Analysis of rainfall and large-scale predictors using a stochastic model and artificial neural network for hydrological applications in southern Africa

Abstract

Rainfall is a major requirement for many water resources applications, including food production and security. Understanding the main drivers of rainfall and its variability in semi-arid areas is a key to unlocking the complex rainfall processes influencing the translation of rainfall into runoff. In recent studies, temperature and humidity were found to be among rainfall predictors in Botswana and South African catchments when using complex rainfall models based on the generalized linear models (GLMs). In this study, we explore the use of other less complex models such as artificial neural networks (ANNs), and Multiplicative Autoregressive Integrated Moving Average (MARIMA): (a) to further investigate the association between rainfall and large-scale rainfall predictors in Botswana, and (b) to forecast these predictors to simulate rainfall at shorter future time scales (October-December) for policy applications. The results indicate that ANN yields better estimates of forecasted temperatures and rainfall than MARIMA.

River Flooding Forecasting and Anomaly Detection Based on Deep Learning

“…Over the past two decades, the ANN model has been widely used in various fields of science and engineering. The better performance of ANN applied to rainfall prediction can also be seen in the research of Thirumalaiah et al [13], Tsai et al [5], and Kenabatho et al [14].…”

“…MLP has now become one of the most commonly-used ANN algorithms. The applications of ANN in hydrology are discussed in the following paragraph [35,14].…”

Transfer Function-Noise Modeling of the Dynamic Relationship between Botswana's Monthly Maximum Temperature and Evaporation

“…The ARMA model represented by Equation 4a is a special case of the general multiplicative Autoregressive Integrated Moving Average (ARIMA) models, of order (p,d,q)x(P,D,Q)s, that account for both the non-seasonal and seasonal dependencies in the time series. These models have the general form (Box and Jenkins, 1970;1976;Box et al, 1994;Huang and Wu, 2014;Kenabatho et al, 2015;Park and Koo, 2015):…”

“…As the number one step of the identification process, the pre-whitening procedure of Box and Jenkins (1970) involves fitting an Autoregressive Moving Average (ARMA) model to the differenced input series. Autocorrelation Functions (ACFs) and Partial Autocorrelation Functions (PACFs) were used to fit the ARMA model to the differenced input series (Box and Jenkins, 1970;1976;Lungu, 1991;Lungu and Sefe, 1991;Box et al, 1994;Bierkens and Knotters, 1999;Jain and Lungu, 2003;Lungu et al, 2003;Huang and Wu, 2014;Kenabatho et al, 2015;Park and Koo, 2015). The 95% confidence limits of the autocorrelation and partial autocorrelations functions were approximated by ±2/n 1/2 (Harvey, 1993;Diggle and Chetwynd, 2011).…”

“…It is also a Ramsar and a World Heritage Site. The models constructed in this study can be used to predict the output series (evaporation) with a lead time of a few months, which will be valuable information, mainly to those responsible for the management of the limited water resources in Botswana, agriculture and other related works (Manzione et al, 2010;Kenabatho et al, 2015).…”