, a baseline model for the entire Western North Pacific (WNP) basin is established. Then the entire WNP basin is divided into five sub-regions, with probabilistic forecasts provided for each sub-region, aiming to capture the influence of large-scale environmental indicators on typhoon formation in terms of timing and location over the next two weeks.
Preliminary test results indicate that using BFCE as the loss function increases the frequency of typhoon formation predictions, resulting in more high-probability forecast values. Additionally, the Reliability Diagram aligns more closely with the diagonal line. Test results for the five sub-regions indicate that typhoon formation primarily occurs to the east of the Philippines, where the forecast skill is higher. Regardless of whether the WNP basin is divided into sub-regions during model development, the typhoon formation forecast skill for this area is the highest, with an Area Under the ROC Curve (AUC) of approximately 0.84, indicating good discriminatory ability. Future work will involve a more detailed examination of the impact of each indicator on typhoon formation. Additional findings will be presented at the conference.