Artificial Neural Networks and MEC
Artificial Neural Networks (ANNs) are used to predict salinity
at various locations within the Delta. Work is continuing to
improve the accuracy of the models and develop ANNs which are
applicable over a wider range of inputs. These new ANNs were
used to perform Marginal Export Cost (MEC) estimates.
Response of ANN-Generated Salinity
to Varying SAC Flows and Exports
As part of the ANN evaluation process some plots were developed
to help exhibit the relationship between Sacramento River flow,
exports, and salinity at Contra Costa Canal and Jersey Point.
Figure 7.1 shows how Contra
Costa EC differs with varying SAC flow and CVP pumping when all
other inputs are kept constant. Figure
7.2 shows how Contra Costa EC varies with changing SAC flow
and SWP pumping. Figures 7.3 and
7.4 show similar plots for Jersey
Marginal Export Cost (MEC) Estimates
ANN can be used to estimate MEC for changes in exports. The
June 1997 Annual Report described how ANNs can be used to predict
salinity and estimate MEC. This technique was used to estimate
MEC for a continuous 500 cfs reduction in pumping. The same technique
was used to estimate MEC with the pumping needed for the 1991
Drought Water Bank.
MEC has been defined as the extra water needed to carry a
unit of water across the Delta to the pumping plants for export
while maintaining a constant salinity at a given location. MEC
varies widely and it is highly dependent on location and antecedent
conditions. Incremental export increases may increase salinity
in some areas while decreasing salinity in others. Modeling studies
using DSM2 and ANNs can be used to study the complex interrelationship
of flows and salinity in the Delta.
The Continuous Impulse Marginal Export Cost (CIMEC) method
was used to study carriage water under historic conditions. Two
investigations were performed. The first investigation looked
at the effects of decreasing exports by 500 cfs, and then recalculated
the SAC flow needed to maintain salinity at a constant level.
The second experiment attempted to quantify the MEC associated
with the pumping required by the 1991 Drought Water Bank.
Both investigations used ANNs which used CVP, DXC position,
SAC, SJR, and SWP as inputs and both ANNs were trained with DSM2
salinity output. The period studied for the first investigation
used historic data for the five-year time starting January 1989
to November 1994. The second study used 1991 historic data.
Jersey Point and Contra Costa Canal were chosen as the study
locations because they represent two interior Delta locations
which have salinity standards which often control Delta operations
and where the salinity/outflow relationship is complex.
Effects of a 500 cfs Reduction in SWP
Jersey Point salinity, historic flows, and DXC gate position
were used to obtain a baseline case by using the CIMEC SAC flow
estimation methodology to calculate SAC flow for January 1989
through December 1994.
The historic SWP pumping data was modified by reducing the
SWP exports by 500 cfs. SAC flow was then recalculated so that
salinity at Jersey Point remained at historic levels.
Figure 7.5 shows how the 500 cfs
reduction affected the calculated SAC flow values. Monthly carriage
water was calculated using the following equation:
- C.W.= (dExports-dCalculated SAC flow) or
- C.W.=((Hist.SWP pumping-500cfs)-(Hist. SWP pumping))-
- ((Calc SAC with exports reduced by 500cfs)-
- (Calc SAC for historic conditions))
Figure 7.6 shows monthly calculated
carriage water at Jersey Pt. expressed in cfs and as a percentage
of export reduction. Carriage water percentage can be zero, negative
or positive. A zero value implies that there is a one-to -one
correspondence between incremental increases in pumping and the
incremental increase in SAC flow needed to maintain salinity
levels at a given station. A negative percentage implies that
dSAC flow is less than dExports, while a positive carriage water
percentage implies that dSAC flow is greater than dExports when
SAC flow is adjusted to keep salinity constant.
The monthly percent carriage water at Jersey Point is positive
but showed some variation which may be attributed to the varying
flows and DXC position.
The average carriage water value for the period was defined
as: Avg C.W. = (monthly calculated C.W.)/ (Dexports) and was
found to be about 9 percent.
This experiment was repeated for the same period using salinity
at Contra Costa Canal. Figure 7.7
shows how the 500 cfs reduction affects the estimated SAC flow
values when SAC flow is calculated using historic salinity at
CCC. Monthly carriage water values were calculated and are shown
in Figure 7.8.
When CCC salinity is assumed to be controlling, the monthly
carriage water ratio varies from -90 percent to 60 percent. The
monthly percent carriage water with CCC controlling is much more
volatile than the carriage water value observed when Jersey Pt
salinity is used. The average carriage water for the entire period
with CCC controlling was 15 percent.
Estimate of Carriage Water for 1991
Drought Water Bank (DWB) Pumping
The second part of this experiment was to estimate the MEC
associated with the 1991 Drought Water Bank (DWB) pumping. MEC
was estimated once for Jersey Point EC controlling and for CCC
Jersey Pt. historical salinity, historic rim inflows and exports
and DXC position were used to calculate SAC flow using the CIMEC
method described previously in this report. This calculated SAC
flow was used as the baseline case.
The 1991 historic exports for SWP were then modified by subtracting
the SWP exports attributed to the 1991 Drought water Bank from
the historic SWP export data, and SAC flow was recalculated.
Monthly carriage water values were calculated and the results
are shown in Figure 7.9. The plot
shows the monthly pumping made for the 1991 Drought Water Bank
(DWB), the carriage water attributable to the 1991 DWB pumping,
and the ratio of (C.W. for DWB pumping)/(avg DWB pumping).
The average carriage water for the 1991 Drought Water Bank
Pumping period with
Jersey Point controlling came out to be 8.9 percent.
The process was repeated using CCC historic salinity. CCC
historic salinity, historic rimflows and pumping, and DXC position
were used to calculate SAC flow using the CIMEC method. The 1991
DWB exports were subtracted from SWP pumping and the new reduced
SWP values were used to recalculate SAC flow. Carriage water
was calculated monthly and the results are shown in Figure
7.10. Figure 7.9 and show
how calculated carriage water estimates can vary depending on
controlling location and changing monthly conditions.
The average carriage water for 1991 DWB pumping with CCC controlling
was calculated to be 14 percent.
These studies show how existing models can be used to further
examine the relationships between salinity at a given location,
rimflows, and gate operations. These preliminary results show
that as we continue to gain an understanding of these complex
flow/salinity/gate operation relationships, opportunities to
further optimize Delta operations will present themselves.
Synthetic ANN Development
Historic based flows and gate positions with DSM2 generated
salinities have been used to train ANNs with considerable success.
However, historic bias in the training set, or incomplete training
data sets could adversely affect ANN development (see 1997 Annual
Synthetic randomly generated training sets could be used to
create ANNs which can give accurate results for a range of inputs
without the errors associated with historic bias.
Synthetic input patterns will be generated by randomly varying
all the flows and gate positions through each input's allowable
range. These inputs will then be fed into DSM2 to generate salinity
values. Finally, ANNs will then be trained on the synthetically
generated inputs and salinities.
If the synthetic ANN training process is successful, and a
methodology for randomly choosing input patterns and ensuring
an adequately sized training pattern can developed, synthetic
ANNs could be an important tool for modeling salinity in the
Goto: 1998 Annual
Goto: Annual Reports