520 |
|
|
|a The transformation of rainfall into runoff involves many highly complex hydrological components that require various hydrological data and topographical information. These data are hard to obtain and not consistent. Therefore, hydrologic tank and artificial neural networks models that require only rainfall and runoff data were proposed. The selected study area is Bedup Basin, Sarawak, Malaysia, a rural catchment in humid region. A new global optimization method named as particle swarm optimization (PSO) was proposed, and compared with shuffle complex evolution and genetic algorithm techniques for calibrating the tank models' parameters automatically. PSO is also hybrid with neural network to form particle swarm optimization feedforward neural network (PSONN) to overcome the slow convergence rate and trapping at local minima problems. PSONN performance is then compared with multilayer perceptron and recurrent networks, that used backpropagation algorithm. Models performances are measured using coefficient of correlation (R) and Nash-Sutcliffe coefficient (E2). Generally, artificial neural networks performance is slightly better than tank model. Results of tank model calibration indicate that PSO method appeared to be the best based on its robustness, reliability, efficiency, accuracy and smallest variability in boxplots. Shuffle complex evolution follows as the second best and the third best is genetic algorithm for both daily and hourly runoff simulation. Among multilayer perceptron, recurrent and PSONN investigated, recurrent network forecasts daily and hourly runoff most accurately, followed second best by multilayer perceptron and lastly PSONN. PSONN has proven its remarkable capability to simulate daily and hourly runoff with an acceptable accuracy. This study revealed that artificial intelligence methods especially PSO, have offered a real prospect for an efficient, simple, cheaper, more flexible, and well suited to model flood processes.
|