Typhoon Key Parameter Identification And Forecast Assessment

Typhoons are a highly destructive natural disaster,causing significant economic losses and human casualties worldwide every year.The accurate and effective acquisition of information on key typhoon parameters is the basis of many disaster prevention and mitigation engineering practices in typhoon-prone areas.Due to the complex structure of typhoons and the large spatial and temporal scales,it is difficult to obtain adequate observation information through traditional land-based equipment.Satellite clouds image contain a wealth of atmospheric information,allowing for continuous and simultaneous observation of typhoons from a global perspective,and are therefore gaining importance in typhoon research.In addition,large-scale numerical simulation techniques have been widely developed in related fields.However,the relevant research is mainly focused on meteorological disciplines,and similar research is lacking in civil engineering disciplines.However,the difference in object and perspective between engineering applications and meteorological science makes it extremely necessary to conduct relevant research in engineering fields.In this paper,deep learning techniques are used to identify and predict key parameters of typhoons in the northwest Pacific based on satellite cloud maps and atmospheric and marine environmental data.This paper will develop an objective and automated identification of key typhoon parameters using over 200 typhoon counts in the Pacific Northwest,100,000 satellite cloud images and best record datasets of typhoons from multiple meteorological departments.A deep convolutional neural network-based typhoon fingerprinting technique is proposed.The results show that the deep convolutional neural network can accurately identify typhoon fingerprints for different ranges/number of typhoons in satellite cloud maps with an accuracy of over 96% and demonstrate the effectiveness of the image pyramid technique in solving different scale features.In addition,it is found that the model identification focuses on the main part of the typhoon,i.e.the typhoon's eye and spiral rainband area,through the heat map technique;currently there is still a large variability in the positioning of the typhoon center by different national meteorological departments.This paper proposes a typhoon centering technique based on a target detection model.The deep learning models used can be applied to different types of satellite cloud maps and have good overall performance,among which the Yolov4 model has the best performance,with a prediction accuracy of over 99% and an optimal longitude error of 0.38° and latitude error of 0.34°,respectively.This paper further employs an improved multi-classification/regression deep convolutional neural network for typhoon intensity estimation.For typhoon intensity identification,a DCNN model is proposed according to typhoon rank,central pressure and maximum wind speed as indicators.The results show that the accuracy of the model proposed in this paper is up to 94%.Finally,the model visualization technique is used to explore the features affecting typhoon intensity,where the barometric pressure model focuses on the periphery of the typhoon body,while the typhoon magnitude and wind speed models are distributed to the typhoon body itself.In order to anticipate the damage of typhoon attacks and to evaluate the future wind-resistant design,this paper presents a study on the short-term prediction and mediumand long-term typhoon hazard analysis of key typhoon parameters using deep learning techniques based on historical data of the Northwest Pacific Ocean and environmental elements such as sea surface temperature,humidity and vorticity from different meteorological models for the next 60 years.A Long and Short Memory Neural Network(LSTM)model,which can capture both long-term and short-term information,and a Convolutional Long and Short Memory Neural Network(Conv LSTM),which takes into account temporal and spatial correlation,are proposed for short-term forecast typhoon parameters.The results show that the Conv LSTM model outperforms the LSTM and other traditional forecasting methods under different advance prediction lengths,and has the advantages of high adaptability,fast computation and convenient operation,which makes it more suitable for engineering practice problems;In addition,this paper further uses deep learning models such as Conv LSTM,variational self-encoder and deep convolutional neural network to simulate the generation/number and full path of future typhoons to investigate the changes in number,generation location,track,landfall and intensity.Compared to statistical dynamics methods,the deep learning models are able to consider a large number of meteorological factors in a non-linear manner,making them more efficient and suitable for prediction and assessment of typhoon extreme winds and climate.This master's thesis is an innovative attempt to integration the disciplines of wind engineering,meteorology and artificial intelligence;The results provide an important reference basis for short-term emergency planning and medium-and long-term disaster prevention and mitigation assessment of typhoons,and are of high practical engineering value.