基於循環神經網路的颱風路徑模擬初步研究

如今的資料數據量龐大，同時資料的複雜性也激增，對於處理資料時經常耗時過長的問題，近年來逐漸興起的機器學習技術成為了主流的解決方式，同樣，在土木工程領域，使用機器學習，也就是人工智慧化的方式來對資料進行分析處理，是未來的必經之路。本研究目的為利用循環神經網路(RNN)中可預測時間序列及長短期記憶(LSTM)模型具有長期記憶的特性，將之應用在颱風領域上，以模擬一個颱風生成後的移動路徑。本研究以颱風聯合警報中心（JTWC）觀測之颱風資料為時間序資料來源，首先對所有資料標準化處理，然後以循環神經網路來訓練學習所有資料，找出他們的特徵與關聯性，以此來預測颱風移動路徑中時間步上颱風移動速度、方位角、氣壓差的變化，最後以 RMSE 與 loss 值作為模型準確程度的判斷依據，進而建立一套可以在颱風領域應用的模擬模型。本模型主要由輸入層、3 層 LSTM 層、全連接層（Dense 層）、融合層（Merge 層）及輸出層所構成，原始颱風資料經過標準化處理後進入輸入層，接著進入 LSTM 層，通多 LSTM 的輸入閥、遺忘閥、輸出閥的運算後，即得到了 LSTM 層的輸出值以及記憶在 LSTM 層中的長短期記憶矩陣，這個長短期記憶矩陣會持續的影響後面的 LSTM 層，一直到最後的一層 LSTM 輸出到全連接層，重新擬合後輸出到融合層，最後到輸出層，完成模擬。本模型在透過 RMSE 比較後，得出以 RNN 模型預測單個颱風移動路徑數據中，模擬方位角與模擬颱風移動速度的結果都較為優秀，而颱風氣壓差數據較差。對於完全模擬颱風移動速度與方位角有優秀的結果，颱風氣壓差的結果則比較差。綜上所述，本模型在颱風模擬領域具有一定的潛力，可用於西太平洋生成的颱風之路徑模擬、對於台灣耐風規範研究起到助力。

關鍵字

深度學習；循環神經網路；長短期記憶模型；聯合颱風警報中心資料

並列摘要

Nowadays, the amount of data is huge, and the complexity of data is also surging. For the problem of time-consuming data processing, machine learning technology has become the mainstream solution in recent years. Similarly, in the field of civil engineering, use of machine learning, which is the artificial intelligence way of data analysis and processing, is the only way in the future. The purpose of this study is to use Long Short Term Memory (LSTM) model of Recurrent Neural Network (RNN), which has the ability to predict time series data and preserve long term memory, to simulate the moving path of typhoon. In this study, we used the typhoon data observed by JTWC as the source of time series data. Firstly, we standardized all the data, and then RNNs were trained using the standardized data to find out their characteristics and correlations, so as to predict the changes of typhoon speed, azimuth and pressure difference at every time steps along typhoon paths, Finally, the RMSE and loss values were used to judge the accuracy of the model, and then a set of simulation models that can be applied in the field of typhoon was established. This model is mainly composed of an input layer, three LSTM layers, a fully-connected layer, a merge layer and an output layer. The original typhoon data enters the input layer after standardized processing, and then enters the LSTM layer. After calculating, the input valve, forgetting valve and output valve of LSTM, the output values of LSTM layer and the long-term and short-term memory matrix in LSTM layer are obtained. This long-short-term memory matrix will continue to affect the later LSTM layer until the last LSTM layer is output to the fully connected layer. From there values output to the fusion layer after refitting, and finally to the output layer to complete the simulation. Three RNNs were trained individually to simulate the three subject values mentioned previously and their accuracies were compared according to their RMSEs. The results of the simulated azimuth angle and the simulated typhoon movement speed are better, while the typhoon air pressure difference data is poor. For the complete simulation, the typhoon movement speed and azimuth angle have good results, and the air pressure difference results are slightly worse. In summary, this model has certain potential in the field of typhoon simulation. It can be used to simulate the path of typhoons generated in the western Pacific Ocean, and contribute to the study of Taiwan's wind resistant design regulations.