Rainfall Forecast under Rainy Season Based on Deep Learning and Transfer Learning
| Autor: | Pin-Xuan Chen, 陳品璇 |
|---|---|
| Rok vydání: | 2019 |
| Druh dokumentu: | 學位論文 ; thesis |
| Popis: | 107 This study uses the radar echo data and rainfall station data provided by the Central Weather Bureau to select the required range and location from the data through data pre-processing, form a data sample, and then convert the data into time-space related data. Type. Because the weather data is too large and complicated, if we train the materials according to the traditional deep learning method, it takes a lot of time. Therefore, this study uses the transfer learning method to migrate a trained model through transfer learning. To various locations to reduce the time cost of the experiment. The model used in this study is a Convolutional Neural Network model combined with a Long-term and Short-Term Memory model, which is a deep learning model with spatio-temporal relationships. It is input into the radar echoes of three locations, such as Taipei, Yu_Zhong and Shang_De_Wen, and uses the Yu_Zhong model as the migration learning. Based on the radar echo data of the past 60 minutes, the rainfall in the next 60 minutes is predicted. The experimental results show that the model error of Minzhong is the smallest, and the model error of Shangdewen is the largest. Therefore, the experiment assumes that Yu_Zhong is plain and the terrain is simple. The Shang_De_Wen is located at the mountainside, the terrain is more complicated than Yu_Zhong, and the rainfall characteristics extracted from the plain cannot be applied in the mountains. In order to verify the hypothetical content and retrain the Shang_De_Wen model, the test error can indeed be reduced, so the experiment verified that the complexity of the terrain will affect the rainfall error. |
| Databáze: | Networked Digital Library of Theses & Dissertations |
| Externí odkaz: |
|