A neural network model for flood forecasting for small hydro plants

Artificial Neural Networks (ANNs) provide a quick and flexible way to create models for streamflow forecasting and have been shown to perform well in comparison with conventional hydro logical models. This research applied multi-layer feedforward error backpropagation ANNs for real-time reservoir daily and hourly inflow forecasting. The proposed ANN models are trained by the Levenberg-Marquardt Backpropagation (LMBP) technique, coupled with an early stop method to avoid overfitting. A dataset partition method, which keeps the statistical properties of the training and the monitoring datasets as close as possible, is introduced to avoid under fitting. The method redistributes input/output patterns, in term of streamflow magnitude, into the training dataset and the monitoring dataset by breaking down the time series of the original data into subsets. The research introduced several indicators to cope with the snowmelt affected streamflow forecasting and overcome the limitation of snow information availability. The performance of the daily time step ANN model is compared to an operational conceptual model (UBC Watershed Model) and a one time step lag model. The hourly time step A NN models are compared to a black-box model: Multi-Input Single Output Linear Model (MISOLM). The overall results of the research show that the ANN technique is practicable and effective for real-time streamflow and flood forecasting; the ANN models have higher simulation accuracy than the other referenced models. The models developed have been implemented in BC Hydro. The real-time test of the models showed that ANN is a promising method for snowmelt affected streamflow and flood forecasting.Applied Science, Faculty ofCivil Engineering, Department ofGraduat