|
Machine Learning Model Development for Water Level Forecasting at P.1Station, Chiang Mai Province |
|---|---|
| รหัสดีโอไอ | |
| Creator | Supachai Mukdasanit |
| Title | Machine Learning Model Development for Water Level Forecasting at P.1Station, Chiang Mai Province |
| Contributor | Tawee Chaipimonplin |
| Publisher | Rajamangala University of Technology Lanna |
| Publication Year | 2568 |
| Journal Title | RMUTL Engineering Journal |
| Journal Vol. | 10 |
| Journal No. | 2 |
| Page no. | 34-48 |
| Keyword | Water Level Forecast, Machine Learning, Artificial Neural Network, Exponential MovingAverage |
| URL Website | https://engsystem.rmutl.ac.th/journal/ |
| ISSN | 3027-7426 |
| Abstract | This research applies machine learning models to forecast water levels at the P.1 station (Nawarat Bridge), Chiang Mai Province, both for 6 and 9 hours in advance. The objectives are to identify suitable variables and to create models for forecasting water levels at the P.1 station. The study utilizes historical hourly water level data from the P.1 and P.67 stations, combined with Moving Average (MA) and Exponential Moving Average (EMA) data covering the years from 2017 to 2024, which has amounted to a total of 66,180 records. The dataset is divided into a training set (80%) and a testing set (20%). The experiment design involves creating artificial neural network models based on historical data from one station (P.1) and two stations (P.1 and P.67). The models consist of those using only historical data, those using historical data combined with MA, and those using historical data combine with EMA, resulting in a total of 12 models. The structure of each model was optimized to achieve the best forecasting results. The results indicate that the best model for the 6-hour forecasting is the P.1_6 + P.67_6 + EMA model. This model utilizes 18 input variables, with 6 and 2 nodes in the first and second hidden layers, respectively, and 1 output node. This model achieved a Mean Absolute Error (MAE) of 0.0405, a Root Mean Square Error (RMSE) of 0.0578, and a coefficient of determination (R²) of 0.9859. For the 9-hour forecasting, the best model is the P.1_9 + P.67_9 + EMA model, which also employs 18 input variables, with 5 and 4 nodes in the first and second hidden layers, respectively, and one output node. This model achieved a MAE of 0.0562, an RMSE of 0.0776, and an R² of 0.9746. Both models utilize data from two stations combined with EMA. |
