Summary and Conclusions

Nonlinear empirical models were developed using artificial neural networks (ANN) and multivariate adaptive regressive splines (MARS) to simulate wind data on Upper Klamath Lake. Both types of models were able to successfully simulate the wind on Upper Klamath Lake based on wind and other meteorological variables measured nearby. By using the MARS method of selecting model input variables prior to using the ANN, some insight into the dependence of wind on various meteorological variables at other land-based sites around the lake was gained. The most successful models were dependent on a diverse set of input variables measured at several sites around the lake.

MARS and ANN were used successfully to simulate wind data at two floating raft sites on Upper Klamath Lake. Both methods were able to capture most of the variability in the 10-minute wind time series. Two different preprocessing methods to smooth the input variables were investigated, but the preprocessing did not improve the performance of the models greatly, and the extra step of smoothing the data may not be necessary. Gap-filling models based on ANN performed moderately better than those based on MARS. Although the ANN approach does not directly reveal the nature of dependencies between input and output variables as well as the MARS approach, the greater accuracy achieved with the ANN approach may justify the added opacity. The set of explanatory variables that contributed to the “best” models as determined by the MARS algorithm included wind from all sites around the lake, as well as wind divergence and curl, and air temperature. The best models for the raft sites included both east-west and north-south wind components from other land-based sites and comprised from 11 to 15 independent variables.

The winds simulated with the ANN gap-filling models were used in the spatially variable wind forcing for the hydrodynamic model of Upper Klamath Lake. The error statistics for the simulated currents, when the simulated winds were used to force the model, were comparable to those for the currents that were simulated when only observed winds were used to force the model. It can be concluded that this technique is adequate for filling gaps of several days in the wind data collected from the rafts on the lake. A numerical tracer experiment indicated that the errors in transport that result when simulated winds are used to drive the hydrodynamic model probably are small, particularly when averaged over periods of several days or more.

The ANN and MARS were used to accurately reconstruct wind data at the Williamson River Delta site WMR on a daily timestep over several years. Both ANN and MARS historical models performed similarly. The datasets long enough to be useful as input to these models were more limited than those available for input to the gap-filling model. Significant explanatory variables selected by the MARS algorithm included wind, solar radiation, and relative humidity at AgriMet sites (AGKO and KFLO), as well as wind and sky cover at the Klamath Falls Airport site (KMLT). The most successful historical wind model simulated the north-south and east-west components separately. The accuracy was higher for this historical wind model than for the gap-filling model, and indicates the potential to simulate the wind on a daily timestep going backward in time at least to the year 2000, when the AgriMet data were first collected. Given the long data record at KMLT, there is a potential to develop an historical wind model based only on data collected at that site to reconstruct the wind at sites around UKL back to 1959, but that is beyond the scope of the current work and the accuracy of such a model has not been investigated for this report.

First posted October 27, 2010