Results of Historical Wind Models

Similar to the gap-filling models, the MARS algorithm was used to select the most significant explanatory variables for the ANN and the MARS historical wind models. These selected variables are shown in table 5. The model of the north-south wind component at WMR selected a greater diversity and larger number of input variables than that selected for the model of the east-west component—notably, relative humidity at AGKO and sky cover at KLMT were added. Solar radiation was selected by the MARS algorithm as a significant variable for all three dependent variables in the historical models (table 5).

Using the same statistical measures as above, the historical wind models are compared in table 6. For all models, NASH values ranged from 0.63 to 0.87 while BIAS values ranged from -0.24 to 0.11 m/s. HIST2 models had higher NASH values and lower BIAS values than HIST1 values, indicating a better statistical fit to measured data. Thus, the historical wind models that use the individual components of the wind as dependent variables (both ANN and MARS models) perform better than the historical wind models that use wind speed only as the dependent variable. Figure 10 displays the close fit to measured data that simulated HIST2 models obtained. As was the case for the gap-filling models, peak winds are the most difficult for the models to simulate consistently, and the models for the dominant (north-south) component of the wind perform better.

Winds reconstructed with the ANN_HIST2 model and measured wind at WMR during the calibration and validation period were aggregated on a weekly basis in order to visualize wind patterns for multiple years simultaneously (fig. 11). To create these “vector-bar” plots, the mean daily north-south wind components and east-west wind components were first converted to wind speed and direction. Next, the daily mean wind direction occurring each day was compartmentalized into eight directional bins. The maximum wind speed occurring in the most populated direction bin was used to determine the height of each bar. The mean of the wind-direction values associated with maximum wind-speed values determined the direction of the arrows. The maximum wind speeds and associated wind direction were used because they are assumed to be most environmentally significant for Upper Klamath Lake. The seasonal reconstructed winds generally matched the seasonal change in measured winds at WMR (red and black “vector-bars”). A closer agreement generally was observed during summer, when the winds are weaker. In cases where simulated wind direction differed from measured wind direction, the difference was infrequently more than 45 degrees.

To demonstrate the use of the ANN_HIST2 model to reconstruct a long-term record, the model was used to reconstruct the daily wind record at WMR from January 1, 2004, through June 12, 2005, prior to the time when data were collected at this site (fig. 12). This approach has the potential to expand the length of the wind record at WMR or other sites around the lake backward in time to at least 2000 (when collection of the AgriMet data began) for the purpose of statistical investigation of the relation between wind and water quality in the lake.

First posted October 27, 2010