Inflow Forecasting of Pa Phayom Reservoir using Genetic Programming

Abstract

Accuracy in reservoir inflow forecasting is significantly crucial for reservoir operation. This article presents the application of Genetic Programming (GP), which relies on genetic evolution and makes computer self-programming, for reservoir inflow forecasting at Pa Phayom reservoir for 1 day, 1 week, and 1 month ahead, respectively. Daily, weekly, and monthly data of rainfall and reservoir inflow at Phayom reservoir were collected for 10 years from 2007 to 2016 and then they were used as input data for 4 case studies in each time span of forecasts. For all study cases, the data sets were divided into two parts, i.e. the first 70% of data sets were for the training process and the other 30% of data sets were for the testing process. Models’ performance was evaluated using 3 statistical indices: correlation coefficient (r), root mean squared error (RMSE), mean absolute error (MAE), and combined accuracy (CA). Varying each GP’s parameter for studied data sets was conducted to determine the optimal parameters. It was found that 1 day ahead reservoir inflow forecasting gave a satisfying performance with values of r (0.878 and 0.069), RMSE (0.031 MCM/day and 0.883 MCM/day), MAE (0.058 MCM/day and 0.024 MCM/day), and CA (0.110 and 0.101) for training and testing processes, respectively. The reservoir inflow forecasting for 1 week and 1 month ahead showed acceptable results with correlation coefficient (r) between 0.5 and 0.7 for training and testing processes and gave values of RMSE and MAE close to those values of 1 day ahead reservoir inflow forecasting, However, they could not forecast reservoir inflow during floods optimally. The model should be kept up to date as more information is added.