The traditional air quality model initialization approach, in which one "cycles" the model, using one forecast to initialize the next, suffers from potential model drift, in which errors in one forecast simulation (possibly due to errors in the meteorological, emissions, or air quality simulations within the system) propagate into the next. For the MCNC/BAMS Real-Time Ozone Forecast System, we have attempted to ameliorate these errors by assimilating chemical observational data from monitor data provided by through EPA's AIRNOW project. Several algorithms have been suggested in the literature for this; however, we have incorporated a new algorithm that attempts to address these problems with the prior algorithms:

*It attempts to incorporate as much of the monitor data as possible (rather than looking only at a one-hour snapshot).

*It attempts to incorporate the monitor data in a relatively unbiased fashion, instead of being overly influenced by the last data reported by a given monitor.

*It takes into account the nature of the three-dimensional daytime mixing in the daytime atmosphere, and have an appropriate three-dimensional influence.

*It takes into account wind-driven transport from the times of the various monitor-observations up to the time of model-initialization.

*Where there are multiple adjacent observations, the algorithm should incorporate all of them "fairly."

*It should not attempt to extrapolate into large data-void areas.

*It should be positive definite.

We have employed the algorithm in our summer-season daily MM5/SMOKE/MAQSIP-RT air quality forecasts. Here we give a description of the algorithm and some caveats about the limitations of the algorithm.