A new weather generator based on spectral properties of surface air temperatures
Introduction
The need for long time series of climate variables for agricultural and hydrological applications has resulted in the widespread use of stochastic weather generation models. These models provide data to augment the existing record at a site or, through interpolation of model parameters, provide climate information where measured data are not available (Johnson et al., 1996, Wilks and Wilby, 1999). Such models, referred to as weather generators, have several inter-connected components and usually simulate multiple variables including maximum and minimum daily surface air temperatures (Tmax and Tmin), solar radiation (R), and precipitation occurrence (Po) and amount (Pa).
The most widely used weather generator has been the autoregressive model introduced by Richardson (1981) and Richardson and Wright (1984), which is based on the multivariate autoregressive process described by Matalas (1967). While many studies have altered the original approach through changes in the way the parameters are computed (e.g., Schoof and Robeson, 2003), inclusion of additional variables (e.g., WXGEN, Nicks et al., 1990; GEM, Hanson and Johnson, 1998), or relaxation of normality constraints for the added variables (e.g., Parlange and Katz, 2000), the basic structure of the model has remained unchanged. As shown in Schoof and Robeson (2003), even with improvements to the model parameterizations, autoregressive weather generators still occasionally produce fundamental simulation errors, such as negative diurnal temperature range (DTR) (e.g., Tmin greater than Tmax). Additionally, Harmel et al. (2002) have indicated that monthly Tmax and Tmin probability distributions are generally skewed, and that generating temperatures with the normal distribution can lead to physically unlikely values. In this study, we present an alternative to autoregressive weather generators based on spectral methods. It is anticipated that by removing normality constraints and focusing on the relationship between Tmax and Tmin, some of the aforementioned problems may be alleviated.
The remainder of the paper is structured as follows. In Section 2, we briefly describe the data used for parameterizing and evaluating the weather generator as well as the study area in which the weather generator is applied. In Section 3, we describe the components of the weather generator. The weather generator is evaluated against observed data and compared to the autoregressive model in Section 4. Finally, in Section 5, we summarize our findings and discuss implications for current weather generator applications.
Section snippets
Study area and data
The weather generator described in this paper is motivated by current interdisciplinary research being conducted as part of the Southeast Climate Consortium, a group consisting of members from six universities in the Southeast USA focused on climate variability and risks to agriculture, forestry, and water resources in the region (see http://secc.coaps.fsu.edu). Within the southeastern USA, climate variability exhibits strong links to El Niño/Southern Oscillation (ENSO), which, in turn, impacts
Precipitation occurrence
The spectral weather generator adopts a two-state Markov chain model to generate a binary precipitation occurrence series (Buishand, 1978, Racsko et al., 1991). Although Wilks (1999) demonstrated that the first-order Markov process (with 1-day dependence) was generally appropriate for the eastern USA, we found that a second-order Markov process was more successful at simulating precipitation occurrence in our study area, especially for the two southernmost stations, which are located on the
Weather generator evaluation
We compare sequences generated by the model with observed data to evaluate model performance. We also compare our generated temperature data with that produced by an autoregressive (AR) model to place the differences in the context of weather generators. The AR model used in this study is described by:where the left hand side of the equation represents the current day's values, the vector with subscript i − 1 refers to the previous day's values, ɛi is a 2 × 1 vector of
Summary and conclusions
We have combined existing methods for generating daily precipitation occurrence with an innovative spectral approach to generating daily maximum and minimum air temperatures. The spectral weather generator was applied to data from nine stations in the southeast USA and the generated data was compared to both observed data and data generated by a variant of the commonly used AR weather generator.
Evaluation of the precipitation occurrence process revealed that the mean number of wet days was
Acknowledgements
COAPS receives its base support from the Applied Research Center, funded by NOAA Office of Global Programs awarded to Dr. James J.O’Brien. Additional support is provided by the USDA, CSREES and the USDA-Risk Management Agency through the Southeast Climate Consortium. We would also like to thank Dr. Daniel S. Wilks of Cornell University for providing Fortran code used for estimation of mixed-exponential parameters and Dr. Scott Robeson of Indiana University for comments on an earlier draft of
References (19)
Some remarks on the use of daily rainfall models
J. Hydrol.
(1978)- et al.
A serial approach to local stochastic weather models
Ecol. Model.
(1991) Interannual variability and extreme-value characteristics of several stochastic daily precipitation models
Agric. Forest Meteorol.
(1999)- et al.
A resampling procedure for generating conditioned daily weather sequences
Water Resour. Res.
(2004) - et al.
Application of Markov models to area-average daily precipitation series and interannual variability in seasonal totals
Climate Dyn.
(1993) - et al.
A quantitative evaluation of ENSO indices
J. Climate
(2003) - et al.
El Niño-Southern oscillation impacts on winter vegetable production in Florida
J. Climate
(1999) - et al.
ENSO influences on agriculture in the southeastern United States
J. Climate
(1998) - et al.
GEM (generation of weather elements for multiple applications): its application in areas of complex terrain
Cited by (21)
A new framework for multi-site stochastic rainfall generator based on empirical orthogonal function analysis and Hilbert-Huang transform
2019, Journal of HydrologyCitation Excerpt :For instance, the Wiener-Khinchin theorem states that the autocorrelation and the spectral density function of a stationary time series form a Fourier transformation pair. Actually Fourier analysis has been used in dealing with weather data series, for example, the Fourier spectra of observed monthly and annual time series were used to modulate the generated daily precipitation for preserving low-frequency variability (Chen et al., 2010; Schoof et al., 2005) developed a spectral weather generator to simulate the temperature based on its Fourier spectral properties. Paschalis et al. (2013) and Peleg et al. (2017) used Fast Fourier Transform (FFT) to model the areal precipitation in the frequency domain.
A conditional stochastic weather generator for seasonal to multi-decadal simulations
2018, Journal of HydrologyCitation Excerpt :Wilks (2008) illustrated the capability of interpolating weather generator parameters to arbitrary locations (e.g., on a grid) using local weighted regressions; Wilks (2009) subsequently offered a method to synchronize gridded synthetic weather series on observed weather data. Approaches to producing weather sequences that deviate from climatology have included the implementation of seasonal correction factors, perturbation of parameters or input data, and spectral approaches (Caron et al., 2008; Kilsby et al., 2007; Hansen and Mavromatis, 2001; Schoof et al., 2005; Qian et al., 2010). Nonparametric weather generators have an improved ability to capture nonlinearities between variables and sites.
Application of the multivariate spectral weather generator to the contiguous United States
2008, Agricultural and Forest MeteorologyCitation Excerpt :These results suggest that the second-order Markov chain coupled with the mixed-exponential distribution reproduces the basic characteristics of the precipitation climatology, but that interannual variability is underestimated on average. These results are in accord with other studies suggesting that additional low-frequency information (e.g., Johnson et al., 1996; Katz and Parlange, 1998; Schoof et al., 2005) or additional high-frequency information, such as conditioning on circulation (Katz and Parlange, 1993), within stochastic models may reduce overdispersion. Most WGs are able to replicate observed means and variances of individual variables.
Experimental drought threat forecast for Florida
2007, Agricultural and Forest MeteorologyEvaluation of multi-site precipitation generators across scales
2020, International Journal of Climatology