AERONET is a global network of ground-based sun-photometers that measure atmospheric aerosol properties (Holben et al., 1998). Cimel CE318, the standard instrument of AERONET, measures direct sun and sky radiances at a number of fixed wavelengths within the visible and near-infrared (VNIR) spectrum to retrieve columnar aerosol optical properties. More precisely, AERONET provides continuous cloud-screened observations of spectral AOD, precipitable water, and inversion aerosol products in diverse aerosol regimes. Inversion products are retrieved from sky radiance measurements performed with different viewing geometries (Dubovik and King, 2000). Together with AOD, these measurements provide inputs for inversion algorithms to derive essential aerosol microphysical and radiative properties such as volume size distribution, refractive index, SSA, and scattering phase function.

In the frame of this study, Level 1.5 (L1.5) Version 3.0 (Giles et al., 2019) AERONET products from the retrieval algorithm were employed for the under-study dust events. The selection of L1.5 instead of L2 products was made to increase data availability, acknowledging the slightly larger uncertainties (0.01–0.02) associated with the absence of full calibration, especially at low AOD conditions. More specifically, AOD along with AE, SSA, Asymmetry Parameter (ASY), and Precipitable water (PW) were used. To further ensure the reliability of aerosol retrievals and minimize potential cloud contamination, we considered only measurements with a sky residual parameter < 5 %.

Broadband solar radiation measurements from ground-based pyranometers were also used in the frame of this study, providing timeseries of GHI and DHI surface solar irradiances. More precisely, the afore-mentioned observations were used to identify the cloudy and clear-sky conditions (for SZA < 80°, GHI-DHI < 120 W m⁻² denoted the presence of clouds) during two selected events for two different stations, Lampedusa and Agia Marina Xyliatou. In addition, they were used to directly validate the model-derived radiation parameters, GHI and DHI, as well as to calculate the produced PV energy. The specific sensors provide high-quality measurements, with uncertainties not exceeding 2 % for hourly totals and 1 % for daily totals over the wavelength range 280–2800 nm at a 1 min temporal resolution.

The dust events across the Mediterranean Basin, which form the primary focus of this study, were identified through AERONET observations collected from all monitoring stations in the region. As a first step, the dust episodes were identified relying on two main criteria: (i) AOD at 500 nm > 0.2 and (ii) AE at 380–500 nm < 0.3 (e.g., Basart et al., 2009). In addition, IASI and MODIS satellite AOD data (Callewaert et al., 2019), as well as ModIs Dust AeroSol (MIDAS; Gkikas et al., 2021) and LIVAS (Amiridis et al., 2015) climatological data were used to confirm the presence of mineral dust particles in the atmosphere. Spatial maps were generated based on these thresholds to detect 20 d intervals in which dust affected more than 10 AERONET stations, allowing the identification of widespread dust activity across the region. Slightly more flexible criteria were used for the identification of dust events originating from the Middle East (at least 6 impacted sites, AE < 0.5). These modified criteria were applied because (1) fewer AERONET stations were at the transfer path of dust, and (2) we noticed that AE during dust events from the Middle East was generally larger than 0.3, possibly due to the presence of smaller particles in the aerosol mixture when dust originates from the Middle East.

A detailed analysis was conducted to identify representative stations, ensuring broad spatial coverage of the region, with priority given to those with available inversion and vertical profile data from lidar or ceilometer observations. The selected stations under study are shown in Fig. 1 (also see Fig. 1 in Papetta et al., 2026a for the different cases in detail). For periods when multiple sub-events occurred, only the most intense was considered.

Furthermore, the arrival height of the dust plumes was estimated from co-located vertical profile observations from ground-based lidars, when available (e.g., from the Aerosol, Clouds and Trace Gases Research Infrastructure and the European Aerosol Research Lidar Network (ACTRIS/EARLINET; Pappalardo et al., 2014). Considering this information, the performance of HYSPLIT (Stein et al., 2015) back-trajectory analysis for the representative stations enabled the confirmation of the sources and transport pathways of the dust plumes.

The four selected dust events covered a wide range of transport regimes and intensities, from a widespread Saharan outbreak with maximum observed daily average AOD reaching 1.08 (Event A), to an efficient mid-tropospheric long-range transport event (Event B), with the maximum daily average AOD reaching 0.97, a spatially confined Middle Eastern outbreak (Event C) with a maximum AOD of 1.17 and a complex multi-source, basin-wide event with moderate dust loading, with daily average AOD not exceeding 1, affecting big part of the Mediterranean (Event D).

While the present study focuses on the radiative effect during these dust events, it is important to note that, as shown in Papetta et al. (2026a), although dust is the dominant species in the aerosol mixture for the studied cases, other aerosol species coexist with dust and affect the effective optical properties that are measured by AERONET. Thus, the presented results correspond to the attenuation of SSR during intense dust events, also considering the role of other pollutants that are either local or are being mixed with dust as it is transferred over the sites of the measurements.

An extended description of the afore-mentioned selected events and the respective analysis on the aerosol properties is given in Papetta et al. (2026a). The periods and AERONET station locations affected during the selected dust events for which the investigation is conducted are presented in Table 1, along with the location that experienced the highest observed AOD, as derived from the daily-mean AOD at 500 nm within the full event period.

Towards estimating the radiative effect of the selected dust events over the afore-mentioned ground-based stations, we performed RT simulations utilizing the uvspec model of the libRadtran package (Mayer and Kylling, 2005; Emde et al., 2016). To solve the radiative transfer equations, the DIScrete Ordinates Radiative Transfer (DISORT) solver was employed (Stamnes et al., 1988, 2000; Buras et al., 2011). The RT calculations were performed in the SW spectral range of 280–3000 nm, relevant to the spectral range of PV systems, using a 16-stream pseudospherical approximation exploiting the extraterrestrial solar spectrum with a 1.0 nm step suggested by Kurucz (1992). The scattering and absorption of solar radiation from atmospheric gases and aerosols, as well as the reflectance of the Earth's surface, were also taken into consideration in the RT simulations. Regarding the molecular absorption, the REPresentative wavelength radiative TRANsfer method (REPTRAN; Buehler et al., 2010; Gasteiger et al., 2014) was employed in the default spectral resolution of 15 cm⁻¹.

The standard US atmospheric profile (Anderson et al., 1986) was utilized for all simulations to parametrize the trace gas concentrations, in which the default profiles of ozone (TCO), water vapor (TCWV), and carbon dioxide (TCCO2) were scaled to explicitly defined columnar values. More precisely, TCWV was adjusted with AERONET PW retrievals while constant climatological values of 410 ppm for CO₂ and 300 DU for O₃ concentrations were set. Solar zenith angle (SZA) and the location coordinates were also provided for each station. The datasets regarding the aerosol and other atmospheric parameters for each station were provided at half-hour intervals, which was selected to be the frequency of the RT simulations in order to produce a complete daily dataset. Finally, the surface albedo was set to 0.2 for all simulations.

To investigate the dust radiative effects, simulations were performed for two configurations (a) based on realistic conditions, separately for the two spectral regions 280–500 and 500–3000 nm (Sect. 2.4) and for (b) the total wavelength range of 280–3000 nm, for which a sensitivity analysis was performed regarding the different input parameters (Sect. 2.5). Finally, the broadband SW GHI and DNI at the bottom of the atmosphere (BOA), considering clear-sky (cloudless) conditions, were calculated in order to further investigate the effect of the different dust events on solar energy production from PV panels. A flowchart illustrating the above-mentioned methodology is presented in Fig. 2. On the left side, the inputs representing the realistic atmospheric conditions (AOD, AE, SSA, ASY and PW) are listed for the two spectral regions as described in (a), allowing the RT simulations to account for the spectral dependence of aerosol radiative effects. On the right side, the parameters (AOD, AE, SSA, ASY, and SZA) used in the frame of a sensitivity analysis described in (b) are presented. A more detailed description of the sensitivity analysis is provided in Sect. 2.6. Finally, the dark and light grey colours indicate the different wavelength ranges of the RT model inputs that were used to calculate the corresponding GHI and DNI, whose integration over the respective spectral intervals yielded the total broadband BOA irradiances. The sensitivity analysis (shown in dark blue) was performed over the full 280-3000 nm wavelength range.

A first set of simulations was performed with realistic aerosol optical properties (AOD, AE, SSA, and ASY) and other atmospheric parameters (e.g., PW) acquired from the AERONET database, as described in Sect. 2.2. The aerosol properties in the RT simulations for the two spectral areas (280–500 nm and 500–3000 nm) were provided at two different wavelengths, with respect to the spectral range.

In the ultraviolet-visible (UV-VIS) range (280–500 nm), AOD, SSA, and ASY retrieved at 440 nm were assumed to be representative and were therefore used in the simulations, while in the visible-near-infrared range (500–3000 nm), AOD at 500 nm and SSA and ASY at 675 nm were applied. The spectral dependence of AOD was accounted for by using the AE over 380–500 nm and 440–675 nm for the respective spectral intervals. Additional atmospheric parameters were specified to better represent the atmospheric conditions, with CO₂ set to 410 ppm and O₃ to 300 DU in all cases.

The different spectral bands of the simulations aimed to capture the effect of the spectrally dependent aerosol optical properties. As previously shown in several studies (e.g., Kazadzis et al., 2016; Mok et al., 2018; Fountoulakis et al., 2019; Papetta et al., 2026a), ASY and SSA generally exhibit greater spectral dependence at wavelengths below 500 nm. A more accurate representation of the spectral variability of these properties would demand measurements with a finer spectral resolution at the full range of the spectral simulations. Such measurements are available at very few stations globally. Given that the most commonly used types of panels are sensitive at wavelengths shorter than approximately 1200 nm (Senthilarasu et al., 2015; Kouklaki et al., 2023), uncertainties due to the inaccurate description of aerosol properties at such wavelengths are not as significant as the uncertainties at shorter wavelengths.

Finally, to quantify the dust radiative effect, the RT simulations were repeated for aerosol-free atmospheric conditions, remaining the rest of the atmospheric conditions unchanged with respect to the initial setup (e.g., PW, CO₂, and O₃). The radiative impact of the local (dust-dominated) aerosol mixtures was estimated through the comparison between the SSR for sky conditions with aerosol load and pristine sky conditions.

Towards the quantification of the impact of dust on PV power potential due to the attenuation of the SSR, we conducted a sensitivity analysis to investigate how different dust optical properties, within the range of measurements during the selected events, affect energy production. The sensitivity relied on different input parameters (AOD at 500 nm, AE for 440–670 nm, SSA at 675 nm, ASY at 675 nm, and SZA) in the simulations, which are presented in Table 2 along with the wavelength that they were derived and used for. Therefore, to account only for the effect of these specifically defined parameters, the water vapor was derived from the US-standard atmosphere profile in all simulations.

The RT simulations in this case were performed for the total wavelength range of 280–3000 nm. Additionally, to save computational time from time-expensive line-by-line calculations of spectral molecular absorption, the sufficiently accurate band parameterization of correlated-k KATO distribution (Kato et al., 1999) was used.

The PV power generation simulations were conducted using the Global Solar Energy Estimator (GSEE; Pfenninger and Staffel, 2016), a widely used open-access model designed for rapid calculations of the PV energy production. The model includes modules that enable the simulation of a complete PV system, incorporating parameters such as the installed capacity, technology (c-Si, CdTe), tracking (fixed, 1- or 2-axis), tilt angle, and azimuth orientation. The user specifies the coordinates of the project location and provides as input a time series of solar radiation data, and optionally, ambient temperature and surface albedo.

For fixed panels, the tilt angles were calculated using the default, built-in latitude linear function included in the GSEE model, considering also a south orientation of the PV panels. More precisely, the optimum tilt angles (OTAs) of the fixed-tilt systems were calculated using the following function (for latitudes > 25° and latitudes < 50°): 1OTA=Lat.⋅0.76+3.1

A detailed table with the OTAs listed for each station along with the related coordinates, is included in the Supplement S1.

Transposition to the PV plane is done considering the GSEE default optimal tilt angle model (http://www.solarpaneltilt.com/#fixed, last access: 1 September 2026) and the BRL output (direct and diffuse components calculated considering solar angles), internally by GSEE. GSEE requires the GHI as input, and optionally DHI to calculate the diffuse fraction (kd): 2kd=DHIGHI

If DHI is not available, the GSEE includes an implementation of the BRL diffuse fraction model. As a next step, GSEE internally calculates the plane-of-array irradiance, by applying trigonometric calculations, which also account for the surface albedo and the ground-reflected component of irradiance. While under moderate aerosol conditions the parameterization of the diffuse component is quite accurate, under high dust loads, BRL underestimates DHI. Under very high dust loads (∼ 1), the underestimation of the diffuse SSR fraction can result in errors in the simulated energy production of the order of 2 % and 5 % for fixed and solar-tracking panes respectively (Papadimitriou et al., 2026). More uncertainties are introduced due to uncertainties in the GSEE input parameters (SSR, surface albedo, temperature), as well as on uncertainties in the parameterization of losses on the panels and the inverter system, and uncertainties related to dust deposition on the panels.

In our simulations, we defined PV plants using both available panel technologies and for the mounting approach, either fixed or 2-axis tracking. Moreover, we calculated the energy production per unit of nominal capacity (1 kWp).

The METAL-WRF system reproduces the relative abundances of the various dust minerals in the atmosphere (Solomos et al., 2023). This is accomplished by coupling the GOCART-AFWA dust emission scheme with prognostic tracers for nine mineral species -illite, kaolinite, smectite, calcite, quartz, feldspar, hematite, gypsum, and phosphorus- as well as iron. Mineral composition data for source regions are derived from the high-resolution GMINER30 (Nickovic et al., 2012) and FERRUM30 (Nickovic et al., 2013) databases. The system simulates the full life cycle of each mineral, accounting for transport, gravitational settling, and a specialized wet-scavenging module developed specifically for the mineral atmospheric life cycle as described in Solomos et al. (2023). As a result, the model produces spatially explicit fields of mineral-specific dust mass, which we use to identify source signatures and track their evolution during transport across the Mediterranean.

In METAL-WRF, aerosol optical properties are computed by combining mineral-resolved dust mass concentrations with wavelength-dependent refractive indices. Optical parameters (extinction, SSA, asymmetry factor) are diagnosed using Mie-based lookup tables and passed to the Rapid Radiation Transfer Model for GCM (RRTMG; Mlawer et al., 1997, 2016; Mlawer and Clough, 1998) to calculate aerosol radiative forcing.

Recently, the METAL-WRF model has been upgraded to incorporate the direct radiative impact of the minerals in dust. The radiative transfer scheme of METAL-WRF has recently been validated in recent work by Spyrou et al. (2025) and revealed good agreement between the model and the measurements. In this work, we compare the PV power potential as it has been estimated by coupling the GSEE model with libRadtran simulations and the METAL-WRF GHI output, for the dust event of April–May 2022.

The results presented in this paper are mainly affected by uncertainties in radiative transfer modelling and uncertainties in solar power potential modelling.

Uncertainties related to RT modelling: Uncertainty in RT simulations lies mostly in the uncertainty in the input parameters (Schwander et al., 1997; Weihs and Webb, 1997). Uncertainties in the used AERONET AOD values are ∼ 0.01 (VIS)–0.02 (UV) (Giles et al., 2019; Sinyuk et al., 2020), while the use of AE at specific wavelength bands adds more uncertainty in the parameterization of AOD since AE does not capture the AOD spectral behaviour perfectly, especially at wavelengths that are outside of the considered AE ranges (e.g., shorter UV wavelengths as discussed in Kazadzis et al., 2016). Overall, these latter uncertainties are less significant in the calculation of the GHI and DNI integrals. Another significant uncertainty factor in RT modelling is the spectral description of the SSA. Uncertainties in AERONET SSA retrievals are of the order of 0.03 for AOD values (at 440 nm) exceeding 0.4 (Dubovik and King, 2000). However, using a fixed SSA value (SSA at 440 nm) for the simulations at wavelengths below 500 nm introduces more uncertainties, since for dust, SSA decreases fast with decreasing wavelength in the UV spectral region (e.g., Kazadzis et al., 2016). The latter assumption results in uncertainties, especially regarding the ratio between the diffuse and direct components of the irradiance, as moving from the VIS to the UV-B region. Nevertheless, the contribution of very short UV-B wavelengths to the DHI and DNI integrals is relatively small, considering the small contribution of UV wavelengths in the VIS region. Uncertainties in the used asymmetry parameter are generally less impactful on the retrieved irradiances relative to uncertainties in AOD and SSA. Similarly, uncertainties in the used aerosol and atmospheric properties profiles are also estimated to be within the overall uncertainty budget of the RT simulations due to the factors discussed above (see e.g., Fountoulakis et al., 2022). An accurate quantification of all the above-mentioned would need a lot of effort and is outside the scope of this study.

Uncertainties related to PV power potential modelling: GSEE models the solar power potential, i.e., the maximum level of energy that can be produced by a specific type of solar panel. Uncertainties in the simulated PV power potential depend primarily on the input irradiances. If both components (direct/diffuse) are known, then uncertainties arise from the fact that the model assumes a standard geometry for the diffuse irradiance, i.e., uses a simple isotropic-type assumption when converting irradiance to the plane of the array. Depending on the atmospheric composition, however, the distribution of the diffuse radiation in the sky changes, resulting in photons reaching the panels from different directions, which changes the efficiency of the diffuse component to produce energy. If only GHI is known, then the distribution in its two components is estimated using Boland-Ridley-Lauret (BRL; e.g., Lauret et al., 2013), which introduces further uncertainties, especially under intense dust events. According to Papadimitriou et al. (2026), the ratio between the diffuse component and GHI can decrease by ∼ 0.4 (relative to its initial value) when retrieved from BRL for a dust optical depth of ∼ 1. Furthermore, GSEE considers a fixed relationship between the total amount of the two solar irradiance components and their spectral composition. Obviously, spectral composition changes depending on atmospheric compositions (e.g., in the presence of small reflective aerosol particles, the contribution of shorter wavelengths in the diffuse component of solar radiation is increased). It is also impossible to parameterize the exact temperature of the different components of a PV installation. Using the ambient temperature as a proxy for the GSEE simulations also adds uncertainties. Since the bibliography is poor regarding the impact of the aforementioned factors, it is not easy to make, even a rough estimation of the overall uncertainty budget. More uncertainties arise if we try to model the actual power production (instead of the power production potential) since factors such as soiling on the panels, module performance parameters, system losses, uncertainties related to panel orientation, etc., can also play a significant role (e.g., Thevenard and Pelland, 2013; Sepúlveda-Oviedo, 2025).

Our analysis across the four selected dust events (March 2021, June 2021, November 2021, and April 2022) revealed a consistent reduction in surface solar radiation and corresponding PV output across all Mediterranean and Middle Eastern stations.

The strongest dust-induced attenuation was systematically observed in the direct component of solar radiation (DNI), with maximum losses frequently exceeding 60 %–80 % during intense dust episodes (Fig. 3a). Stations located closer to major dust source regions or along dominant transport pathways (such as SEDE_BOKER, Cairo_EMA_2, CUT-TEPAK, and Ben_Salem) experienced the most severe reductions, particularly during March 2021 and June 2021 events. This denotes that a decrease in DNI reflects the strong extinction of the direct solar beam by mineral dust through a combination of scattering and absorption processes, which becomes increasingly efficient as dust optical depth increases. In contrast, GHI exhibited more moderate reductions, typically ranging between 5 % and 25 % across most of the stations (Fig. 3b). This behaviour highlights the compensating role of the diffuse radiation component under dusty conditions, whereby part of the attenuated direct radiation is redistributed into the diffuse field, partially offsetting total GHI losses.

The afore-mentioned combined radiative effects were reflected directly into substantial reductions in PV power output (Fig. 4) for fixed-tilt and two-axis tracking systems.

For fixed-tilt PV systems (Tracking = 0, Fig. 4a), maximum daily power losses typically ranged between 10 % and 45 %, depending on station location and event intensity. Observed losses were systematically higher during March and June 2021 events, in agreement with the elevated dust loads and stronger DNI attenuation observed during these periods. For two-axis tracking systems (Tracking = 2, Fig. 4b), PV power losses were significantly amplified, frequently exceeding 50 %–80 % at stations strongly affected by dust intrusions. This pronounced sensitivity arises from the strong dependence of tracking systems on the direct solar component, which is disproportionately reduced during dust events. Consequently, although tracking systems generally outperform fixed installations under clear-sky conditions, their relative vulnerability to intense dust episodes becomes evident during high-AOD conditions. The resulting irradiance and PV energy losses are expressed relative to aerosol-free atmospheric conditions. More specifically, the afore-mentioned losses were calculated based on the following formula: 3Losses%=Rclear-sky-Rall-skyRall-sky⋅100 where R denotes the respective parameter in each case; DNI and GHI for radiation losses (presented in Fig. 3) and solar energy for PV losses (presented in Fig. 4).

To further investigate the relationship between solar energy losses and the corresponding global and direct radiation losses, we directly compared the findings presented in Figs. 3 and 4 for both fixed-tilt and two-axis tracking systems. The analysis incorporates all events with respect to the respective maximum (daily average) AOD (Fig. 4c) to identify differences in system performance due to changing atmospheric conditions. Overall, the afore-mentioned comparison (Fig. 5) demonstrates a consistent relationship between PV solar energy and radiation losses, for both the global and direct components, as previously presented in Fig. 4. More precisely, solar energy losses increased systematically for both system configurations with increasing radiation losses, as expected, since radiative attenuation is the primary driver of PV performance degradation under these conditions. Solar energy losses refer to the decrease in PV energy production. The percentage losses have been calculated relative to the PV output under aerosol-free skies (Eq. 3), with all other conditions kept identical.

In the case of direct irradiance, this effect is more pronounced (exhibiting losses of up to ∼ 80 %, corresponding to approximately 55 % higher than global radiation losses), indicating a tighter dependence than that of the diffuse component. Additionally, increasing AOD is consistently associated with higher losses in radiation and, therefore, energy, while the observed residual variability suggests that other parameters might also be associated with these observations. Finally, two-axis PV tracking systems exhibit higher solar energy losses for comparable radiation attenuation than fixed-tilt systems, reflecting their sensitivity to these effects, and more precisely their stronger dependence on the direct irradiance under high aerosol loading conditions in the atmosphere.

In addition to the analysis described in Sect. 3.1 based on real atmospheric conditions during the four dust events in the Mediterranean Basin, a sensitivity analysis was performed to quantify how aerosol optical properties and solar geometry jointly influence PV energy production. The simulations considered representative aerosol conditions of Mediterranean dust intrusions, as described in Sect. 2.6 and Table 2, and evaluated their radiative impact across a wide range of SZAs. Unless otherwise stated, for the analysis, we assumed a south-oriented PV system (azimuth 180°), which corresponds to the typical optimal configuration for fixed PV installations in the Northern Hemisphere.

The sensitivity analysis was structured to isolate the influence of key atmospheric and geometrical parameters such as latitude, AOD, AE, and PV azimuth orientation. The resulting PV energy losses, expressed relative to aerosol-free atmospheric conditions based on Eq. (3), are summarized in this section.

Across all scenarios, SZA emerges as the dominant driver of the observed PV energy losses. As shown in the figures of this section and in the Supplement (Figs. S2–5), aerosol-induced attenuation remained relatively small under high solar elevation conditions (SZA lower than approximately 30–35°), in most cases leading to losses that do not exceed 10 %–15 %. However, SZA increases toward early morning, late afternoon, or wintertime conditions, leading to substantial atmospheric optical path length increases, resulting in progressively stronger attenuation of solar radiation and, therefore, energy output. Consequently, PV energy losses increased rapidly with increasing SZA, particularly under enhanced aerosol loading.

The latitude sensitivity analysis (Fig. S2), conducted for 25, 35, and 45° N under identical aerosol conditions, revealed only minor differences in the resulting PV energy losses when expressed as a function of SZA. In order to further quantify the relative importance of each parameter, additional difference maps were computed, representing the incremental change in PV energy losses between selected parameter ranges while keeping all other variables fixed. This indicates that, for a given SZA, aerosol-induced radiative attenuation is primarily governed by solar geometry and relatively independent of geographic latitude. Although in general higher latitudes experience larger SZAs on average throughout the year, the modest variations in solar incidence angles on the plane of the PV panels across the examined, relatively narrow latitudinal range, representative of the Mediterranean region, lead to nearly identical instantaneous PV energy losses, as they are primarily governed by the incident solar geometry. The corresponding difference analysis confirms that inter-latitudinal variations remain small, typically not exceeding 5 %–6 % across the examined conditions Fig. 6. However, some differences due to latitude may, in real atmospheric conditions, also derive from other parameters, such as water vapor and temperature, which are not accounted for in the present sensitivity analysis.

In contrast, the AOD sensitivity analysis clearly demonstrates the strong amplification of PV energy losses with increasing aerosol concentration (Fig. S3). Based on the findings of the latitudinal sensitivity and the negligible differences across the different cases, a representative latitude of 35° N was adopted for all subsequent sensitivity analysis. For relatively moderate aerosol levels (AOD = 0.2), losses remain limited at low and moderate SZAs but increase significantly beyond approximately 55–60°. Under elevated dust concentrations (AOD = 0.6), strong attenuation develops already at intermediate SZAs (45–50°), while losses exceeding 60 %–70 % occur at the highest SZAs considered. For extreme aerosol loading (AOD = 1.0), PV losses become substantial across nearly the entire SZA range. This behaviour is further quantified by the difference analysis in Fig. 7, which shows that increasing AOD from 0.2 to 1.0 results in additional PV energy losses exceeding 40 % at large SZAs. Even moderate increases (e.g., AOD 0.6 relative to 0.2) lead to differences on the order of 20 %–30 %, particularly beyond SZA ≈ 50°.

Investigation of the AE impact (Fig. S4) provides insight into the role of aerosol particle size distribution, comparing coarse-mode dominated dust conditions (AE = 0.0–0.2) with slightly more mixed aerosol regimes (AE = 0.2–0.4). The results indicate relatively modest differences between the two regimes, with the coarse-mode dominated cases generally producing slightly higher PV energy losses, particularly at large SZAs. The difference analysis in Fig. 8 indicates that variations in AE result in changes in PV energy losses generally below 3 %–4 %, with slightly higher losses associated with coarse-mode particles at large SZAs.

Finally, regarding the PV orientation sensitivity analysis (Fig. S5), we compared east-facing (90°), southeast-facing (130°), and south-facing (180°) configurations. The results indicate moderate variations in PV energy losses among the different orientations, with the south-facing configuration generally experiencing slightly larger aerosol-induced losses at high SZAs arising from its stronger alignment with the solar trajectory during peak irradiance periods. The difference analysis (Fig. 9) shows that orientation effects can reach up to approximately 30 % at high SZAs when comparing east- and south-facing configurations.

In this section, we compare the PV output that has been estimated from the modelled irradiances with the PV output that has been estimated from ground-based measurements. We focus on two stations affected by dust events of different origin, Sahara (Lampedusa, during event D) and the Middle East (Agia Marina Xyliatou-AMX, during event C), characterized by different optical properties. In particular, we compare modelled PV outputs when GSEE inputs are irradiances that have been measured at the surface level, and irradiances from libRadtran (retrieved as described in Sect. 2.3) and the METAL-WRF model.

In previous works (Spyrou et al., 2025; Papetta et al., 2026a), the efficiency of METAL-WRF in reproducing dust optical properties, as well as in simulating the GHI using its RT scheme, has been evaluated. In this work, we try to assess how accurately it can quantify solar power potential when it is coupled with a PV output model (GSEE), with respect to the reference case where GSEE inputs are ground-based measurements.

As an additional step, we compared the daily PV output for three different cases, when calculated based on ground-based measurements, METAL-WRF, and libRadtran, respectively. For the comparison, we used measurements of the GHI and DHI surface solar irradiance from pyranometers operating at the two selected stations. For the selected period, data under all atmospheric conditions are presented, cloudy and cloud-free days were identified using the temporal variability of the measured GHI (shown in Fig. 10). More precisely, days were identified as clear-sky when GHI and DHI components exceeded 120 W m⁻² for SZAs < 80° and exhibited a smooth diurnal curve with no rapid fluctuations. For instance, in the case of AMX, during days 2, 3, 5 and 7 clear-sky conditions dominated, while in the case of Lampedusa, days 1, 6, 7, 8 and 9 indicated cloudy conditions.

The evolution of the AOD at 500 nm from AERONET and the DOD at 500 nm from METAL-WRF is shown in Fig. 11. In the case of AMX, METAL-WRF underestimates AOD values, although it follows the evolution of the event. As discussed in Papetta et al. (2026a), this is possibly due to anthropogenic particles that are included in the mixture and not considered by METAL-WRF. In the case of Lampedusa, dust in the aerosol mixture is more dominant. DOD from METAL-WRF follows the AOD closely, although differences are occasionally large, especially when AOD is very high. This is associated with the uncertainties in the description of the timing and the magnitude of the events (Solomos et al., 2023).

A comparison between the daily GHI from libRadtran and METAL-WRF for the clear-sky days only of the event is shown in Fig. 12. For the daily GHI values, the full dataset was used without cloudy conditions being filtered out, since METAL-WRF represents all atmospheric conditions. During cloudless sky days (with respect to the ones indicated as clear-sky based on Fig. 10), the comparison between modelled and measured GHI revealed an agreement in most cases better than 10 %. Comparisons for cloudy days were also performed but are not shown here since libRadtran simulations were only performed under clear-sky conditions and the effect of clouds is out of the scope of this study. However, the results between METAL-WRF GHI (which accounts for cloud presence) and ground-based GHI revealed a good agreement for cloudy sky conditions as well.

To further investigate the behaviour of the different irradiance components, such as GHI and DNI from libRadtran and METAL-WRF, we focused on two mostly clear-sky days, one for AMX and one for Lampedusa station. For each case study, we present the ratio between modelled (from libRadtran and METAL-WRF) and measured values of the afore-mentioned components during the selected days (Fig. 13). More precisely, for AMX, the second day of the event is presented, while for Lampedusa, the fifth day was selected (as indicated in Fig. 12), corresponding to 11 November 2021 and 25 April 2022, respectively. In terms of aerosols, in the case of Lampedusa, the load was significantly more enhanced than in the case of AMX, reaching approximately 0.8 compared to 0.07 based on AERONET retrievals (Fig. 11). Based on complementary ground-based observations (Figs. 6–7), an aerosol layer was detected up to 4 km altitude in the case of Lampedusa, whereas in the case of AMX, it did not exceed 2 km. The timeseries of the ratios during the selected cases revealed an overall quite good agreement between modelled and measured parameters, with larger differences being observed for higher SZA's, specifically in the case of AMX, while a threshold of SZA < 85° was applied in the comparisons of both cases.

For AMX, the comparison with libRadtran revealed a very good agreement, slightly overestimating the GHI throughout the entire day, while the METAL-WRF model revealed larger deviations for larger SZAs. However, this is not reflected in the daily ratios previously illustrated in Fig. 12. For Lampedusa, the behaviour of the two models was reversed, with the METAL-WRF to measured ratio very close to unity overall, but larger deviations in the morning and the afternoon, when collocated ground-based Ceilometer observations revealed the presence of clouds (see Fig. S7).

In both stations, the DNI component revealed higher deviations than GHI. Considering that the direct component is primarily affected by AOD and the diffuse component is also influenced by other aerosol optical properties, the respective ratios denote that these parameters may not be fully represented in the simulations. Moreover, these effects are more pronounced for higher AOD values, which explains the more pronounced deviations in the case of Lampedusa. For lower aerosol load conditions, additional cases were analysed, which confirmed the findings illustrated for AMX, however, these results are not presented in this study.

The ratio between the PV output from METAL-WRF (PV_wrf) and measurements, as well as the ratio between PV output from libRadtran (PV_lbr) and measurements, is presented in Fig. 14. PV output from libRadtran and from METAL-WRF has been simulated using only GHI as input to GSEE. The PV output from ground-based measurements has been retrieved using either only GHI (PV_ghi), or both GHI and DHI (PV_{ghi_dhi}). In addition, all the comparisons have been conducted for two PV orientations (optimally inclined: axis = 0 and installed on a solar tracker that follows the sun: axis = 2).

Comparing PV_wrf and PV_lbr with PV_ghi provides an estimate of how the differences between the modelled and the measured GHI impact the modelled PV output. Comparing PV_wrf and PV_lbr with PV_{ghi_dhi} results in generally larger deviations from unity (with respect to the ratios with PV_ghi), which are due to the fact that only the GHI has been used to simulate PV_wrf and PV_lbr, i.e., uncertainties in the parameterization of the diffuse component in GSEE when DHI is not available (Papadimitriou et al., 2026).

Although GHI values and model to measurement ratios of GHI have previously been presented for the entire period of the selected dust events for both AMX and Lampedusa (Figs. 10 and 12), in the case of Lampedusa some of the clear-sky days are not presented in Fig. 14. More specifically, days 1 and 3 have been excluded, to avoid biases into the analysis due to missing DHI (or GHI and DHI) measurements that did not allow the calculation of the daily integrals.

In (mostly) cloudless days, for static PVs (axis = 0), the ratios between PV_lbr (or PV_wrf) and PV_ghi are very similar to the ratios between PV_lbr (or PV_wrf) and PV_{ghi_dhi}, showing that even under intense dust events, not including the DHI component in GSEE simulations is not critical for static, optimally inclined PVs. Differences become slightly more significant for the solar tracking PVs, generally not exceeding 10 %.

Over AMX (event C), the dust plume is mixed with fine aerosols, as also discussed in Papetta et al. (2026a). The fine aerosol component is not captured by METAL-WRF, which generally results in overestimation of the GHI, and subsequently an average overestimation in the daily PV output of the order of 10 % (for both fixed PVs and PVs on solar tracking systems) during cloudless days. Over Lampedusa (event D), the aerosol mixture is constituted mainly by dust (see also Papetta et al., 2026a), and the event is much stronger relative to the event over AMX. However, the cloudless sky days are very few to estimate the impact of this variability. In addition, the observed relative differences of AOD at 500 nm from AERONET and METAL-WRF (presented earlier in Fig. 11) are reflected in the ratios of Fig. 14 for the clear-sky conditions. More precisely, for AMX, the difference of the AOD retrievals is more pronounced during the fifth day compared to the second day, which can also explain the lower model-to-measurement ratio differences between the two models during the latter. The same was observed in the case of Lampedusa.

The chemical composition and microphysical properties of dust, and consequently its optical properties, strongly depend on its source region. Therefore, dust originating from the same source region is expected to exert a similar influence on surface solar radiation (SSR) and, consequently, on PV power output. In this context, the selected cases can be regarded as representative of dust events affecting the Mediterranean Basin, rather than as isolated case studies. Such events regularly influence multiple regions across the Basin through recurring transport pathways, making them suitable case studies for assessing regional-scale impacts.