Contacts:


Project Plan:

Start Date: On hold.

Status: On hold.

End Date: On Hold.

Abstract:

During the past 50 years the CBRFC has been producing water supply forecasts for many sites in the CBRFC area of responsibiy. The technique used during most of this period has relied upon multiple linear regression (Statistical Water Supply-SWS). The historical observed flows were regressed against rational variables such as snow water equivalent, precipitation and more recently teleconnection indices such as the SOI and SST. During the past 10-15 years the CBRFC has begun to utilize ESP or Ensemble Streamflow Prediction for some forecasts. It was thought that other techniques should be investigated, analyzed and compared to see if significant improvement could be made.

Other techniques are available that provide forecast capabilities, using rational input variables and observed output flow volumes to calibrate the model. One such technique is based on : Artificial Neural Networks (ANN). This provides a technique that can accommodate non-linear relationships, such as those that are prevalent in the Lower Colorado Basin. The ANN software that will be used is a commercial package written by Ward Systems. A second technique is based on pattern matching or analogs of historical occurrences (KNN). The idea here is that if certain patterns which occurred historically produced a particular volume flow, that similar patterns in the future would produce similar flows. One such technique is based on procedures by Upmanu Lall and Ashish Sharma, call "Nearest Neighbor Bootstrap for Time Series Resampling". This work relies on work done by S. Yakowitz called, "Nearest-Neighbor Methods for Time Series Analysis." The original software was enocded in FORTRAN by Connely Baldwin, Utah State University and was converted to ANSII C by Dave Brandon ( HIC-CBRFC).

The project will select at a minimum of 30 sites across the CBRFC area of responsibility. The same data sets will be used as input to the SWS, ANN and KNN. Historical time series of precipitation and temperature for the same time period will be used to feed the ESP system. The Jackknife standard error will be compared between the SWS, KNN and ANN techniques. The most probable, 10 and 90 percent bound forecast values along with the observed will also be used in order to compare the four techniques.