Early flood warning system for the Glenmore Reservoir with ANN flow preditions
With the objective of developing an early flood warning system for the Glenmore Reservoir in the Elbow watershed, an Artificial Neural Network (ANN) model was developed to predict streamflows in the Elbow River. Artificial Neural Networks (ANN) provide a potentially useful framework for modelling complex non-linear and dynamic systems, including the complex hydrologic processes that influence stream.flows and flood events. The resulting ANN prediction tool then forms the core of the early flood warning system. Real-time measurements of flows, temperatures and precipitation in the Elbow watershed during preceding days are used as inputs to the prediction model. With the premise that events in a neighbouring waters~ed could potentially trigger a warning, a preliminary correlation analysis was undertaken into the use of available precipitation data for neighbouring watersheds. Either due to low correlation or absence of lag-time with the measured precipitation in the Elbow watershed, this approach for flood warning was abandoned. Precipitation measured in neighbouring watersheds would basically provide limited information on the anticipated precipitation in the Elbow watershed. The optimal ANN model recommended for the early warning system is the result of a thorough investigation of multiple design approaches, including the dual ANN framework. Through the training process of candidate ANN models, the study recognized the shortcomings of the ANN formulation in terms of extrapolating flows beyond the training data sets. The very few peak events available for training, each having different antecedent conditions, also presented difficulties for the ANN, whereby the antecedent space was poorly sampled. To overcome this problem, the study used artificially generated data using the SSARR model already calibrated for the Elbow watershed. The SSARR generated events, together with available measured data, provided a better coverage of the antecedent space. The resulting optimal ANN is capable of forecasting flows up to the 1 in 100 years return period with a maximum error of 16%. Sensitivity analysis of each input parameter proved crucial in the selection of this best model. The development of the ANN model supported the influence of topography on precipitation observed in the correlation analysis of neighbouring watersheds. The prediction capability of the ANN model was significantly improved with the use of multiple gauging stations, rather than relying on the precipitation data gathered at a single location. The core elements of the early warning system attained through this study are improved warning time, prediction of streamflows ( addressed with the ANN model) and the identification of an appropriate action to mitigate the predicted flood event. The study addressed the third element with preliminary reservoir drawdown levels, using the GFlood Program, triggered by the forecasts of the optimal ANN model.
Bibliography: p. 156-162
Yadete, H. A. (2004). Early flood warning system for the Glenmore Reservoir with ANN flow preditions (Master's thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca. doi:10.11575/PRISM/13713