Synthetic data generation

PVsyst can generate hourly and sub-hourly weather data from the monthly values saved in a geographic site (.SIT file). The generation is performed using the Meteonorm 9.0 algorithm included in PVsyst, which is based on the model proposed by Collares-Pereira¹ (for irradiances) and Scartezzini² (for temperatures), but modernized.

The detailed description of the Meteonorm algorithm can be found in their theory manual. This page gives a simplified summary of the method.

Synthetic irradiance

Daily values

Daily global horizontal irradiance is generated using a Markov chain model, building on Aguiar et al.³ but reformulated on a clear-sky clearness index basis rather than the classical clearness index. This reformulation was introduced to ensure consistency with the ESRA clear-sky model⁴ used elsewhere in the chain. The Markov transition matrices (9×10×10) were recalibrated using 121 stations spanning all major climate zones.

Hourly values

Hourly values are generated from daily values using the TAG model¹ : a time-dependent autoregressive Gaussian model. It superimposes a deterministic average daily profile — scaled from the clear-sky hourly profile — with a first-order AR(1) stochastic perturbation. The autocorrelation of the AR(1) process is itself a function of the daily clearness index Kt.

Global horizontal irradiance is split into direct normal (DNI) and diffuse horizontal components using the Perez model⁵.

Minute values

The default minute model used is the Remund time series model⁶. The model was trained on 15 BSRN stations across all climate zones. It stores 20 randomly selected measured time series of minute irradiance, normalised by clear-sky irradiance, classified by wind speed (3 classes), cloud cover (10 classes), and solar elevation (5 classes). Synthetic time series are then generated from hourly values by selecting one of these 20 stored profiles depending on weather conditions.

Synthetic temperature

The hourly temperature sequence is generated using a two-stage stochastic model:

1. Daily mean temperature is generated by a first-order autoregressive (AR(1)) process, which introduces realistic day-to-day persistence while preserving the monthly mean and standard deviation for each month.

2. Hourly profile within the day is generated in a way that is physically consistent with the hourly GHI sequence: temperature rises after sunrise and falls after sunset, with the amplitude and timing derived from the irradiance profile. This replaces the assumption of a sinusoidal profile used in older models.

The Meteonorm model is calibrated on a worldwide station dataset, making it applicable to any climate.

Other variables

Wind Velocity

Hourly wind speed is generated at 10 m above ground using a combined deterministic–stochastic model. The daily model uses tabulated normalised hourly profiles for seven terrain and climate categories — ranging from open mid-latitude Europe to alpine valleys, lake sites, tropical regions, and continental USA — conditioned on the daily clearness index K_t and daily global radiation. The model was validated against 30 stations in the USA and 20 in Switzerland, showing good agreement in distribution shape and monthly means.

Note that wind speed is strongly influenced by local topography and that the generated values are not intended for wind power plant design.

Relative humidity

Relative humidity (RH) is generated as a supplementary parameter from the synthesised dew point temperature and air temperature, via standard saturation vapour pressure relationships,⁷. Generation begins with a sunrise-hour reference value derived from the monthly mean. This baseline is then adjusted for three physical factors: nocturnal cloud cover, precipitation history over the preceding days, and a stochastic white-noise term.

Linke coefficient

The Linke turbidity coefficient is a dimensionless index summarising the total optical depth of the atmosphere relative to a clean, dry Rayleigh atmosphere. Monthly mean values are drawn from a global climatology compiled by Solar Consulting Services⁸, derived from MODIS and MISR satellite measurements over the period 2000–2015 and gridded at 0.5 arc-second resolution. These gridded values are altitude-corrected. To reproduce realistic day-to-day variability, the daily values are then perturbed stochastically using a first-order autoregressive process.

PVsyst renormalization

When using the Synthetic Data generation tool monthly weather values are passed to the meteonorm algorithm to generate synthetic hourly or sub-hourly values. However, the sum of these generated synthetic values does not always match the original inputs. In these cases, PVsyst applies a renormalization to scale the generated values so tha monthly averages match.

Updates

The Meteonorm generative algorithm is updated regularly.

• PVsyst 8.1 uses the Meteonorm 9.0 algorithm
• PVsyst 8.0.X used the Meteonorm 8.2 algorithm
• PVsyst 7.3.X and PVsyst 7.4.X used the Meteonorm 8.1 algorithm
• PVsyst 7.2.X used the Meteonorm 8.0 algorithm
• PVsyst 7.0 to 7.1 used the Meteonorm 7.3 algorithm
• PVsyst 6.76 until PVsyst 7.0 used the Meteonorm 7.2 algorithm

PVsyst prior to 6.76

In very early versions, the generation was done by PVsyst directly as follows:

• Global Horizontal Irradiance: For the irradiance, the generation of GHI hourly values from monthly averages was performed using stochastic models developed by the Collares-Pereira¹ team in the 1980s. These models generated a sequence of days, and then a sequence of hours in the day, using Markov transition matrices. These matrices were established in order to produce an hourly sequence, with distributions and statistical properties analogous to real hourly weather data measured on more than 30 sites all around the world.
• Diffuse Horizontal Irradiance: When the beam horizontal or the normal beam were available in the original data, they were used to calculate the diffuse part. Otherwise, the program used the Erbs correlation⁹. At the end of each month, the diffuse values were renormalized in order to match the specified monthly diffuse (scaling of the Kt value): this was not the case in version 4, and in middle Europe the Erbs correlation had a tendency to over-estimate the diffuse.