A Study On Neural Network Based Committee Machines For Typhoon Track Forecast Model In The South China Sea

Typhoons are one of the most devastating and destructive natural disasters.The Northwestern Pacific Ocean is the most active basin for typhoons formation on the planet.Typhoons in this region often affect China,specifically the coast of the South China Sea(SCS).Because of the irresistible destructive power of typhoons,precise forecasting of the typhoon tracks remains one of the top priorities and the most effective approaches for saving the human life and property.The use of artificial neural networks(ANN)as a typhoon track statistical forecasting technique has been fairly an important hotspot for researchers because of the rapid development of the ANN.The ANN based forecasting techniques mainly with good features like non-linearity,fast speed,strong adaptivity and better robustness compared to the traditional dynamical and numerical forecasting techniques or statistical-dynamical forecasting techniques.In past decades,ensemble forecast techniques have been implemented into the operational forecast system.The work in this thesis mainly focus on the applications of ANN and committee machines on typhoon track 24 hours short-term forecasting in the South China Sea.Firstly,we propose an Bayesian neural network(BNN)based Bagging ensemble forecast model.In this model,BNN is employed to learn the history typhoon track data while control the free degree of the model.Meantime,the use of Bagging committee avoids the difficulty in component models' coefficients determination which exists in traditional ensemble forecast models.Furtherly,we propose mixture density network(MDN)based ensemble forecast models using Averaging and Bagging committee machines for multi-patterns typhoon track forecasting.In the training process of the model,we adapt a strategy that dividing the dataset into multiple chunks,each chunk is put into training set one by one after used as test set.Operational forecast experiments in the dataset of typhoon tracks of the SCS reveal that proposed ANN based ensemble forecast model using committee machines have good stability and acceptable generalization ability,which satisfied with the requirements of short term(24h)typhoon track forecasting in the SCS.