Application of principal components regression to streamflow forecasting

TitleApplication of principal components regression to streamflow forecasting
Publication TypeConference Proceedings
Year of Conference1996
AuthorsSmith, S., and Weiss E.
Conference Name64th Annual Western Snow Conference
Series TitleProceedings of the 64th Annual Western Snow Conference
Date PublishedApril 1996
PublisherWestern Snow Conference
Conference LocationBend, Oregon
KeywordsForecasting, Regression, Statistical analysis

B.C. Hydro adapted the techniques of principal component regression, cross-validation and a systematic search for optimal combinations of variables (Garen, 1993), to try to improve to the statistical accuracy of stream- flow volume forecast equations for the glaciated Bridge River basin in British Columbia. The paper will discuss the trade-offs that B.C. Hydro made in balancing the improvements in forecast accuracy using these techniques against other factors, such as preserving continuity in forecast equation structure from month to month, preferring short-term stations located within the basin to longer term stations outside the basin, and using the thirty-year normals period ( 1961-90) versus the historical period of record. Results for the Bridge River system demonstrate the balance that can be achieved between optimizing forecast accuracy and selecting variables and stations which make the forecast equations more physically rational.