|DXC and Flow Salinity Relationships in the Delta|
|Earlier sections of this report described how
early ANNs trained only on historic data could not correctly model the effects of DXC
operation. A Pittsburg ANN trained only with historic flow was used to show how the
historic correlation between DXC operation and salinity could inhibit ANN model
development. It was suggested that training ANNs using DSM simulated salinity could reduce
the errors caused by historic correlation.
This section revisits the historic correlation problem and its effect on ANNs at Pittsburg and Contra Costa Canal (CCC), which were trained using historic rimflows and DXC gate position data, along with DSM2 generated salinities. A methodology for "decorrelating" the training inputs is proposed and this method is applied to ANN development at Pittsburg and CCC. The results of this experiment are then shown.
Historic data and DSM2 generated salinity were used train an ANN to predict EC at Pittsburg. This ANN was then used to predict daily salinities for the period Oct 1976 to Sept 1981. Two cases of DXC operation were simulated. The first case simulated EC using historic flows where the DXC gate was kept closed for the entire simulation period. The second case again used historic flows but the DXC gate was kept open for the simulation period. Figure 1 shows the simulated EC values for this Pittsburg EC ANN.
Historically during periods of high salinity the DXC gate has been opened, and during periods of low salinity the DXC gate has been closed. An operational decision has been made to operate the DXC in this manner to control salinity in the interior Delta and there is no actual cause/effect relationship between DXC gate operation and EC at Pittsburg. However, this apparent relationship is reflected in the training data and this can adversely affect an ANN's ability to accurately learn how flows affect salinity at Pittsburg. ANNs cannot differentiate between relationships caused by operational decisions and relationships dictated by hydrodynamics and salinity transport laws. The relative positions of the DXC gate and Pittsburg indicate that EC at Pittsburg should be relatively insensitive to DXC operation. Since Pittsburg EC should be insensitive to DXC operation any errors induced by historic correlation should be readily apparent.
The effects of the correlation of DXC operation and historic salinity can be seen in Figure 1. Closing the gate often seems to cause lower EC values and opening the gate seems to result in increased EC values for the simulated period.
Similarly another EC ANN was prepared for the Contra Costa Canal location using historic flows and DSM2 generated EC data for training. Figure 2 shows how EC values vary for two simulated cases. The first case simulates EC using historic rimflows while the DXC was kept closed, while the second case simulates EC for the same historic conditions but the DXC gate kept open for the entire period. Daily values were calculated for the period starting Oct 1976 and ending Sept 1981.
Since CCC EC is actually affected by DXC gate position the problem caused by historical correlation is not as readily apparent, but the plot does seem to show that this ANN has not correctly captured the relationship between DXC gate operation and CCC EC. If the rimflows remain constant for both runs we expect the DXC open run to return the lowest EC estimates, and the DXC closed run to consistently return the highest EC estimates. Figure 2 shows that there are several points on the plot where the DXC closed line falls below the DXC open line.
In order to reduce the effects of historic correlation, the historic training set was augmented. The original training set was composed of 16 years of historic flows and gate position data. A second 16 years of data was simulated by using the same historical flows along with inverted DXC gate data. When the gate was opened during the initial 16 year period the gate was kept closed in the second training period. Whenever the gate was closed during the initial 16 year period the gate was opened during the second 16 year training period. The historic flows along with the inverted gate position data were fed into DSM2 as inputs and simulated salinities for the second 16 year period were obtained. It was hoped that this second training set could reduce the error caused by historic correlation seen at Pittsburg and CCC. These two training sets were merged into a 32 year training set consisting of rimflows, DXC position data, and simulated EC values. This augmented training set was used to train new ANNs to predict EC at Pittsburg and Contra Costa Canal.
Figure 3 shows how an ANN trained on Pittsburg EC using the augmented training set responds to varying DXC operation. This plot compares ANN estimated EC for the DXC open and the DXC closed cases for the period from Oct 1976 to Sept 1981. The Pittsburg ANN trained on the augmented training set now correctly indicates that EC at Pittsburg is relatively insensitive to DXC operation. Both runs give EC values that are almost identical.
Figure 4 shows how the ANN trained on CCC EC data using the augmented training set performed for varying DXC operation. Again it is shown that augmenting the training set with the DXC inverted data improves ANN performance. The EC estimate for the DXC closed case is consistently greater than the estimate for the DXC open simulation. Figures 3 and 4 seem to indicate that augmenting the data set with the inverted DXC data allows us to develop ANNs which are capable of modeling the effects of DXC operation at westward Delta locations like Pittsburg and also at an interior Delta location such as Contra Costa Canal.
As a final check our new ANNs trained on the augmented training sets were used to predict EC for the time period from October 1976 to September 1986. The DXC gate was allowed to operate normally and was allowed to open and close during the simulation period. The same DXC operation and rimflows were used in the original DSM2 run. The DSM2 simulated EC was compared to the ANN simulated EC for Pittsburg and Contra Costa Canal. The results of these comparisons can be seen in Figure 5 and Figure 6. These two plots show that in both cases our ANN networks seem to capture the functionality of DSM2 and the effects of DXC operation.
These results show how we can remove the error associated with historic correlation between ANN inputs and salinity by working to make our training sets as unbiased as possible. Typically, ANN development is restricted by the availability of adequate training data, our data sets must be of both sufficient size and quality so that the ANN can learn the required flow salinity relationships within a reasonable number of training cycles.
Synthetic training sets using DSM derived salinity can provide us with an almost unlimited source of training data. The difficult part is creating manageable training sets which sufficiently exercise our model without unintentionally biasing our data.
|Back | Home||Available Studies||Computer Models||Reports||
Section Home Pages
|Delta Modeling||Computer Assistance|