Forecasting summertime surface-level ozone concentrations in the Lower Fraser Valley of British Columbia: An ensemble neural network approach

Empirical models for predicting daily maximum hourly average ozone concentr ations were developed for 10 monitoring stations in the Lower Fraser Valley (LFV) of British Columbia. According to data from 1991 to 1996, ensemble n eural network models increased explained variance an average of 7% over mul tiple linear regression models using the same input variables. Without modi fication, all models performed poorly on days when the observed peak ozone concentration exceeded 82 parts per billion, the National Ambient Air Quali ty Objective. When numbers of extreme events in training data were increase d using a histogram equalization process, models were able to forecast exce edances with improved accuracy. Modified generalized additive model (GAM) p lots and associated measures of input variable importance and interaction w ere generated for a subset of the trained models and used to investigate re lationships between input variables and ozone levels. The neural network mo dels displayed a high degree of interaction among inputs, and it is likely the ability of these model types to account for interactions, rather than t he nonlinearity of individual input variables, that explains their improved forecast skill. Inspection of GAM-style plots indicated that the relative importance of input variables in the ensemble neural network models varied with geographic location within the LFV. Four distinct groups of stations w ere identified, and rankings of inputs within the groups were generally con sistent with physical intuition and results of prior studies.