Forecasting of typhoon wave based on hybrid machine learning models

Abstract

The generation of typhoon waves is associated with both marine meteorological factors and continuous dynamic time series. This study develops a hybrid multi-layer perceptron (HMLP) neural network and a hybrid genetic expression programming (HGEP) model with a switch layer to forecast the typhoon waves. The switch layer transforms the input data as vectors with specified time delay under the precondition of the physical-based essence modeled. Metocean data from 55 typhoons passing through the Fujian and Taiwan sea areas are used as training data. Typhoon Talim was tested on ten test sites in the research area, and the forecast lead time was set to 3h, 6h, 12h, and 24h. The hybrid models can forecast the significant wave height well, with RAE no more than 0.83 and RRSE no more than 0.8. The actually occurred Typhoon Lekima and Typhoon Mitag were tested and observed, and showed an agreement between test data, forecast results, and observed data. The influence factors of forecast performance are discussed. The amount of training typhoons and the similarity of training typhoon tracks to target typhoon have influence on the forecast results. The forecast performance is related to the impact intensity of the typhoon on the test sites.