Coupling Satellite Rainfall Estimates and Machine Learning Techniques for Flow Forecast: Application to a Large Catchment in Southern Africa - Proceedings of 2013 IAHR World Congress

Aug 3, 2026

Upload by Ophelia

Accurate river flow forecasting supports stream and reservoir management, delivering social, economic, and ecological benefits. This work couples satellite rainfall estimates with machine learning for daily flow forecasting, evaluating lead times of 30 and 60 days at Victoria Falls in Southern Africa. Six machine learning models are compared against optimized ARMA and a Fourierseries benchmark. Rainfall inputs generally improve machine learning performance at 30 days, but not at 60 days. Traditional ARMA cannot utilize rainfall information, and at 60 days machine learning shows clear advantages.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages12

Reading Time30 min

AccessFree to download

1 Introduction

Background and data-driven modeling approaches

Study objectives and forecasting setup

2 Data

Discharge data and upstream stations

View metadata, citation and similar [papers at ](papers at core.ac.uk)[core.ac.uk](papers at core.ac.uk) brought to you by CORE provided by Infoscience- École polytechnique fédérale de Lausanne Proceedings of 2013 IAHR World Congress Coupling Satellite Rainfall Estimates and Machine Learning Techniques for Flow Forecast: Application to a Large Catchment in Southern Africa José P. Matos Ph.D. student, Laboratoire de Constructions Hydrauliques (LCH) École Polytechnique Fédérale de Lausanne (EPFL), Switzerland; Instituto Superior Técnico (IST), TechnicalUniversity of Lisbon, [Portugal. Email: jose.matos@epfl.ch](Portugal. Email: jose.matos@epfl.ch) Théodora Cohen Liechti Ph.D. student, Laboratoire de Constructions Hydrauliques (LCH) École Polytechnique Fédérale de Lausanne (EPFL),Switzerland. Email: [theodora.cohen@epfl.ch](theodora.cohen@epfl.ch) Maria M. Portela Professor, Instituto Superior Técnico (IST), TechnicalUniversityofLisbon, Portugal. [Email: mps@civil.ist.utl.pt](Email: mps@civil.ist.utl.pt)[ ](Email: mps@civil.ist.utl.pt)Anton J. Schleiss Full Professor, Laboratoire de Constructions Hydrauliques (LCH) École Polytechnique Fédérale de Lausanne (EPFL), Switzerland. Email: anton.schleiss@epfl.ch ABSTRACT: Accurate river flow forecasting is an important asset for stream and reservoir management, being often translated into substantial social, economic and ecological gains. This contribution aims at coupling satellite rainfall estimates and machine learning techniques for daily flow forecast. Two lead times, of 30 and 60 days, were tested for flows at Victoria Falls, in Southern Africa. Six distinct machine learning models were compared with optimized ARMA models and benchmarked against a Fourierseries approximation. Results show that the addition of rainfall data generally enhanced the performances of machine learning models at 30 days but did not improve forecasts at 60 days. Also, it was shown that traditional ARMA models do not make use of the rainfall information. Regarding a lead time of 60 days, the machine learning models appear to bear great advantages compared to ARMA models which, for such a lead time have shown practically no forecast capabilities. KEY WORDS: Artificial Neural Networks,Flow forecast, Kariba, Support-Vector Regression, Zambezi. 1 INTRODUCTION Accurate river flow forecasting is an important asset for stream and reservoir management, being often translated into substantial social, economic and ecological gains. In the past, considerable advances have been accomplished on this subject, ranging from the development of physical distributed and lumped conceptual approaches to data-driven models. In the last decades, a wealth of alternative data-driven models has been proposed among which autoregressive moving-average (ARMA)(e.g. Anderson, 1977; Mohammadi et al., 2006 ; Valipour et al., 2013), one of the most popular times series models for reservoir design and operation (Wang et al., 2009), autoregressive integrated moving-average (ARIMA)(e.g. Carlson et al., 1970), which is a non-static generalization of the ARMA model (Valipour et al., 2013), artificial neural networks (ANN)(Abrahart and See, 2000; Cigizoglu, 2005; Shamseldin and O'Connor, 2001), genetic programming (GP)(Londhe and Charhate, 2010; Wang et al., 2009), support vector regression (SVR)(e.g. Lin et al., 2006 ; Wang et al., 2009), and k-nearest neighbors (KNN) (e.g. Sivakumar et al., 2002)are examples. The performances reported in literature vary greatly(Wu and Chau, 2010) . Developed in the framework of the African Dams Project (ADAPT)(Mertens et al., 2013), a flow forecast system is proposed and evaluated for the Zambezi River at Victoria Falls. This section drains the over 360 000 km2 UpperZambezi catchment. Not far downstream that section laysKariba, the World’s largest artificial reservoir by volume, whose operation greatly influences the Zambezi basin’s economy and ecosystem. Characterized by its

FAQs

What is the main goal of the study on flow forecasting?

To couple satellite rainfall estimates with machine learning techniques to improve daily river flow forecasts for lead times of 30 and 60 days.

How do satellite rainfall inputs affect model performance at different lead times?

Rainfall data generally enhance machine learning performance at 30 days, but do not improve forecasts at 60 days.

Why do traditional ARMA models underperform in this framework?

ARMA models do not use rainfall information, so they cannot benefit from the added satellite rainfall estimates.

Aug 3, 2026

Upload by Ophelia