|
Short Term Precipitation Forecasting using Recurrent Neural Networks;a Case Study ofSuvarnabhumi Airport |
|---|---|
| รหัสดีโอไอ | |
| Creator | Ragkana Phooseekhiwe |
| Title | 1. Short Term Precipitation Forecasting using Recurrent Neural Networks 2. a Case Study ofSuvarnabhumi Airport |
| Contributor | Suronapee Phoomvuthisarn |
| Publisher | The Association of Council of IT Deans (CITT) |
| Publication Year | 2565 |
| Journal Title | Journal of Information Science and Technology |
| Journal Vol. | 12 |
| Journal No. | 1 |
| Page no. | 13-26 |
| Keyword | Long Short-Term Memory, Predictive analytics, Rainfall prediction, Recurrent Neural Network, Neural Network, Gated Recurrent Unit, Imbalanced data |
| URL Website | https://tci-thaijo.org/index.php/JIST |
| Website title | Journal of Information Science and Technology |
| ISSN | 2651-1053 |
| Abstract | Rainfall is one of the keyimportant factorsaffectinghuman life. Accurate precipitation forecastingallowshumans to better preparefor various activities that will happen in the future.However,in some situations, the availability of weather data is limited, makingit difficultto make accurate forecasting.Currently, much researchhas chosendeep neural networksas an algorithm to train forecasting models. The main idea is to come up with relevant features at the architecture level. Based on this paradigm, it has been shown that the appropriatedeep neural network architecture can flexibly mix and match features that are relevant in predictions. Consequently, most of theexisting research then focuseson some of the techniques to improve the performance of the modelswithout paying much attention on the issues of adding relevant features. However, when the training data is limited, deep neural networks might not scale very well, thus making it difficult to mix and match features for predictions. This imposes a research question of how effective the proposed forecasting models might be when not adding relevant features to the models. Theaims of this research are to develop and compare the efficiencyof various models for short-term precipitation forecasting with and without adding relevant features. The experiment consists of 2 parts: (1) The experiment by exploringwhether themodels with relevant featured providedcan achieve higher accuracy compared to original models in equivalent environments, and (2) The experiment by comparing accuracybetween 5models (ARIMA, ARIMAX, RNN, LSTM and GRU). The weather dataset, which is very limited in quantity,used in this research wascollected from Suvarnabhumi Airport. The results show that adding relevant features can enhance the forecasting performance of the modelwhen the weather data is limited. In addition, the GRU model with relevantfeatured provided is the most effective prediction. |
