Combining statistical machine learning models with ARIMA for water level forecasting: The case of the Red river. (August 2020)

• Record Type: Journal Article
• Authors: Phan, Thi-Thu-Hong; Nguyen, Xuan Hoai
• Is Part Of: Advances in water resources. Volume 142(2020)
• Publication Date: 2020-08
• Subjects: Water level forecasting -- Hybrid model -- Statistical machine learning -- ARIMA -- Long-term univariate time series; Hydrology -- Periodicals; Hydrodynamics -- Periodicals; Hydraulic engineering -- Periodicals; 551.48
• DOI: 10.1016/j.advwatres.2020.103656
• Languages: English
• ISSNs: 0309-1708

Abstract

Highlights: A novel approach that combines ARIMA and Statistical Machine Learning Models for water level forecasting. A new data driven transformation of univariate time series into multi-dimensional data. Multi-step ahead forecasting based on one-step ahead. Testing and comparing forecasting methods on 3 real big datasets. Abstract: Forecasting water level is an extremely important task as it allows to mitigate the effects of floods, reduce and prevent disasters. Physically based models often give good results but they require expensive computational time and various types of hydro-geomorphological data to develop the forecasting system. Alternatively, data driven forecasting models are usually faster and easier to build. During the past decades, statistical machine learning (ML) methods have greatly contributed to the advancement of data driven forecasting systems that provide cost-effective solutions and better performance. Meanwhile, Autoregressive integrated moving average (ARIMA) is one of the famous linear statistical models for time series forecasting. In this paper, we propose a hybrid approach that takes advantages of linear and nonlinear models. The proposed method combines statistical machine learning algorithms and ARIMA for forecasting water level. Observed water level of the Red river at the Vu Quang, Hanoi (3 hourly sampled from 2008 to 2017) and Hung Yen hydrological stations (hourly collected data from 2008 to 4/2015) are used to evaluate the performance of different methods. Experimental results on these 3 real big datasets show the effectiveness of our proposed hybrid models.