A Study on Real-Time Forecasting of Reservoir Inflow Based on Artificial Neural Network

For the most effective operation of multi-purpose reservoir at flood period, the forecasting of inflow must be preceded and a rainfall-runoff modeling is necessary for the forecasting of inflow. However, the rainfall-runoff process is nonlinear and complex so many errors can be occurred by uncertain parameter estimation in modeling procedure. In this study, a neural network theory was adopted for modeling rainfall-runoff process, and a real-time inflow forecast system was developed. The models developed in this study were based on the back-propagation algorithm and Cascade-Correlation algorithm for learning. We applied these models to Soyangang River basin, so we could get forecasted inflow values — 1 hour, 3 hour and 6 hour preceding inflows. In case of the back-propagation algorithm, many trials are required to find out the optimum structure, but Cascade-Correlation algorithm can make the optimum neural network structure automatically at a time. We applied this model to the August '95 flood event at Soyangang River basin by using Cascade-Correlation algorithm and back-propagation algorithm. In order to improve the accuracy of the flood forecasting, the filtering technique has been used at the neural network model. As a result, Cascade-Correlation filtering model shows better forecasting capability.