AN ARTIFICIAL NEURAL NETWORK-BASED DECISION SUPPORT SYSTEM TO EVALUATE HYDROPOWER RELEASES ON SALINITY INTRUSION

摘要

Six reservoirs in North Carolina, USA, discharge into the Pee Dee River, which flows 260 kilometers through South Carolina to the coastal communities near Myrtle Beach. During the Southeast's record-breaking drought from 1998 to 2002, salinity intrusions inundated a coastal municipal freshwater intake, limiting water supplies. The North Carolina reservoirs are currently (2006) undergoing a re-licensing process by the Federal Energy Regulatory Commission for a 50-year operating permit. Stakeholders along the Pee Dee River formed a coalition to determine the necessary flows to protect the freshwater intakes in the future. Salinity intrusion results from the interaction of three principal forces-streamflow, mean tidal water levels, and tidal range. To analyze, model, and simulate hydrodynamic behaviors at critical coastal gages, data mining techniques were applied to more than 15 years of hourly streamflow, coastal water-quality, and water-level data. Artificial neural network (ANN) models were trained to learn the variable interactions that cause salinity intrusions. Streamflow data from the 47,900-square-kilometer basin are used as input to the model as time-delayed variables and accumulated tributary inflows. Tidal inputs to the models were obtained by decomposing tidal water-level data into a "multiply periodic" signal of tidal range and a "chaotic" signal of mean water levels. The ANN models were able to convincingly reproduce historical behaviors and generate alternative scenarios of interest. To make the models directly available to all stakeholders, a user-friendly decision support system was developed as a spreadsheet application that integrates the historical database, ANN models, user controls, streaming graphics, and simulation output.