Comparison of Machine Learning Algorithms for Daily Runoff Forecasting with Global Rainfall Products in Algeria

Rainfall–runoff models are crucial tools for managing water resources. The absence of reliable rainfall data in many regions of the world is a major limitation for these models, notably in many African countries, although some recent global rainfall products can effectively monitor rainfall from space. In Algeria, to identify a relevant modeling approach using this new source of rainfall information, the present research aims to (i) compare a conceptual model (GR4J) and seven machine learning algorithms (FFNN, ELM, LSTM, LSTM2, GRU, SVM, and GPR) and (ii) compare different types of precipitation inputs, including four satellite products (CHIRPS, SM2RAIN, GPM, and PERSIANN), one reanalysis product (ERA5), and observed precipitation, to assess which combination of models and precipitation data provides the optimal performance for river discharge simulation. The results show that the ELM, FFNN, and LSTM algorithms give the best performance (NSE > 0.6) for river runoff simulation and provide reliable alternatives compared to a conceptual hydrological model. The SM2RAIN-ASCAT and ERA5 rainfall products are as efficient as observed precipitation in this data-scarce context. Consequently, this work is the first step towards the implementation of these tools for the operational monitoring of surface water resources in Algeria.