Comment l'adaptation au site permet de réduire l'incertitude liée au rendement énergétique

For the analysis of solar resource interannual variability and associated uncertainty we use the long time series depending on satellite data availability. 

The long time series allow us to get some idea of trends and if the trend is present the uncertainty from interannual variability will capture it.

In the Solargis Evaluate approach for PV simulation we recommend to do the simulation on historical time series, as it will properly propagate into simulation results any trend which comes from solar data, but also other meteorological parameters such as air temperature, water vapor or wind speed.

Such PV simulations give more realistic combined effect of any trends coming from the primary solar and meteorological data. 

Ground measurements are a must for site adaptation. We require GHI (and ideally DNI and DIF) measurements from a pyranometer, and ideally also meteorological parameters such as air temperature, wind speed, precipitation, relative humidity - these also affect the PV yield.

It is very difficult to provide an overview like this, because it depends greatly on the quality of the measured data (instruments, cleaning, calibration, O&M), and the uncertainty of the satellite-based solar model data (which is driven to a great extent by local geographical conditions and weather patterns).

But overall, we can go from 4-8% uncertainty of the satellite model data to 3-5% uncertainty of the site-adapted data. In principle there is no regional dependency (except high mountains or very high latitudes e.g. 60+deg.), if the high quality measurements is available. 

We don't have actual examples, as financing details are typically private information, which we don't get access to. Sample calculations, like the one presented in the webinar, can be made using the uncertainty of the data, and assumptions about project finance variables (energy price, DSCR, debt maturity period, ...).

We use measurements of GHI, DNI, and meteorological variables. All measurements must be of sufficient quality to be used. 

Diffuse (DIF) light is adapted indirectly by adaptation of GHI (and if the DNI measurements are available, also by adaptation of DNI), by recomputing to maintain consistency between GHI, DNI, and DIF.

In addition, our site adaptation techniques take into consideration consistency between GHI and DNI, to avoid physically imposible results such as a negative DIF. 

Satellite-based model data without site adaptation can be used in pre-feasibility, if ground measurements are not available, and then upgraded to site-adapted data in the feasibility stage once measurements are available.

However, you should be careful about the uncertainty and validation of the satellite model data - high uncertainty of satellite data can lead to a large shift after adaptation.

The indication of performance of model in given region can be inferred from the model validation for the sites in that region and surrounding regions with similar environmental conditions.   

P90 and P50 are probabilistic scenarios used in construction of TMY datasets. The number designates the percentage of real cases (years) in which the P90/P50 value of solar resource or PV yield is expected to be exceeded.

The calculation of P90 uses concept of combined uncertainty which merges the uncertainty of the long term model values and uncertainty coming from the interannual variability.

For example, a given P90 yearly sum of solar irradiance can be expected to be exceeded with 90% probability. 

You can read more about probabilistic scenarios here.

Pay attention to the quality indicators the provider (LTA) uses to describe the data (ideally, they should report more than just bias), how uncertainty of the dataset is estimated and whether that uncertainty is realistic (keep in mind the 3% physical limit of uncertainty), how is the ground-measured data processed and quality controlled, and enquire about documentation of the site adaptation process, and the validation of the satellite-based model data (the provider should be transparent).

Besides reduced bias, there are many additional criteria that have to be checked in the site-adaptation results, for example, consistency of GHI-DNI-DIF values, consistency of diurnal profiles, fit of cumulative distribution, seasonal consistency, long-term consistency and others. The site-adaptation should not be focussed just on the bias reduction. 

Yes, this is possible, although rare. If the original satellite model data overestimated the solar resource at the site, the average (i.e. P50) of the site-adapted data will be lower. If this decrease is larger than the increase in P90 due to lower uncertainty post site adaptation, the site-adapted P90 will be lower than the original P90.

However, the adapted P90 will be more accurate, and represent the actual solar resource at the site much better, and the uncertainty of the site adapted data will be much lower, reducing also risks of using the site adapted data. 

There is no target R² value. The quality control of the measurements ensures that all erroneous measurements are excluded. The rest can be considered correct, regardless of their correlation with the satellite model data.

As it was mentioned above, there are several criteria that have to be balanced. The site-adaptation should not lead to situation that one criterion is improved and others are worse then those of original data before site-adaptation.

Depends on the original mismatch of the measurement and model values, which determines the selection of the proper site adaptation method. 

In more complex methods (e.g. adaptation of model input AOD data or cloud properties), the adaptation modifies consistently both GHI and DNI data.

Simpler post-processing methods which are appropriate for correction of smaller mismatch, treat each parameter separately.

But also for these simpler approaches a site adaptation parametrization should be done in  a way that the DNI-GHI consistency is not broken and is within the margins of the physically possible combinations. GHI and DNI adaptation affects DIF, which is indirectly adapted by recomputing to maintain consistency between GHI, DNI, and DIF.

If the ground measurements are short, and seasonal variability cannot be understood well, the uncertainty reduction from the original satellite model data to the site-adapted data will be smaller.

In extreme cases site adaptation can be refused, as the ground measurements do not contain enough information to decrease the uncertainty of the satellite model data.

No, we would not recommend it - as mentioned in the webinar, ground measurements only cover ~1 year, which does not contain information about the variability and trends of solar and meteorological parameters at the site.

Using this data in PV yield estimates would likely lead to inaccuracy (under-/overestimation), and would not allow for consideration of realistic extreme events.

If they are well-maintained, Class B instruments can be used for site adaptation, although we recommend using Class A instruments. The space for uncertainty reduction of site adapted model data is in this case limited, if any.

Using the full time series of adapted data is the better approach, allowing much deeper analysis of variability. However, the industry due to many reasons still prefers to work with P90 scenario, at least in some cases.

The uncertainty is quantified separately for GHI and DNI. If DNI measurements are avaialable, we can adapt DNI, and reduce its uncertainty in addition to GHI.

Moreover, DNI measurements allow for adaptation of aerosols, and can help in reducing the GHI uncertainty.

DIF measurements can also help identify measurement issues during the quality control process, and hence improve the uncertainty of both GHI and DNI.

The limiting factor for the site-adaptation is the period of available ground measurements. As the models can have different performance for individual seasons, the basic requirement for the site adaptation is to have coverage of all seasons by the measurements after quality control.

In other words, the minimum requirement for the site adaptation is one year of measurements, optimally two years and more. This prevents from the overfitting for single season.

The basic assumption for the site adaptation is stability of corrections (derived for the ground measurements period) over the full history of model data. Apparently this is a simplification of the reality.

The site adaptation should be capable to reduce any systematic present in the original model data, therefore one of the primary steps of the site-adaptation is inspection of both

1) measurements and 

2) model for any anomalies.

For the model data we inspect mainly temporal stability, and for measurement we look for anomalies that may not be representative for the whole history (e.g. wild fires, volcano effects).

Only if these preconditions are fulfilled, the proper site-adaptation can be applied.

In addition, if multi-year measurements are available, we explore the year-by-year stability of model validation indicators (e.g. bias, RMSD).

Finally, the uncertainty of the site adapted data takes into account the risk of the overfitting which fades out with longer time period. Thus the achievable uncertainty of site adapted data decreases with longer period of available mesurements.

Ground measurements and satellite data (before and after adaptation) will always overlap. Their difference will depend on the accuracy metrics (bias, RMSD, KSI) of the satellite model data after site adaptation - the yearly sum will specifically depend on the bias of the site-adapted data.

Corrections developed for the overlapping period (usually 1-2 years) are propagated for the full history time series (30+ years of data), which after site-adaptation better represent local conditions and have reduced uncertainty.

The calculation of the basic solar radiation components (GHI, DNI, DIF) does not use albedo.

Albedo becomes relevant in the calculation of global tilted irradiance (GTI, also known as plane-of-array or POA irradiance), when the reflections from the ground and other surfaces need to be calculated.

The sensitivity to the albedo data depends on the GTI calculation approach (view factor vs. raytracing).