Overview

Note

The rainfall model has been subject to more development and testing than the weather generator/model so far. Please check back for updates to the model and documentation.

There are three key “objects” available within RWGEN:

  1. Weather generator

  2. Rainfall model

  3. Weather model

The weather generator is used to perform coupled simulations of rainfall and other weather variables (e.g. temperature). To do this, the weather generator uses the other two objects as its component models:

  1. Rainfall model - Neyman-Scott Rectangular Pulse (NSRP) process

  2. Weather model - regression equations

Regressions in the weather model are conducted according to wet/dry transition states. This means that the weather model depends on the output of the rainfall model (whereas the rainfall model does not depend on the weather model). The rainfall model may be used as a standalone model, but the weather model will typically be used as a component of the weather generator.

The rest of this page explains the core concepts of the weather model component. The rainfall model is described on the rainfall model Overview page, while an overview of usage of the weather generator is provided on its Workflow page.

Single Site Weather Model

The weather model largely follows the structure described by Kilsby et al. (2007) and Jones et al. (2010).

Transformations

Before fitting the regression models, transformations are used to help non-normal variables better approximate a normal distribution. The Box-Cox transformation is used, apart from for sunshine duration (for which a Beta distribution was selected). The weather input time series are additionally standardised/scaled to follow a standard normal distribution (mean of 0 and standard deviation of 1).

Regressions

Regression equations are used to model temperature, vapour pressure, sunshine duration and wind speed based on their previous values and the precipitation “transition state”. The transition state describes whether a day and the preceding day are both wet, dry or different to each other. Five transition states are used in the model:

  • Dry to dry (DD)

  • Dry to wet (DW)

  • Wet to dry (WD)

  • Wet to wet (WW)

  • Dry to dry to dry (DDD)

The final state listed above (DDD) considers the previous two days, rather than just the preceding day. This helps to better simulate longer dry spells.

Note

While the weather model runs on a daily basis, simple sub-daily disaggregation methods are included in the package. The timestep of the weather model output can therefore match that of the rainfall model.

Both average temperature and diurnal temperature range are simulated. Daily minimum and maximum temperatures can be derived from these variables (and written as outputs). Temperature is simulated first (after precipitation), with the other weather variables following.

The precise form of the regression equation used varies depending on the variable and transition state. The equations all include an autoregressive (lag-1) term and sometimes a term related to another variable. For example, when simulating average temperature, a term depending on precipitation is included in the regression equation if either the current or previous day are classified as wet.

Regression coefficients are identified using ordinary least squares.

An error/noise term adds the random component to the regression equations. This random component is simulated from a standard normal distribution and scaled according to the standard error of the regression equation.

Potential Evapotranspiration

Potential evapotranspiration (PET) is calculated using the FAO56 Penman-Monteith method. PET is calculated from the simulated temperature, vapour pressure, wind speed and sunshine duration time series.

Spatial Weather Model

The spatial weather model is very similar to its single site counterpart. The form of the regression equations is the same, but it is possible for some of the parameters to vary spatially. It is also possible to get weather model output for any location in the domain, even if the location does not correspond to an input weather station.

However, at present, only the standard error (used to scale the error/noise term) can vary spatially for a given application. This means that, currently, the model only uses uniform regression coefficients across the domain. This will be updated in future.

Future versions will also include the ability to simulate spatial fields for the error/noise term. Currently the model uses a single random number across the domain, although this number is scaled according to the standard error at each location.