Ragweed Pollen Alarm System (R-PAS) has been running since 2014 to provide pollen
information for countries in the Pannonian biogeographical region (PBR). The aim of
this study was to develop forecast models of the representative aerobiological monitoring
stations, identified by analysis based on a neural network computation. Monitoring
stations with 7-day Hirst-type pollen trap having 10-year long validated data set
of ragweed pollen were selected for the study from the PBR. Variables including forecasted
meteorological data, pollen data of the previous days and nearby monitoring stations
were used as input of the model. We used the multilayer perceptron model to forecast
the pollen concentration. The multilayer perceptron (MLP) is a feedforward artificial
neural network. MLP is a data-driven method to forecast the behaviour of complex systems.
In our case, it has three layers, one of which is hidden. MLP utilizes a supervised
learning technique called backpropagation for training to get better performance.
By testing the neural network, we selected different sets of variables to predict
pollen levels for the next 3 days in each of the monitoring stations. The predicted
pollen level categories (low-medium-high-very high) are shown on isarithmic map. We
used the mean square error, mean absolute error and correlation coefficient metrics
to show the forecasting system's performance. The average of the Pearson correlations
is around 0.6 but shows big variability (0.13-0.88) among different locations. Model
uncertainty is mainly caused by the limitation of the available input data and the
variability in ragweed season patterns. Visualization of the results of the neural
network forecast on isarithmic maps is a good tool to communicate pollen information
to general public in the PBR.