During the spraying seasons of 2014 and 2015, wind velocity and solar radiation (2014 only) were collected at a one meter height above the ground to simulate conditions affecting droplets near a ground-based spray boom. This instrumentation was placed in a cross pattern with sensors at the four cardinal directions (north, south, east, and west) with a fifth sensor in the center (2015 only). Data were collected at 10 Hz to measure the turbulent properties of the wind near the ground.

Measurements of wind velocity profiles moving from upwind sensors to downwind sensors were used to evaluate correlation between the wind measurements. Two periods in which wind direction, on average, was collinear with multiple sensors were investigated. The first period contained five hours of data in which the average wind speed was 3.6 m/s (8 mi/h), while the second period contained 1.5 hours of data with an average wind speed of 1.5 m/s (3.4 mi/h). For the five hour dataset, correlation coefficients of 0.29 and 0.27 were found for wind direction and wind speed measured at two sensors respectively. This value fell when the five hours were broken up into multiple one minute periods. The correlation coefficients rose from less than 0.03 to greater than 0.14 once a lag term was introduced to the data. These results were not observed in the 1.5 hour dataset. Over the 1.5 hour period, the correlation coefficients were found to be less than 0.03. The introduction of a lag term had no clear effect.

The entirety of the datasets that were collected in 2014 and 2015 were investigated to see under what conditions large wind change events were more likely to occur. The datasets suggest that low wind speeds lead to higher probability of large wind changes. As solar radiation increased so did the probability of large changes in wind. As a tolerance on the wind shift was tightened, the probability of wind changes became uniform.

In models that predict spray drift, a popular method to simulate turbulent wind conditions in which the droplet is entrained, is to update the current wind velocities with a random process to achieve new wind velocities. This type of process is known as a random walk. The random walk hypothesis was tested using data collected at 10 Hz, and the average of the collected data to simulate data recorded at 0.5 s, 1 s, 5 s, 10 s, 30 s, 1 min, 5 min, and 10 min. For all tests below five minute averages, the test rejected the hypothesis that wind velocity updates can be independent of previous measurements at greater than 95% confidence. Indicating that updates to the current wind velocity is dependent on previous velocities.

To help reduce the chances of spray drift, prediction models were developed and tested to predict wind direction 30 seconds into the future utilizing current and past measurements. The models tested included a kernel filter that is used for prediction of wind speeds for wind turbines, an autoregressive process (AR), a full ARIMA process, and a hybrid model that includes ideas from ARIMA and Taylor series expansions. The listed models were tested against a “No Model” model in which the predicted value was simply the current observed value. Models were trained over a one hour dataset and tested over a four hour data set. The AR and hybrid models lowered the RMS error value by 9% over the “No Model” model. The AR and hybrid models were outside of a 20 degree tolerance about 12% of the time.

The correlation values between an upwind and downwind sensors indicate that little correlation exists. Along with the predictive models yielding limited results indicate that the wind changes rather randomly. However, results from testing the time series against the random walk hypothesis indicate that wind’s random fluctuations are correlated with one another, but these correlations are not seen using linear correlations. Further effort is needed to better model the wind process.