Prediction of salinity in San Francisco bay delta using neural network
T. Dhiliphan Rajkumar, M.L. Johnson · 2002
We used an artificial neural network approach to predict salinity intrusion, a complex problem in the Sacramento-San Joaquin River delta in California (USA). The inputs comprise the flows from the Sacramento and San Joaquin rivers, and stage data from different locations around the bay and delta region. Different neural network architectures and training algorithms were applied to this problem to find the optimal architecture to satisfy all possible scenarios. Out of all training algorithms tested, the back propagation method using the Levenberg-Marquardt algorithm was the best predictor of salinity intrusion. The neural network was composed of three layers with a hidden layer of neurons consisting of three times the number of input neurons. Predicted salinities were within ten percent of the actual salinity at the Carquinez strait (RSAC054) measured for two periods of time, April 1997 and August 1998. Two selected management scenarios consisting of increased pumping at the federal and state water projects were evaluated to determine the resulting change in salinity at Carquinez strait. Increased pumping by fifty percent resulted in an increase in salinity of twenty percent. The speed with which these predictions can be evaluated indicate that a neural network approach could be used to evaluate a large number of potential management scenarios to determine their general effects on salinity intrusion into the delta.