In Fig. 5.10 the target diagram for PM10 results is shown. In this case, modelling results comply at 66 % with the unit circle criterion, even if the overall BIAS
is around 33 % of the average value and the average correlation between modelled
and actual values is about 0.5.
The four non-complying stations have high values of BIAS and Root Mean
Square Error (RMSE), which could be related to an overestimation of background
values. The target diagram for NO 2 shows that 69 % of the values comply with the
unit circle criterion, while values are similar to PM10 for the other performance
indicators.
5.4.3 Design of Experiments and Source/Receptor Models
Ten emission sets were defined to train the RIAT+ Artificial Neural Networks for
the Great Porto Area. These scenarios have to contain all possible relationships
between precursor emissions and the various air quality indices. Ideally, the number
of scenarios is determined by checking the incremental improvements to the ANN
results of adding additional scenarios to the training dataset.
Starting from the 2009 Portuguese emission inventory, three different emission
levels were considered: B (base case), L (low emission reductions) and H (high
emission reductions).
The B (base) case considers the evolution of 2009 emissions taking into account
the fulfilment of CLE2020 increased by 15 %. The H (high reduction) case is
associated to the Maximum Feasible Reduction of emissions in 2020 (MFR2020),
further decreased by 15 %. These bounds guarantee that the optimal plausible
reductions will lie within those present in the training dataset. The L (low reduction)
scenario results from averaging B and H emission values.
The procedure to implement these S/R models requires two steps. In the first step
the best ANN structures were chosen on the basis of maximum correlation and
minimum RMSE, considering a series of different possible configurations (i.e.
different network structure, activation function and number of cells). Then, in a
second step the best structure was applied to the whole domain. The quality index
considered in this application was the PM10 annual mean. Table 5.3 presents the
best ANNs parameters selected for PM10 Neural Network.
To validate the results from the ANN, output values are compared to the results
calculated by the CTM. The scatter plot in Fig. 5.11 shows the comparison for an
independent validation set which consists of 20 % of the available grid cells not
used in ANN training. The good performance of the ANN, with a
Normalised RMSE of 0.34 and a correlation coefficient of 0.95 confirms that the
ANNs have a sufficient capability to simulate the nonlinear S/R relationship
between PM10 mean concentration and the emission of its precursors.
Further analyses confirmed this conclusion. For instance, the average correlation
between TAPN and the ANN surrogate model in terms of AQI variations with
respect to the base scenario is about 0.93.
98
C. Carnevale et al.
is around 33 % of the average value and the average correlation between modelled
and actual values is about 0.5.
The four non-complying stations have high values of BIAS and Root Mean
Square Error (RMSE), which could be related to an overestimation of background
values. The target diagram for NO 2 shows that 69 % of the values comply with the
unit circle criterion, while values are similar to PM10 for the other performance
indicators.
5.4.3 Design of Experiments and Source/Receptor Models
Ten emission sets were defined to train the RIAT+ Artificial Neural Networks for
the Great Porto Area. These scenarios have to contain all possible relationships
between precursor emissions and the various air quality indices. Ideally, the number
of scenarios is determined by checking the incremental improvements to the ANN
results of adding additional scenarios to the training dataset.
Starting from the 2009 Portuguese emission inventory, three different emission
levels were considered: B (base case), L (low emission reductions) and H (high
emission reductions).
The B (base) case considers the evolution of 2009 emissions taking into account
the fulfilment of CLE2020 increased by 15 %. The H (high reduction) case is
associated to the Maximum Feasible Reduction of emissions in 2020 (MFR2020),
further decreased by 15 %. These bounds guarantee that the optimal plausible
reductions will lie within those present in the training dataset. The L (low reduction)
scenario results from averaging B and H emission values.
The procedure to implement these S/R models requires two steps. In the first step
the best ANN structures were chosen on the basis of maximum correlation and
minimum RMSE, considering a series of different possible configurations (i.e.
different network structure, activation function and number of cells). Then, in a
second step the best structure was applied to the whole domain. The quality index
considered in this application was the PM10 annual mean. Table 5.3 presents the
best ANNs parameters selected for PM10 Neural Network.
To validate the results from the ANN, output values are compared to the results
calculated by the CTM. The scatter plot in Fig. 5.11 shows the comparison for an
independent validation set which consists of 20 % of the available grid cells not
used in ANN training. The good performance of the ANN, with a
Normalised RMSE of 0.34 and a correlation coefficient of 0.95 confirms that the
ANNs have a sufficient capability to simulate the nonlinear S/R relationship
between PM10 mean concentration and the emission of its precursors.
Further analyses confirmed this conclusion. For instance, the average correlation
between TAPN and the ANN surrogate model in terms of AQI variations with
respect to the base scenario is about 0.93.
98
C. Carnevale et al.