On-line short-term streamflow forecasting using neural networks

Previous studies showed that a hybrid multi-layered perceptron network (HMLP) coupled with a modified recursive prediction error (MRPE) training algorithm has a very good prediction performance. The network has a compact structure and can be trained rapidly. HMLP network does not require multiple training epochs. Based on these promising properties of HMLP and MRPE, the current study investigates the capability of the network to perform on-line short-term streamflow forecasting. A portable on-line short-term streamflow system will be built, which includes a flow sensor to measure streamflow and a HMLP network to model and at the same time forecast the streamflow. The simulation study showed that the system starts to give good forecasting results after 150 training samples for 3-h ahead and good 24-h ahead after 300 samples. These results suggest that the system is suitable for short-term on-line forecasting.