Operational planning of water resource systems like reservoirs and hydropower plants calls for real-time forecasting of reservoir inflow. Reservoir daily inflow forecasts provide a warning of impending floods or drought conditions and help to optimize operating policies for reservoir management based on a fine time scale. The aim of this study was to determine the best model for daily reservoir inflow prediction through linear regression, exponential smoothing and artificial neural network (ANN) techniques. The Hedi reservoir, the third largest reservoir in south China with a 1.144 x 109 m 3, was selected as the study site. The performance of these forecasting models, in terms of forecasting accuracy, efficiency of model development and adaptability for future prediction, were compared to one another. All models performed well during the dry season (inflow with low variability), while the non-linear ANNs were superior to other models in frontal rainy season and typhoon season (inflow with high variability). The performance of ANN models were hardly affected by the high degree of uncertainty and variability inherent to the rainy season. Stepwise selection was very helpful in identifying significant variables for regression models and ANNs. This procedure reduced ANN's size and greatly improved forecasting accuracy for ANN models. The impact of training data series, model architecture and network internal parameters on ANNs performances were also addressed in this study. The overall evaluation indicates that ANNs are an effective and robust tool for input-output mapping under more extreme and variable conditions. ANNs provide an alternative forecasting approach to conventional time series forecasting models for daily reservoir inflow prediction.

No description available.