Abstract
The set of hourly averaged condensation nuclei (CN) data collected at Mace Head during 1991-1994 was examined for relationships that might exist between CN number concentrations and the more commonly measured meteorological variables, including tides. CN number concentrations at Mace Head can be characterized by typically low ``background'' levels (less than about 700 particles cm super(-) super(3) ) when the wind is from the west, somewhat higher ``background'' levels (1000-4000 particles cm super(-) super(3) ) when the wind is from the east, with sporadic bursts of short-lived discrete ``events'' of more than 10,000 cm super(-) super(3) for several hours. These events occur typically during early afternoon and are normally associated with slack winds and anomalously warm, dry air. They appear to be independent of pressure, wind direction and precipitation. They can occur any time during the year, although the strongest events tend to occur during spring and autumn. Large-amplitude low tides also occur predominantly in the early afternoon during this observing period. We present evidence that large CN concentration events occur preferentially after exceptionally low tides during daylight. A neural network was employed to train the standard meteorological variables to predict CN concentrations. Baseline forecasts of CN counts for the final 180 days of the observing period were made using lagged values of all other variables. Further forecasts were made with some variables removed from the predictor set. The best correlation between the predicted values and the verifying data over the 180 days was 0.67, which was obtained from a 1-hour forecast using knowledge of all variables except temperature. Other variables whose removal improved the forecast (or whose presence degraded it) were pressure and wind speed. The best predictors of CN values were wind direction, relative humidity, and time of day. An elementary ``nearest neighbor,'' or ``historical analogue'' approach to predicting the same set of CN values generated lower correlations with the verifying data but generated a much more accurate probability distribution function.