What is the difference between LARS-WG and KNN in generating weather data?

The K-nearest neighbor based weather generators is that they do not produce new values but merely reshuffle the historical data to generate realistic weather sequences. LARS-WG in which the sequence of dry and wet series of days is modelled first while the precipitation amounts and other variables are generated conditioned on the wet or dry status. Both WGEN and LARS-WG, however, have difficulty in reproducing the annual variability in monthly means of the variables. Further, they cannot simultaneously simulate weather data at multiple sites (Sharif and Burn, 2007). WGEN and LARS-WG are single-site models and therefore cannot simultaneously simulate weather data at multiple sites. Moreover, they require specification of model parameters and generally have difficulty in reproducing the annual variability in monthly means of the variables.

Rackso et al. (1991) used predefined distributions for modelling of wet and dry series whereas semi-empirical distributions are used in LARS-WG (Semenov and Barrow, 1997; Semenov et al., 1998). LARS-WG can be used for the simulation of weather data at a single site under both current and future climate conditions. These data are in the form of daily time series for climate variables, namely precipitation (mm), maximum and minimum temperature (C), and solar radiation (MJm-2day-1). Another advantage of LARS-WG is that the output of 15 GCMs with different scenarios has been incorporated into the model to deal better with the uncertainties of GCMs.

The most promising nonparametric technique for generating weather data is the K-nearest neighbour (k-NN) resampling approach. The works of Lall et al. (1996), Rajagopalan and Lall (1999) and Buishand and Brandsma (2001) describe various forms of k-NN resampling. A K-NN algorithm typically involves selecting a specified number of days similar in characteristics to the day of interest. One of these days is randomly resampled to represent the current day's weather. A k-nearest neighbor (k-NN) resampling scheme is presented that simulates daily weather variables, and consequently seasonal climate and spatial and temporal dependencies, at multiple stations in a given region.

Since LARS-WG uses every observation in the modelling process, it is expected to perform better than models that are based on fitting of a predefined distribution to the observed data. Semenov et al. (1998) confirmed the superiority of LARS-WG through a performance evaluation of WGEN and LARS-WG at 18 sites from different parts of the world. LARS-WG matched the observed data more closely than did WGEN, which may be attributed to the use of more complex distributions in LARS-WG. Both generators, however, had difficulty in reproducing the annual variability in monthly means of the variables.

Burn and Sharif (2005) reported the application of the basic K-NN model for the assessment of water resources risk and vulnerability to changing climatic conditions. Application of K-NN model for the generation of daily and hourly weather data has been presented by Sharif et al. (2007).