Within the MEGAPOLI and FRANCIPOL projects, on-line high-sensitivity proton transfer reaction – mass spectrometers (PTR-MS, Ionicon Analytik GmbH, Innsbruck, Austria) were used for real-time (O)VOC measurements. As this instrument has widely been described in recent reviews (Blake et al., 2009, and references therein), only a description of analytical conditions relating to ambient air observations is given here.

During these two intensive field experiments, a PTR-MS was installed in a small room located on the roof of the LHVP site (15 m a.g.l.). For the MEGAPOLI winter campaign, PTR-MS measurements performed by the Laboratoire de Chimie Physique (LCP, Marseille, France) have already been described in Dolgorouky et al. (2012). As those performed by the Laboratoire des Sciences du Climat et de l'Environnement (LSCE) during FRANCIPOL have not yet been described elsewhere, more technical details are presented below.

Air samples were drawn up through a Teflon line (0.125 cm inner diameter) fitting into a Dekabon tube in order to protect it from light. A Teflon particle filter (0.45 µm pore diameter) was settled at the inlet to avoid aerosols and other fragments from entering the system. The PTR-MS was operating at standard conditions: a drift tube held at 2.2 mbar pressure, 60 ∘C temperature with a drift field of 600 V to maintain an E/N ratio of ∼ 130 Townsend (Td) (E: electrical field strength [V cm-1]; N: buffer gas number density [molecule cm-3]; 1 Td = 10-17 V cm2). First water cluster ions H3O+H2O (at m/z 37.0) and H3O+H2OH2O (m/z 55.0) were also measured as well as NO+ and O2+ masses to indicate any leak into the system and assess the PTR-MS performances.

(O)VOC measurements performed in a full-scan mode were enabled to browse a large range of masses (m/z 30.0–m/z 150.0). Eight protonated target masses were considered here: methanol (m/z = 33.0), acetonitrile (m/z = 42.0), acetaldehyde (m/z = 45.0), acetone (m/z = 59.0), methyl vinyl ketone (MVK) + methacrolein (MACR) + isoprene hydroxy hydroperoxides (ISOPOOHs) (m/z = 71.0), benzene (m/z = 79.0), toluene (m/z = 93.0) and xylenes (m+p-, o-) + C8 aromatics (m/z = 107.0). With a dwell time of 5 s per mass, a mass spectrum was obtained every 2 to 10 min for MEGAPOLI and FRANCIPOL campaigns, respectively. Around 80 % of PTR-MS data were validated. Missing data were partly due to background measurement periods and calibrations. The PTR-MS background for each mass was monitored by sampling zero air through a catalytic converter heated to 250 ∘C to remove chemical species. Daily background values were averaged and subtracted from ambient air measurements.

In order to regularly ensure the analytical stability of the instrument, injections from a standard containing benzene (5.7 ppbv ± 10 %) and toluene (4.1 ppbv ± 10 %) were performed approximately once a month from March to November 2010. These measurements have shown that the analyzer stability remained stable during the year with variations within ±10 %. In addition, two full calibrations were performed before and during the intensive field campaign with a gas calibration unit (GCU, Ionicon Analytik GmbH, Innsbruck, Austria). The standard gas mixture provided by Ionicon contained 17 VOCs at 1 ppmv. These calibration procedures consisted of injecting defined concentrations (in the range from 0 to 10 ppbv) of different chemicals (previously diluted with synthetic air) with a relative humidity at 50 %. Gas calibrations allowed to determine the repeatability of measurements, expressed here as a mean coefficient of variation. This coefficient was less than 5 % for most of the masses. Slightly higher coefficients were observed for m/z 69.0 (isoprene) and m/z 71.0 (crotonaldehyde) with 5.6 and 5.2 %, respectively. Observed differences between both full calibration procedures were from 1.1 % for methanol to 9.8 % for toluene, thus illustrating good analyzer stability over time.

Detection limits (LoDs) were calculated as 3 times the standard deviation of the normalized background counts when measuring from the catalytically converted zero air. For the MEGAPOLI winter campaign, LoDs ranged from 0.020 to 0.317 µg m-3 (0.007–0.238 ppb), whereas they were estimated between 0.020 and 0.330 µg m-3 (0.007–0.248 ppb) during the FRANCIPOL intensive campaign. The analytical uncertainty on all data was estimated by taking into account errors on standard gas, calibrations, blanks, reproducibility/repeatability, linearity and relative humidity parameters. The measurement uncertainty was estimated at ±20 % in agreement with previous studies (Gros et al., 2011; Dolgorouky et al., 2012).

Two different automated gas chromatographs equipped with a flame ionization detector (GC-FID) were used in order to continuously measure light (C2–C6) VOCs in ambient air. The AirmoVOC C2–C6 analyzer (Chromatotec, Saint-Antoine, France), provided by LSCE, was installed near the PTR-MS at the LHVP site from January to February 2010 (e.g., MEGAPOLI period). An in-depth description of the analyzer, sampling setup and technical information (sampling flows, preconcentration, desorption–heating times, types of traps and columns, etc.) can be found in Gros et al. (2011). For each half-hour analysis, more than 20 VOCs were monitored. A certified standard gas mixture (NPL, National Physical Laboratory, Teddington, Middlesex, UK) containing on average 4.00 ppbv of major C2–C9 NMHCs was used for calibration procedures. The injections of this standard allowed checking compound retention times, testing the repeatability of atmospheric measurements and calculating average response factors to calibrate all the measured ambient hydrocarbons. Detection limits were in the range of 0.013 (n-hexane)–0.060 µg m-3 (iso-/n-pentanes) (0.004–0.020 ppb). The analytical uncertainties on LSCE measurements were estimated from laboratory tests (i.e., memory effects, repeatability, accuracy of the gas standard) to be ±15 % (Gros et al., 2011).

From March to November 2010 (e.g., FRANCIPOL period), a GC-FID coupled to a thermo-desorption unit was in operation at the Les Halles subway station monitored by the regional air quality network AIRPARIF. Air samples were drawn up at 2.7 m a.g.l. A total of 29 hydrocarbons from C2–C9 were measured during this experiment. A full calibration was performed once a month with a standard gas mixture containing only propane during 6 h. As the FID response is proportional to the Effective Carbon Number (ECN) in the molecule, calibration coefficients were calculated for each compound and regularly checked so that they drifted no more than ±5 % (tolerance threshold). In addition, a zeroing was carried out every 6 months using a zero air bottle in order to detect any instability or problem with the GC-FID system. LoDs were assessed at 0.024 µg m-3 for all the selected compounds, except for n-hexane (0.013 µg m-3) (0.022–0.004 ppb). We were unable to perform a comprehensive calculation of uncertainties due a lack of sufficient information provided to us by AIRPARIF. Nevertheless, based on previous experimental tests, the NMHC measurements were provided by AIRPARIF with an accuracy of ±15 %, which is the same level of uncertainty as we calculated for the LSCE GC-FID data.

Some ancillary pollutants and parameters were also measured and used as independent tracers with the aim of strengthening the identification of VOC emission sources derived from the receptor modeling.

Black carbon (BC) was measured using a seven-wavelength (370, 470, 520, 590, 660, 880 and 950 nm) AE31 Aethalometer (Magee Scientific Corporation, Berkeley, CA, USA) with a time resolution of 5 min. BC data were acquired by this instrument from 15 January to 10 September 2010. Raw data were corrected using the algorithm described in Weingartner et al. (2003) and Sciare et al. (2011). BC concentrations issued from fossil fuel and wood-burning (BCff and BCwb, respectively) were assessed in accordance with their own absorption coefficients using the Aethalometer model described by Sandradewi et al. (2008).

Carbon monoxide (CO) measurements were performed using an analyzer based on infrared absorption (42i-TL instrument, ThermoFisher Scientific, Franklin, MA, USA) with a time resolution of 5 min. Nitrogen monoxide and dioxide (NO, NO2) were measured by chemiluminescence using an AC31M analyzer (Environment SA, Poissy, France) and ozone (O3) was monitored with an automatic ultraviolet absorption analyzer (41M, Environment SA, Poissy, France). NO, NO2 and O3 measurements were provided with a 1 min time resolution by the local air quality network AIRPARIF. In addition, gas chromatography – mass spectrometry (GC-MS) measurements were performed to measure C3–C7 OVOCs, including aldehydes, ketones, alcohols, ethers and esters during the MEGAPOLI winter campaign. This instrument has been described in detail by Roukos et al. (2009). As the measurement frequency was different for each analyzer, a common average time was defined to get all datasets on a similar time step of 1 h.

Standard meteorological parameters (such as temperature, relative humidity as well as wind speed and direction) were provided by the French national meteorological service Météo-France from continuous measurements recorded at the Paris-Montsouris monitoring station (14th district – 48∘49′ N, 02∘20′ E), located about 2 km away from the LHVP site.

In order to determine the air masses' origin, 5-day back trajectories were calculated every 3 h from the PC based version of the HYbrid Single-Particle Lagrangian Integrated Trajectory (HYSPLIT) model (Stein et al., 2015) with Global Data Assimilation System (GDAS) meteorological field data. Back trajectories were set to end at Paris coordinates (49∘02′ N, 02∘53′ E) at 500 m a.g.l.

Developed about 20 years ago, PMF is an advanced multivariate factor analysis tool widely used to identify and quantify the main sources of atmospheric pollutants. Concerning VOCs, PMF studies have been conducted in urban (e.g., Brown et al., 2007 – Los Angeles, California, USA; Lanz et al., 2008 – Zürich, Switzerland; Morino et al., 2011 – Tokyo, Japan; Yurdakul et al., 2013 – Ankara, Turkey) and rural areas (e.g., Sauvage et al., 2009 – France; Lanz et al., 2009 – Jungfraujoch, Switzerland; Leuchner et al., 2015 – Hohenpeissenberg, Germany). For this current study, the PMF 5.0 software developed by the EPA (Environmental Protection Agency) was used in the robust mode from ambient air VOC measurements from January to November 2010. A more detailed description of this PMF analysis is given in Appendix A.

PMF mathematical theory was extensively described in Paatero and Tapper (1994). Briefly, this statistical method consists in decomposing an initial chemically speciated dataset into factor profiles and contributions. Equation (1) summarizes this principle in its matrix form: X=GF+E, where X is the input chemical dataset matrix, G is the source contribution matrix, F is the source profiles matrix and E is the so-called residual matrix.

The initial chemical database used for this statistical study contains a selection of 19 hydrocarbon species and masses divided into 10 compound families: alkanes (ethane, propane, iso-butane, n-butane, iso-pentane, n-pentane and n-hexane), alkenes (ethylene and propene), alkyne (acetylene), diene (isoprene), aromatics (benzene, toluene, xylenes plus C8 species), alcohol (methanol), nitrile (acetonitrile), aldehyde (acetaldehyde), ketone (acetone) and enones (methyl vinyl ketone, methacrolein and isoprene hydroxy hydroperoxides), which have been measured from 15 January to 22 November 2010 (n = 6445 with a 1 h time resolution). This combination of hydrocarbon species and masses is similar to that from Gaimoz et al. (2011) except for iso-butene. Each missing data point was substituted with the median concentration of the corresponding species over all the measurements and associated with an uncertainty of 4 times the species-specific median, as suggested in Norris et al. (2014). The proportion of missing values ranges from 19 % (especially for compounds measured by PTR-MS) to 41 % (only for isoprene). This high percentage for isoprene can be mainly explained by analytical problems on GC-FID in July. Despite this limitation, it was decided to take into account this compound as isoprene is a key tracer related to biogenic emissions.

The uncertainty matrix was built upon the procedure described by Norris et al. (2014), adapted from Polissar et al. (1998). This matrix requires the method detection limit (MDL, here in µg m-3) and the analytical uncertainty (u, here in percent) for each selected species. MDLs were calculated as 3σ baseline noise and in some cases were homogenized to keep consistency in uncertainty calculations. Species MDLs were ranged from 0.013 to 0.060 µg m-3 (0.004–0.020 ppb) for NMHCs measured by GC-FID and from 0.020 to 0.330 µg m-3 (0.007–0.248 ppb) for VOCs measured by PTR-MS. Their analytical uncertainties were, respectively, estimated at 15 and 20 % and kept constant over the experiments.

Single-species additional uncertainties were also calculated using an equation-based on the signal-to-noise (S/N) ratio. As a first approach, Paatero and Hopke (2003) suggested categorizing a species as “bad” if the S/N ratio was less than 0.2, “weak” if it was between 0.2 and 2 and “strong” if it was greater than 2. Bad variables are excluded from the dataset, weak variables get their uncertainties tripled while uncertainties of strong variables stay unchanged. Here, all species exhibited a S/N ratio greater than 3, except for isoprene which had a ratio of 1.7 due to its 41 % missing values recorded. To address this lack of isoprene data, several empirical tests (e.g., simulating an averaged seasonal/diurnal cycle of isoprene or increasing the analytical uncertainty of raw data from 15 to 30 %) were conducted within PMF simulations with the aim of better modeling the variability of this compound. As a consequence of these tests, no significant improvement on the quality of modeling isoprene was observed. Finally, isoprene is still categorizing as strong here. In addition to isoprene, acetonitrile exhibited a S/N ratio less than 3 (S/N ratio of 2.7). It was the only VOC to be defined as weak because it may be eventually contaminated with local emissions from laboratory exhausts (although visible spikes of acetonitrile were excluded from the initial dataset). Keeping in mind these limitations for isoprene and acetonitrile, it was decided to include these two compounds into the PMF model because they are considered relevant tracers for biogenic and wood-burning activities, respectively. ΣVOC was defined as “total variable” and automatically categorized as weak to lower its influence in the final PMF results. No optional extra modeling uncertainty was applied here.

All the Q values (Qtrue, Qrobust, Qexpected), scaled residuals, predicted vs. observed concentrations interpretation and the physical meaning of factor profiles were investigated to determine the optimum number of factors. Although some mathematical indicators pointed towards a five-factor solution, a source mixing solvent use activities, natural gas and background emissions was detected. In order to split each emission source individually, the six-factor solution was then investigated and chosen in terms of interpretability and fitting scores. More technical details are reported in Appendix A, Sect. A2.

PMF output uncertainties were estimated using two error estimation methods starting with DISP (dQ-controlled displacement of factor elements) and finally processing to BS (classical bootstrap). The DISP analysis results were considered validated: no error could be detected and no drop of Q was observed. As no swap occurred, the six-factor PMF solution was considered sufficiently robust to be used. Bootstrapping was then carried out, executing 100 iterations, using a random seed, a block size of 874 samples and a minimum Pearson correlation coefficient (R value) of 0.6. All the modeled factors were well reproduced through this bootstrap technique over at least 88 ± 2 % of runs, hence highlighting their robustness. A low rotational ambiguity of the reconstructed factors was found by testing different degrees of rotations of the solution using the Fpeak parameter (Fpeak = 0.1).

As emphasized in Paatero and Tapper (1994), a PMF analysis does not require a priori knowledge on the chemical nature of factor profiles. To help strengthen the identification of VOC emission sources derived from this statistical tool, near-field additional measurements (at limited source points) were worthwhile. These in situ observations were performed close to specific local emission sources in real conditions as far as possible. They aim at providing chemical fingerprints (considered as reference speciation profiles) of three significant VOC sources representative within the Paris area: road traffic (i), residential wood-burning (ii) and domestic natural gas consumption (iii). The speciated profiles of these different anthropogenic sources and their representativity are given here. All the technical details of these experiments are reported in Sect. S2 in the Supplement.

Road traffic was considered to be one of the most significant sources of primary hydrocarbons in many megacities (Seoul – Na, 2006; Los Angeles – Brown et al., 2007; Zürich – Lanz et al., 2008), including Paris (Gros et al., 2007, 2011; Gaimoz et al., 2011). The accurate characterization of the vehicle fleet footprint is therefore important. Consequently, near-field VOC measurements (within the PRIMEQUAL–PREQUALIF program) were conducted inside a highway tunnel located about 20 km southeast of inner Paris center in autumn 2012.

This first experiment has the advantage of supplying a realistic assessment of the average chemical composition of vehicular emissions, as these in situ measurements were performed under on-road real driving conditions. Most of VOCs emitted from road traffic are representative of local primary emissions (due to their relatively short lifetimes). Photochemical reactions leading to changes in the initial composition of the air and to the formation of secondary products can be considered of minor importance.

VOC levels during traffic jam periods (07:00–09:00 and 17:30–19:00 LT) were considered as the most representative values of vehicular emissions. In order to omit any local background, nighttime values (as suggested in Ammoura et al., 2014) were subtracted from the peak VOC concentrations. The mass contribution of 19 selected compounds was calculated and reported in Fig. 2. Each compound is expressed in terms of weight out of the weight (w/w) of the total VOC (TVOC) mass. The two predominant species measured inside the highway tunnel were toluene (19.4 %, 27.1 µg m-3 (7.1 ppb) on average) and iso-pentane (18.6 %; 26.0 µg m-3 (8.7 ppb)). The next most abundant VOCs were aromatics (benzene, C8) and oxygenated compounds (acetaldehyde and methanol) accounting for 5.0 to 10.0 % (5–9 µg m-3) (1.1–4.6 ppb). In addition, significant contributions of light alkenes (ethylene, propene, acetylene; 3.3–4.5 %; 4.6–6.2 µg m-3) (4.2–5.4 ppb) and alkanes (such as butanes, n-pentane, n-hexane, methyl alkanes) were also noticed. These observations were found to be consistent with the literature (Na, 2006; Araizaga et al., 2013, concerning NMHCs and Legreid et al., 2007, for OVOCs measurements) and more importantly with the study from Touaty and Bonsang (2000), for which iso-pentane, ethene, acetylene, propene and n-butane were considered as the major aliphatic compounds observed in the same highway tunnel in August 1996 (aromatics and OVOCs were not measured during this study).

Residential wood-burning activities have been shown to be a significant source of (O)VOCs to local indoor and outdoor air pollution during winter months in urban areas (Evtyugina et al., 2014). Currently, only a few studies about the characterization of VOCs from wood-burning have been conducted in Europe (Gustafson et al., 2007; Gaeggeler et al., 2008). After being subject to lively debates within the Île-de-France region, an in-depth investigation of this source would therefore appear necessary to better understand its emission specificities and its potential impacts on atmospheric chemistry.

In order to complete the information on wood-burning activities, VOC measurements (within the CORTEA–CHAMPROBOIS program) were performed at a fireplace facility located ∼ 70 km northeast of the inner Paris center in March 2013. On-line (PTR-MS) and off-line (sampling flasks analyzed later on at the laboratory with a GC-FID) measurements were performed. These in situ observations represent a more qualitative (predominant species identification) than quantitative approach as the resulting speciation profile is based on a limited number of data. As illustrated in Fig. 3, 19 VOC species could be detected. C2 hydrocarbons (ethylene, acetylene), C6–C8 aromatics (benzene, xylenes) and oxygenated species (methanol, acetaldehyde and acetone) can be considered as predominant compounds from domestic wood-burning. This finding is still consistent with intensive field studies of wood-burning performed in Europe (Barrefors and Petersson, 1995; Gaeggeler et al., 2008; Evtyugina et al., 2014; Nalin et al., 2016).

Natural gas is predominately composed of methane (CH4) accounting for at least 80 % of the total chemical composition. It is also a mixture of other pollutants including lightweight VOCs and lower paraffins (approximately 10 % by volume).

In a first approach to determine the speciation profile from natural gas used in Paris, near-field samplings were performed from a domestic gas flue using three stainless-steel flasks, which have been analyzed by GC-FID at the laboratory. Main results (Fig. 4) show a large dominance of alkanes, such as ethane (∼ 80 %), propane (∼ 11 %) and heavier hydrocarbons (like butanes, pentanes) ranging from 4.5 to 0.4 %. Ethane and propane therefore appear as a significant profile signature of natural gas leakages (Passant, 2002).

The speciation profile of Factor 1 (see Fig. 8a) exhibits high contributions of alkanes, such as pentanes (iso-/n-) and n-hexane with on average ∼ 50 % of their variabilities explained by this factor. Aromatic compounds (toluene, xylenes plus C8, benzene; ∼ 35 %) and light alkenes (ethylene, propene), which are considered as typical combustion products, are also the predominant species in this factor.

To evaluate the relevance of this factor, a comparison between speciated profiles from tunnel measurements (Fig. 2) and PMF simulations was done and is reported in Fig. 9. Traffic profiles are in general coherent and consistent amongst themselves, thus allowing to label this factor as a motor vehicle exhaust source. Indeed, a good agreement is observed between these two profiles for the major species such as iso-/n-pentane, toluene, ethylene and propene. Instead, significant differences in mass contributions for ethane (almost a factor of 10), acetylene (considered as a key combustion compound emitted from traffic not identified in the PMF profile), isoprene (represented by evaporative sources) and oxygenated species were found. These differences can potentially be explained by several reasons. Firstly, the proportion of VOC emitted from traffic may be different depending upon the types of vehicles/engines/fuels (Montero et al., 2010). VOC emissions can also be dependent on the use of vehicles (age, maintenance), driving situations and thermal conditions (hot soak). Secondly, the vehicle fleet composition is different in the center of Paris and in the highway tunnel. Although the proportion of passengers cars and light duty vehicles (LDVs) accounts for 60–90 % of the total composition of the fleet in circulation in both cases, the share of two-wheelers and heavy goods vehicles can be different. Indeed, heavy vehicles are subject to traffic limitations prohibiting their entry in Paris, whereas they are allowed in the highway tunnel (5 %). The proportion of two-wheelers is significant in Paris (10–20 %) (Airparif, 2013) while they represent less than 2 % of the total vehicles in the tunnel. Finally, PMF artefacts cannot be excluded. We suppose that the contribution of ethane in the motor vehicle exhaust factor is overestimated (same as the wood-burning factor) at the expense of the mixed natural gas and background source (for which ethane contributions seem to be underestimated).

Factor 1 displays fair correlations with nitrogen monoxide/dioxide (NO / NO2), carbon monoxide (CO) and black carbon (from its fossil fuel fraction), which are known to be relevant vehicle exhaust markers (0.53<r<0.64). The average contribution of this factor is rather stable throughout the year (5.8 µg m-3; see Fig. 10, panel 1 and Sect. S7 in the Supplement). A smaller contribution is found during winter (3.2 µg m-3), whereas the highest emissions from motor vehicle exhaust occur in autumn (8.6 µg m-3 on average) with a contribution of up to 10.1 µg m-3 in September. This seasonal cycle has already been observed and described in Bressi et al. (2014) for the road traffic source of fine aerosols in Paris. The diurnal variation of this source is characterized by a double wave profile with an initial increase at 07:00–10:00 LT and a second increase at the end of the afternoon between 16:00 and 19:00 LT. These increases correspond to morning and evening rush-hour traffic periods. After 21:00 LT, the absolute contributions of this factor stay quite high (7.0–8.0 µg m-3) due to several factors: ongoing emissions (until midnight), lower photochemical reactions and atmospheric dynamics (the shallower boundary layer leads to more accumulation of pollutants at night). Lower contributions are generally displayed during late mornings/early afternoons and at night. This reduction in factor contributions could be mainly explained by dilution and OH oxidation processes of more reactive species, which are not being balanced by additional vehicular emissions. This pronounced cycle has already been reported in previous studies (Gaimoz et al., 2011, and references therein). The temporal source strength variation is usually much more pronounced during weekdays than the weekend.

The profile of Factor 2 (see Fig. 8b) exhibits a high contribution from propane and iso-/n-butanes, with more than 47 % of their variabilities explained by this factor. It was already identified by Gaimoz et al. (2011) and is used here as reference profile from gasoline evaporation emissions (including storage, extraction and distribution of gasoline or liquid petroleum gas (LPG)). The generic term “evaporative sources” is here used to take into account these types of evaporative emissions. Factor 2 also includes a significant proportion of isoprene (20 %). This finding is still consistent with the conclusions of Borbon et al. (2001), which have shown that traffic activities emit a small amount of isoprene. In the same way, oxygenated compounds (acetaldehyde (4 %), acetone (6.6 %)) were found in fugitive evaporative emissions in agreement with what was observed during the highway tunnel experiment (see Fig. 2).

Among independent tracers used, only NO displays a fair agreement with this factor (r=0.35). A correlation between F2 and F1 can also be noted (r=0.36), thus indicating that these two factors are related to a common source (e.g., road traffic). This source is in the range of ∼ 3.9 µg m-3 over the whole studied period (see Fig. 10, panel 2). The annual trend of F2 seems to be consistent with the motor vehicle exhaust factor (F1), even though its monthly change remains ambiguous. Indeed, lower evaporative contributions are recorded both in winter and in early summer with minimum average contributions in June and July (1.7 µg m-3). This finding was already identified by Frachon (2009). However, this value in June is somewhat puzzling as road traffic emissions are usually significant (4.9 µg m-3). In July, propane and butanes (iso-/n-) values were missing due to analytical problems on the operating GC-FID. Consequently, these compounds were simulated by the PMF model (e.g., missing values were virtually substituted by median values) which may underestimate the contribution of this factor during this specific period of time. However, high contributions of this source occur in August (6.6 µg m-3). Although exhaust emissions are not particularly important, this observation could be eventually explained by gasoline storage and distribution sources, which may have increased with higher temperatures during that month. Maxima temperatures have generally been in the range of 16 to 32 ∘C. The source contribution is on average higher in autumn (6.1 µg m-3) with a contribution of up to 6.3 µg m-3 in October.

The diurnal variation of this factor contribution is characterized by a nighttime minimum, an increase from 07:00 to 10:00 LT (consistent with the motor vehicle exhaust factor, F1) and a much slower decrease in emissions during the afternoon than those observed for the vehicle combustion profile. This second factor therefore represents the emissions of less reactive species (OVOCs, propane, butanes), for which concentrations cannot be expected to be consumed photochemically in short transport times. The temporal source strength variation is less pronounced on weekends than weekdays, which is typical of mobile source activity patterns.

According to the Copert IV (European Environment Agency, EEA) program for the calculation of air pollutant emissions from road transport, gasoline evaporation emissions can be explained by the evaporation of VOCs due to temperature, vehicle refueling, running losses, diurnal and hot soak reactions (when a hot engine is switched off). It was speculated that hot engines would emit more in the morning than in the evening, considering typical conditions of active inhabitants going to and from their workplace. Fugitive gasoline emissions from the loading of tank trucks, transportation and unloading from tank trucks at service stations and distributions depots can also be likely sources of this factor. In summary, this source depends on several parameters (related to road traffic conditions, the vehicle fleet composition, economic activities and meteorological observations), which can make the interpretation of its seasonal variability difficult.

In Paris, domestic wood-burning represents a non-negligible part (about 5 %) of the energy consumption by fuel used for home heating (Airparif, 2011). The chemical profile of this source (Factor 3), shown in Fig. 8c, is mainly dominated by acetylene with approximately 80 % of its variability explained by this factor. It also includes ethylene (57.4 %), benzene (22.7 %) and oxygenated compounds, such as acetonitrile, acetaldehyde and methanol (with 18.3, 12.6 and 8.2 %, respectively). Acetonitrile is a hydrocarbon commonly used as a marker of biomass burning (Holzinger et al., 1999). All these chemical species typically reflect an anthropogenic source related to wood combustion processes (Lanz et al., 2008; Leuchner et al., 2015) in agreement with the fireplace emission profile (see Sect. 2.4.2, Fig. 3). No full comparison between both speciation profiles was possible as the fireplace profile was based on a limited number of data. With this mind, only a qualitative approach allowed to identify predominant species emitted from this source and confirm the term “wood-burning” assigned to this factor.

Biomass burning emissions are well correlated with black carbon originating from residential wood-burning (BCwb) and carbon monoxide, a long-lived compound especially emitted from combustion reactions (0.6<r<0.7). In addition, they co-vary well with naphthalene (m/z 129.0 measured by PTR-MS), a known polyaromatic hydrocarbon emitted from combustion processes (industry, tailpipe emissions) including wood-burning (Purvis and McCrillis, 2000). As expected, wood-burning contributions display a distinct cycle with a winter maximum (20.5 µg m-3 on average) and a summer minimum (3.3 µg m-3). Average contributions of this factor are rather stable in both spring and fall (6.9 and 5.9 µg m-3, respectively).

Wood-burning emissions linked to home/building heating are obviously highly dependent on meteorological conditions and particularly on cold temperatures. A clear negative relationship between the wood-burning factor and temperature is found (r = -0.56). The diurnal variation of this source exhibits a double wave profile. Average contributions increase from sunrise to a maximum in midmorning and decrease until 16:00–17:00 LT. At the end of the day, a second increase is observed with another maximum contribution at 19:00–21:00 LT. This diel cycle can be explained by domestic behaviors. An important finding is that the diurnal pattern of this source is fairly comparable to that of the motor vehicle exhaust factor. However, the wood-burning factor does not display any distinct weekly variation. High contributions are observed all week (without any distinction between weekdays and weekends) compared to motor exhausts, for which vehicular emissions are less pronounced on weekends than weekdays. In addition, it exhibits poor correlations with NO, NO2 and BCff (r = 0.30, 0.29 and 0.19, respectively), thus indicating that the wood-burning factor is completely independent of the motor vehicle exhaust source.

The profile of Factor 4 (see Fig. 8d) exhibits a high contribution from isoprene, a known chemical marker of biogenic emissions, with more than 79 % of its variability explained by this factor. In addition, this factor profile includes isoprene's oxidation products (methyl vinyl ketone (MVK), methacrolein (MACR) and isoprene hydroxy hydroperoxides (ISOPOOHs)), methanol and acetone. These oxygenated compounds have a large contribution from biogenic emissions (Kesselmeier and Staudt, 1999; Guenther, 2002). It also accounts a significant contribution of some light alkenes (e.g., ethylene and propene), which can be evenly emitted by plants (Goldstein et al., 1996). Consequently, this factor F4 is termed “biogenic factor”. Amounts of light alkanes (butanes, iso-pentane, n-hexane) and acetonitrile were also found in this profile and could be attributed to a mixing with other temperature-related sources or artefacts from the PMF model (Leuchner et al., 2015).

Biogenic emissions are directly related to solar radiation (Steiner and Goldstein, 2007) and ambient temperature (r>0.7). For that reason, the highest biogenic factor contributions occur in summer (10.5 µg m-3 on average) with a contribution of up to 14.3 µg m-3 in July. Daily mean contributions gradually increase from 09:00 LT. A slight delay is observed in comparison with diurnal temperature/solar radiations variations (for which values increase from sunrise at 06:00 LT). We assume that chemistry affects this factor as it takes part in the formation of secondary species (MVK + MACR + ISOPOOHs, for instance) from the oxidation of primarily emitted compounds (isoprene, OVOC). Diurnal contributions reach their maximum at the end of the day (at 19:00 LT). Highest nighttime contributions of this source can be explained by the presence of oxygenated species (long-lived compounds already present in the atmosphere and/or secondarily formed from the oxidation of isoprene) in the profile combined with lower photochemical reactions and atmospheric dynamics (a low PBL height) at night.

The profile of Factor 5, shown in Fig. 8e, is associated with a large contribution of selected OVOCs (acetaldehyde, methanol and acetone) with on average ∼ 33 % of their variabilities explained by this factor. Significant contributions from aromatic compounds (toluene, xylenes plus C8 and benzene) and some alkanes (pentanes, butanes, propane and n-hexane) are also observed. Toluene, in addition to road traffic, is a good marker for solvents originating from an industrial source (Buzcu and Fraser, 2006). Benzene, due to its toxic and carcinogen nature, was regulated in recent years and was strongly limited in solvent formulations. Current standards establish limits in benzene concentrations at 0.1 % in cleaning products. However, PMF results point out the presence of benzene in this factor, suggesting that this compound might potentially still be in use by some manufacturers. Finally, the presence of these aforementioned species illustrates that this profile could be linked to industrial emissions, although a mixing of different sources cannot be excluded.

This factor co-varies well with ethanol, butan-2-one (also called methyl ethyl ketone – MEK), isopropyl alcohol or even ethyl acetate (0.68>r>0.52, respectively) – four organic compounds that were measured by GC-MS during the MEGAPOLI campaign (January–February 2010). These species are often used as solvents, diluents or cleaning fluids in industrial processes (Zheng et al., 2013). Some manufactories can consume fossil fuels for their activities, which may explain the fairly good correlation between this factor and black carbon originating from fossil fuels (BCff, r=0.50). Indeed, these fossil fuels could be used by industries as diverse as paints, paintings inks and lacquers (Tsai et al., 2001; Cornelissen and Gustafsson, 2004).

The highest contribution of this factor is observed during winter (14.2 µg m-3) with a contribution of up to 15.5 µg m-3 in January. In winter, factor contributions increase at 06:00 and reach their maximum between 11:00 and 19:00 LT (15–20 µg m-3) before a long and gradual decline in the evening (see Fig. 10, panel 5 – top right). Higher contributions in winter can be explained by lower photochemical reactions (combined with weaker OH concentrations/UV radiations) and atmospheric dynamics. Indeed, a shallower PBL (and consequently, less intense vertical dynamics) leads to more accumulation of pollutants and thus to higher source contributions. The daily wintertime variability of this source is in agreement with the diel cycle of independent tracers (ethanol, butan-2-one).

Reconstructed contributions associated with this factor are also significant in summer (12.6 µg m-3 in July), which could be mainly explained by the evaporation of solvent inks, paints and other applications during that month due to higher temperatures. In spring/summer/autumn, factor contributions also increase at sunrise, but reach their maximum between 08:00 and 10:00 LT (typical of anthropogenic activities). They progressively decrease during the afternoon (see Fig. 10, panel 5 – bottom right). This gradual decline (not earlier observed in winter) is influenced by greater photochemical reactions and more intense vertical dynamics during these three seasons, leading to dispersion and dilution processes (and consequently, lower source contributions during the afternoon).

The temporal source strength variation is much more pronounced during weekdays than the weekend, except on Saturday morning. These diel and weekly patterns seem to be consistent with industrial source activities.

The profile of Factor 6, shown in Fig. 8f, is mainly dominated by ethane with around 45 % of its variability explained by this factor. It also contains propane (14.7 %) and light alkanes (butanes), which are key long-lived compounds known to be associated with natural gas leakages. Such species have already been identified in the natural gas experiment (see Sect. 2.4.2, Fig. 4), thus allowing to confirm the identification of this profile. The diel pattern of this factor is mainly based on the diurnal variation of ethane, which is characterized by a nighttime maximum and a midafternoon minimum. Mainly due to its low reactivity, the behavior of ethane can be interpreted as homogeneous species levels during the night under a shallow inversion layer, then followed by concentration reductions caused by the increase of the PBL and vertical mixing – leading to dispersion and dilution processes. Average contributions of this factor were significantly higher when the PBL was low (∼ 11.0–14.0 µg m-3) and lower when the PBL was high (∼ 6.0 µg m-3).

This F6 profile is also characterized by the presence of oxidized pollutants (OVOCs including acetone and methanol) and aromatic compounds (like benzene), which have relatively long atmospheric residence times of respectively 53, 12 and 9 days (assuming OH = 2.0×106 molecules cm-3) (Atkinson, 2000). Because of their low reactivity, all the species of this factor tend to accumulate in the atmosphere and show significant background levels, especially in the Northern Hemisphere. The resulting emissions can be considered as a partly aged background air, implying a possible regional background and/or a long-range (intercontinental) transport.

The average contribution of this mixed source (combining both natural gas and background emissions) is in the range of 9.2 µg m-3 during the whole studied period. Lowest source contributions were observed in winter, which does not fit with those reported in the literature. As mentioned in the motor vehicle exhaust and wood-burning sections (Sect. 3.4.1 and 3.4.3, respectively), PMF artefacts cannot be ruled out. Indeed, a problem with the distribution of ethane (considered as the key species of the mixed source) within PMF factors was raised. We assumed that higher ethane contributions were partly assigned to the motor vehicle exhaust and wood-burning factors. Consequently, we supposed that the natural gas and background factor contributions were underestimated (especially in winter) for the benefits of the wood-burning factor (another source significantly contributing during this season).

The highest contributions occur in spring (13.3 µg m-3) when the Paris region is mostly influenced by prevailing air masses originating from the north and the northeast parts of Europe passing over Germany and the Benelux area (see Fig. 5). These continental imports constitute background events, which significantly impact baseline levels of ethane and oxygenated species. Slightly lower reconstructed mass contributions of this factor F6 were also observed in autumn. This fact can be explained by the consumption of natural gas (for home heating) during this season as average temperatures are progressively going down. No significant continental influences occur during the fall period as main air masses are coming from the west, south and southeast sectors, thus illustrating the importance of local pollution emissions during this season.