Combined Multiple Neural Networks and Genetic Algorithm with Missing Data Treatment: Case Study of Water Level Forecasting in Dungun River - Malaysia

摘要

We consider water level forecasting in Dungun River where the collected data contain missing values. Therefore, we cannot utilize a prediction technique to forecast the water level directly. To overcome this difficulty, we used Ordinary Linear Regression (OLR) and mean substitution to handle the imperfect data and to make the data meaningful. ARIMA and SARIMA are well known techniques and widely used in time series forecasting. Unfortunately, they produce a linear regression model that may improper model for water level forecasting. Instead, Backpropagation Neural Network (BPNN) and Nonlinear Autoregressive Exogenous Model (NARX) are alternative techniques to address the issue of linearity in regression. Nevertheless, they also have difficulties to determine the optimal network and regression coefficients/weights due to the randomness of their initial weights. Under this circumstance, we proposed Multiple BPNN and Genetic Algorithm (GA) to overcome the limitation of ARIMA/SARIMA, standalone BPNN and NARX. Our experiment showed that our proposed technique is superior compared to ARIMA, SARIMA, BPNN and NARX.