HYBRID TECHNIQUE BETWEEN DESIGN OF EXPERIMENTS AND ARTIFICIAL NEURAL NETWORKS FOR RAINFALL-RUNOFF MODEL CALIBRATION METHOD

Calibration is one of standard procedures to be conducted before the application of hydrology models. Some rainfall-runoff models have many model parameters which cause more difficulties and require longer time in the calibration processes. This paper illustrates the application of a proposed hybrid technique between Design of Experiments (DOE) and Artificial Neural Networks (ANN) for calibrating the parameters of rainfall-runoff models. The DOE is used to select the appropriate sample experiments based on the range of model parameters, and the ANN is used to optimize the value of model parameters. A Mock rainfall-runoff model was used to illustrate the application of the proposed technique. As the model has six model parameters, the model calibration requires 32 runs in the linear full factorial design experiments or 77 runs in the curvature full factorial design experiments. The error back-propagation technique (BP) was utilized in this approach to synthesize the suitable networks for reflecting the relationships between goodness-of-fit criterion, the sum of absolute errors as the inputs and model-parameters as the outputs. Standard statistical techniques of goodness of fit, such as the Nash-Sutcliffe Efficiency, NSE and the sum of absolute error, |E| were used to measure the differences between simulated and observed runoffs. Observed runoff and climatic data, including rainfall from 1990 to 2010 for the Babak River Basin in Lombok, Indonesia were used in the calibration process; while, data from 2011 to 2016 were used for verification of the model. The results indicate the proposed technique gave more accurate calibrated parameters than the trial-and-error method. In addition, the proposed method requires less time for model calibration. The application of the proposed technique is not limited for calibrating rainfall-runoff models; however, it can be used to calibrate any kind of mathematical models.