期刊名称：Journal of Theoretical and Applied Information Technology
印刷版ISSN：1992-8645
电子版ISSN：1817-3195
出版年度：2021
卷号：99
期号：23
语种：English
出版社：Journal of Theoretical and Applied
摘要：Machine learning has been deemed to be a powerful approach in forecasting hydrological events such as flood using time-series historical data. A flood can be forecast in a manner of lead time whereby short-term forecast is up to 2 days, the medium forecast is between 2 to 10 days, and the long-term forecast is more than 10 days and several months of forecasts will have a seasonal lead time. Even though the determination of forecast lead time is normally bound with the purpose of operation i.e., daily operations or strategical, but the determination of time-series scale pattern to be input into the forecast model still impose a challenging task as it involves availability and variability of the data. Commonly, the hydrological data has a dynamic nature with non-stationary and non-linear characteristics. Therefore, it is important to choose dominant input to provide an accurate forecast. The objective of this study is to investigate the effects of different time-series scales of rainfall data from eight rainfall stations in Kelantan River towards the accuracy of forecasting water level at Kuala Krai station. Pre-processing techniques based on Mutual Information (MI) are also introduced to cater the variability of the data in finding the most dominant features as input to the forecast model. There are four scale patterns that have been investigated which consist of 7 days, 10 days, 14 days, and monthly. The forecasting analysis of all scale patterns were run against three machine learning models which are Artificial Neural Networks (ANN), Long-Short Term Memory (LSTM), and Adaptive Neuro-Fuzzy Inferences System (ANFIS) model. The results show that monthly scale pattern achieve the best performance compared to other scale patterns and LSTM is the best model for forecasting monthly water level. It indicates that longer time-series of scaled pattern may provide better forecasting accuracy and able to capture more information of the seasonal characteristics of the rainfall. Thus, it will largely benefit the flood management in reducing the flood risk and controlling its resources.
关键词：Flood Forecasting;Machine Learning;Rainfall;Time-Series Scale;Water Lev
