Figure 1.
Pipeline in between input, output, a conceptual model and its prediction. Dashed arrow corresponds to the need of output for the training of the model.
Figure 1.
Pipeline in between input, output, a conceptual model and its prediction. Dashed arrow corresponds to the need of output for the training of the model.
Figure 2.
Map of Oslo with its residential (blue) and industrial (dark gray) areas, its roads (primary, secondary, motorway and trunk) and its air quality (AQ) station network measuring (purple circle). GIS data come from Open-Street Map.
Figure 2.
Map of Oslo with its residential (blue) and industrial (dark gray) areas, its roads (primary, secondary, motorway and trunk) and its air quality (AQ) station network measuring (purple circle). GIS data come from Open-Street Map.
Figure 3.
observation timeseries between 4 December 2015 00:00:00 and 8 December 2015 00:00:00 of the nine monitoring stations in the municipality of Oslo.
Figure 3.
observation timeseries between 4 December 2015 00:00:00 and 8 December 2015 00:00:00 of the nine monitoring stations in the municipality of Oslo.
Figure 4.
Permanent location of station 504 at Smestad (left). Temporary location of station 504 from 21 May 2015 to 8 February 2017 (right).
Figure 4.
Permanent location of station 504 at Smestad (left). Temporary location of station 504 from 21 May 2015 to 8 February 2017 (right).
Figure 5.
Interpretation of the predictive qq (pqq)-plot adapted from [
45].
Figure 5.
Interpretation of the predictive qq (pqq)-plot adapted from [
45].
Figure 6.
Sample of observation (dots) timeseries with the RFreg prediction (line), the error prediction at CI95 (gray area) following Eaamm 2010 at station 7 between 4 December 2016 and 8 December 2016 (upper left), following Wager 2014 at station 7 between 4 December 2016 and 8 December 2016 (upper right), following Lu 2019 at station 7 between 4 December 2016 and 8 December 2016 (lower).
Figure 6.
Sample of observation (dots) timeseries with the RFreg prediction (line), the error prediction at CI95 (gray area) following Eaamm 2010 at station 7 between 4 December 2016 and 8 December 2016 (upper left), following Wager 2014 at station 7 between 4 December 2016 and 8 December 2016 (upper right), following Lu 2019 at station 7 between 4 December 2016 and 8 December 2016 (lower).
Figure 7.
Predictive qq-plot for the training/testing period 2017–2018 (left). Predictive qq-plot for the validation period 2015–2016 (right). The different methods are colored in red for Eaamm 2010, green for Wager 2014 and blue for Lu 2019. Each curve corresponds to one station.
Figure 7.
Predictive qq-plot for the training/testing period 2017–2018 (left). Predictive qq-plot for the validation period 2015–2016 (right). The different methods are colored in red for Eaamm 2010, green for Wager 2014 and blue for Lu 2019. Each curve corresponds to one station.
Figure 8.
Sample of observations (dots) timeseries with the RFreg prediction (line), the error prediction at CI95 (gray area) following Lu 2019 at station 11 between 4 December 2016 and 8 December 2016 (left), following Lu 2019 at station 504, one year earlier, between 4 December 2015 and 8 December 2015 (right).
Figure 8.
Sample of observations (dots) timeseries with the RFreg prediction (line), the error prediction at CI95 (gray area) following Lu 2019 at station 11 between 4 December 2016 and 8 December 2016 (left), following Lu 2019 at station 504, one year earlier, between 4 December 2015 and 8 December 2015 (right).
Figure 9.
Results at Station 11. Predictive qq-plot for the training/testing period 2017–2018 (left). Predictive qq-plot for the validation period 2015–2016 (right). The different methods are colored in red for Eaamm 2010, green for Wager 2014 and blue for Lu 2019.
Figure 9.
Results at Station 11. Predictive qq-plot for the training/testing period 2017–2018 (left). Predictive qq-plot for the validation period 2015–2016 (right). The different methods are colored in red for Eaamm 2010, green for Wager 2014 and blue for Lu 2019.
Figure 10.
Results at Station 504. Predictive qq-plot for the training/testing period 2017–2018 (left). Predictive qq-plot for the validation period 2015–2016 (right). The different methods are colored in red for Eaamm 2010, green for Wager 2014 and blue for Lu 2019.
Figure 10.
Results at Station 504. Predictive qq-plot for the training/testing period 2017–2018 (left). Predictive qq-plot for the validation period 2015–2016 (right). The different methods are colored in red for Eaamm 2010, green for Wager 2014 and blue for Lu 2019.
Figure 11.
Illustration of the artifact of RFreg on structural modeling error with dataset A and subset type 1. Artfact on the prediction is shown on the left. Artifact on the error is presented on the right. Vertical gray lines represent output subset used for the training phase.
Figure 11.
Illustration of the artifact of RFreg on structural modeling error with dataset A and subset type 1. Artfact on the prediction is shown on the left. Artifact on the error is presented on the right. Vertical gray lines represent output subset used for the training phase.
Figure 12.
Illustration of the artifact of RFreg on structural modeling error with dataset B and subset type 2. Artifact on the error (purple) and Wager 2014 error prediction CI95 (dark gray) is shown on the left. Artifact on the error (purple) and Lu 2019 error prediction CI95 (dark gray) is shown on the right. Vertical gray lines represent output subset used for the training phase.
Figure 12.
Illustration of the artifact of RFreg on structural modeling error with dataset B and subset type 2. Artifact on the error (purple) and Wager 2014 error prediction CI95 (dark gray) is shown on the left. Artifact on the error (purple) and Lu 2019 error prediction CI95 (dark gray) is shown on the right. Vertical gray lines represent output subset used for the training phase.
Figure 13.
Same as
Figure 12 for the second run of the experiment.
Figure 13.
Same as
Figure 12 for the second run of the experiment.
Figure 14.
Same as
Figure 12 for the third run of the experiment.
Figure 14.
Same as
Figure 12 for the third run of the experiment.
Figure 15.
Same as
Figure 12 for the fourth run of the experiment.
Figure 15.
Same as
Figure 12 for the fourth run of the experiment.
Figure 16.
Predictive qq-plot for the prediction period with dataset B and subset type 2. The different methods are colored in red for Wager 2014 and blue for Lu 2019.
Figure 16.
Predictive qq-plot for the prediction period with dataset B and subset type 2. The different methods are colored in red for Wager 2014 and blue for Lu 2019.
Table 1.
Metadata of the nine monitoring stations measuring located in the metropolitan region of Oslo.
Table 1.
Metadata of the nine monitoring stations measuring located in the metropolitan region of Oslo.
Name | ID | Municipality | Coordinates (Lon/Lat) | Area Class | Station Type | EOI |
---|
Alnabru | 7 | Oslo | (10.84633, 59.92773) | suburb | near-road | NO0057A |
Bygdøy Alle | 464 | Oslo | (10.69707, 59.91898) | urban | near-road | NO0083A |
Eilif Dues vei | 827 | Bærum | (10.61195, 59.90608) | suburb | near-road | NO0099A |
Hjortnes | 665 | Oslo | (10.70407, 59.91132) | urban | near-road | NO0093A |
Kirkeveien | 9 | Oslo | (10.72447, 59.93233) | urban | near-road | NO0011A |
Manglerud | 11 | Oslo | (10.81495, 59.89869) | suburb | near-road | NO0071A |
Rv 4, Aker sykehus | 163 | Oslo | (10.79803, 59.94103) | suburb | near-road | NO0101A |
Smestad | 504 | Oslo | (10.66984, 59.93255) | suburb | near-road | NO0095A |
Åkebergveien | 809 | Oslo | (10.76743, 59.912) | urban | – | – |
Table 2.
Metadata of the nine monitoring stations measuring located in the municipality of Oslo.
Table 2.
Metadata of the nine monitoring stations measuring located in the municipality of Oslo.
ID | Component | 2015–2018 |
---|
Coverage (%) | Valid (%) |
---|
7 | | 99 | 93 |
464 | | 99 | 90 |
827 | | 99 | 99 |
665 | | 99 | 98 |
9 | | 92 | 92 |
11 | | 99 | 99 |
163 | | 97 | 96 |
504 | | 98 | 98 |
809 | | 99 | 90 |
Table 3.
Parameters of the measurement error expression. and are expressed in , is a percentage. Values from NILU come from an internal technical report. Parameters and are divided by 1.96 in order to get a generalization on the whole Normal distribution and not only at the confidence interval 95-percentile (CI95). Parameter is a threshold and does not require any division by 1.96.
Table 3.
Parameters of the measurement error expression. and are expressed in , is a percentage. Values from NILU come from an internal technical report. Parameters and are divided by 1.96 in order to get a generalization on the whole Normal distribution and not only at the confidence interval 95-percentile (CI95). Parameter is a threshold and does not require any division by 1.96.
Entity | | | |
---|
TÜV | 0 | 4.35/1.96 | 0 |
NILU | 5.64/1.96 | 5/1.96 | 112.8 |
Table 4.
Characteristics of “truth” datasets with expression between input and output, linear: , non-linear: , with if , and otherwise.
Table 4.
Characteristics of “truth” datasets with expression between input and output, linear: , non-linear: , with if , and otherwise.
ID | Length | Input | Pdf | |
---|
Distribution | Parameters |
---|
A | 2991 | 1 | uniform | – | linear |
B | 2991 | 10 | exponential | [0.01:0.055] | non-linear |
Table 5.
Characteristics of training datasets subset, built-up from “true” datasets.
Table 5.
Characteristics of training datasets subset, built-up from “true” datasets.
Type | Length | Index Range | Choice of the Indexes |
---|
1 | 1000 | [1001:2000] | non-random, no replacement |
2 | 1000 | [1:2991] | random, no replacement |
Table 6.
Metrics of the prediction of the nine monitoring stations measuring for the testing phase and the validation phase.
Table 6.
Metrics of the prediction of the nine monitoring stations measuring for the testing phase and the validation phase.
| Testing | Validation |
---|
ID | rmse | Bias | | rmse | Bias | |
7 | 12.92 | 0.10 | 0.76 | 15.49 | 4.70 | 0.76 |
464 | 10.13 | −0.13 | 0.79 | 13.82 | −5.21 | 0.80 |
827 | 11.33 | 0.42 | 0.74 | 15.37 | 0.79 | 0.70 |
665 | 16.26 | −0.53 | 0.71 | 3.34 | −2.04 | 0.67 |
9 | 7.34 | 0.13 | 0.85 | 12.35 | −3.20 | 0.80 |
11 | 18.22 | 0.81 | 0.57 | 22.01 | 2.71 | 0.54 |
163 | 10.31 | −0.29 | 0.76 | 13.91 | 0.17 | 0.72 |
809 | 8.16 | 0.33 | 0.79 | 9.98 | 1.18 | 0.80 |
504 | 9.66 | −0.10 | 0.83 | 19.94 | −11.95 | 0.77 |