Emergency generators that provide electricity to office buildings, grocery stores, or warehouses when the main power supply is interrupted rarely exceed the ambient air quality standards. However the large number of multi-megawatt diesel-powered generators supporting a typical data center have the potential to produce concentrations well above the 1-hour NO_2 National Ambient Air Quality Standard (NAAQS). The intermittent nature of the testing and maintenance of these generators means that the traditional approach of modeling them as continuous sources will produce a gross over-estimate of their impact on the ambient air. Also, the operating power levels and number of generators operating simultaneously vary with the particulars of the testing protocol. If each unique combination of power level and number of engines operating simultaneously is considered a mode, a typical installation may require 15 modes of operation to describe its annual operations. Over the past ten years the six data centers with up to 44 generators each that have been built in a small community in Eastern Washington have required a more accurate assessment of their impact on air quality. An approach that combines the Monte Carlo method with traditional modeling has been found to produce a better estimate of the impacts. The recommended recipe to follow for 1-hour NO_2 evaluation is to: 1. Run a dispersion model for each mode assuming continuous operation over the years of interest, saving the hourly output. 2. Characterize each day by the highest hourly concentration at every receptor for each mode, mimicking the form of the 1-hour NO_2 NAAQS. 3. Create a distribution by randomly selecting, without replacement, the days a mode will run and record the modeled concentrations at all receptors on those days. Repeat for all modes. On days where more than one mode is randomly selected, compute the maximum at each receptor to remain consistent with the form of the NAAQS. Fill in the rest of the days in the analysis period with zeros representing those days with no emissions. At each receptor compute the 98th percentile of that distribution. 4. Repeat step 3 1000 times to produce a distribution of 98th percentiles at each receptor. The median of that distribution at each receptor is the best estimate of the 98th percentile and the distribution will also provide, for the first time, the ability to characterize the uncertainty associated with the overall modeling exercise. This paper describes these processes in sufficient detail to allow replication and describes the results of both an initial numerical experiment and an application to a data center. Including the Monte Carlo method roughly doubles the computing time for an air quality analysis.
展开▼