A NOVEL METHOD TO ESTIMATE HIDDEN NEURONS IN ENSEMBLE FLOOD FORECASTING MODEL

Machine learning model such as neural networks have been widely adopted to provide flood forecast. Self-adaptability in neural networks enable them to learn pattern by their own and adjusted the connection between the neurons, but still producing error. In many neural network adaptations for flood forecasting, the number of hidden neurons is normally randomly selected which can caused overfitting problem in the network. In this study, a novel method to estimates the hidden neuron is proposed to overcome this problem. This method integrates the evaluation of various convergence theorem criteria with grid search to estimates the hidden neuron. By having this integration, optimal number of fix hidden neurons can be determined. This method is used in the ensemble model that based on neural networks to forecast the water level based on rainfall data. Based on the performance measurement using Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and Nash Sutcliffe Efficiency (NSE), it is found that the integration of convergence theorem and grid search can be used to fix the number of hidden neuron and has reduce the error that led to overfitting in the forecast.