Optimization of a Model to Predict Salinity Intrusion in San Francisco Bay Estuary Using a Genetic Algorithm

Salinity intrusion is a significant issue in the San Francisco Bay Delta because this estuary contributes substantially to fresh water supply for much of northern California. Traditionally, salinity along the estuary has been measured using monitoring instruments installed at various locations within the Sacramento-San Joaquin River Delta system, and these measurements are used to predict future conditions and evaluate the water management alternatives. The California Department of Water Resources Delta Simulation Model (DWRDSM) is used for this type of analysis. DWRDSM is an unsteady onedimensional hydrodynamic and salt transport model. It can provide estimates of salinity at almost any location within the Bay-Delta system. However, like many numerical models, when simulating the dynamics of a large complex system, the processing time can be extremely long, which makes it difficult to evaluate a large number of management scenarios in a short time. Statistically-based models of the flow and salinity have been tried and found lacking; classical time series analysis is often linear or requires transformation to a stationary time series that renders the resulting model unsuitable, since the salinity intrusion problem is a complex non-linear problem. Consequently, there remains a need for a reliable and fast method to predict salinity intrusion in the Sacramento-San Joaquin River Delta system. In this paper, such a method is advocated: Genetic Algorithms (GA) are used to optimize a previously built Artificial Neural Network (ANN) that reliably predicts periodic salinity intrusion in the San Francisco Bay Estuary. Results indicate that the GA provided an efficient method of optimizing the ANN model to predict salinity. The reliability of the ANN using the GA includes prediction of salinity to an accuracy of ±10% at hourly intervals for any time period in the year.