14:00 〜 14:15
[AAS02-08] Integrating Encoder-Decoder Neural Networks and Multi-Model Guidance for Typhoon Track Forecast Uncertainty Estimation
Monte Carlo approaches are widely used to estimate uncertainties in tropical cyclone (TC) track forecasts. However, these methods often struggle to capture situation-dependent and spatiotemporal error correlations. To address these limitations, this study proposes an artificial intelligence (AI)-based approach using a Recurrent Neural Network (RNN) with an Encoder-Decoder architecture. The model utilizes data from the Central Weather Administration’s (CWA) official TC forecasts, as well as ensemble forecasts from global (i.e., ECMWF and NCEP) and regional (i.e., WRF) weather prediction models.
Our preliminary results show that the RNN-based method effectively captures scenario-specific uncertainty variations. For instance, tropical cyclones in mid-to-high latitudes with faster translation speeds exhibit smaller cross-track forecast errors. The prediction intervals (PIs) generated by the model encompass approximately 70% and 95% of observed errors within ±1 and ±2 standard deviations, respectively. Incorporating large-scale environmental indices, such as steering flow and monsoon circulation, further improves the accuracy of uncertainty estimates. These findings underscore the potential of AI-based techniques in enhancing TC track forecast uncertainty estimation and improving the reliability of operational TC track forecasting. Detailed results will be presented during the conference.
Our preliminary results show that the RNN-based method effectively captures scenario-specific uncertainty variations. For instance, tropical cyclones in mid-to-high latitudes with faster translation speeds exhibit smaller cross-track forecast errors. The prediction intervals (PIs) generated by the model encompass approximately 70% and 95% of observed errors within ±1 and ±2 standard deviations, respectively. Incorporating large-scale environmental indices, such as steering flow and monsoon circulation, further improves the accuracy of uncertainty estimates. These findings underscore the potential of AI-based techniques in enhancing TC track forecast uncertainty estimation and improving the reliability of operational TC track forecasting. Detailed results will be presented during the conference.