Hybrid particle swarm optimization and group method of data handling for short-term prediction of natural daily streamflows

Hydrological forecasts have been developed since the earliest civilizations allowing to plan actions such as agriculture, grain storage, and the construction of reservoirs to supply water during long periods of drought. These forecasts are becoming increasingly essential given the growing dependence on water resources in the most diverse activities such as hydroelectric power generation. In this study, we develop a hybrid approach to forecasting the daily flow of the Zambezi River at the Cahora Bassa dam in Mozambique. These forecasts use daily historical data on flow, evaporation, relative humidity, and rainfall to predict the weather one day in advance. The model employs the seven past days as inputs to the Group Method of Data Handling (GMDH) algorithm optimized by the Particle Swarm Optimization (PSO) algorithm. GMDH is a neural network composed of neurons arranged in several layers, consisting of polynomial functions with two variables combined in a cascade to produce the output at the end of the network. The PSO promotes a search for optimal GMDH parameters to minimize the error values between the predicted and the actual river flow observed values. The simulations are performed 25 times to reduce the effects of the random values characteristic of the tested models. The results obtained by the proposed approach are compared with the other neural networks, such as extreme learning machine (ELM) and Multi-layer perceptron (MLP). The models’ performances were compared using five metrics, statistical tests, and uncertainty analysis. These results show that the GMDH model produced better flow prediction capability than the other two models.