A high‐resolution bilevel skew‐tstochastic generator for assessing Saudi Arabia's wind energy resources

Abstract

Saudi Arabia has recently established its renewable energy targets as part of its "Vision 2030" proposal, which represents a roadmap for reducing the country's dependence on oil over the next decade. This study provides a foundational assessment of the wind resource in Saudi Arabia that serves as a guide for the development of the outlined wind energy component. The assessment is based on a new high-resolution weather simulation of the region generated with the Weather Research and Forecasting (WRF) model. Furthermore, we propose a spatiotemporal stochastic generator of daily wind speeds that assists in characterizing the uncertainty of the energy estimates. The stochastic generator considers a vector autoregressive structure in time, with innovations from a novel biresolution model based on a skew-t distribution with a low-dimensional latent structure. Estimation of the spatial model parameters is performed using a Monte Carlo expectation-maximization (EM) algorithm, which achieves inference over approximately 184 million points and enables to capture the spatial patterns of the higher order moments that typically characterize high-resolution wind fields. Our results identify regions along the western mountain ranges and central escarpments that are suitable for the deployment of wind energy infrastructure. According to the assessment, between 30 and 70% of the national electricity demand could be met by wind energy.

Discussion on A high‐resolution bilevel skew‐tstochastic generator for assessing Saudi Arabia's wind energy resources

“…Wind speed, when treated as a stochastic process, is likely to be non-separable in space and time, and the stencil neighbourhood scheme used in Tagle et al (2020) does allow for nonseparability. The model also allows for spatially varying dynamics, which is important since dynamics do tend to vary spatially.…”

“…The approach taken is pragmatic, but in other applications one might need to take a different stance when weighing the use of a pure non-Gaussian model against that of a latent Gaussian model. Figure 2a in Tagle et al (2020) depicts only mild departures from Gaussianity, so a latent Gaussian model might indeed be appropriate here. Cseke, Zammit-Mangion, Heskes, and Sanguinetti (2016) fitted a spatiotemporal log-Gaussian Cox process model where the spatial dimension was approximately 10,000, and where the data spanned a few hundred time points.…”

“…To obtain gradients with respect to , , and , these parameters need to be explicitly represented in the expression for . I therefore now draw on a connection between the distributions used by Tagle et al (2020) and the reparameterization trick of Kingma and Welling (2014) by expressing the Monte Carlo samples via the parameterization of Azzalini and Capitanio (2013, Chapter 2),…”

“…This discussion paper is broadly divided into two parts. In the first part I discuss some aspects of the work by Tagle et al (2020) at a conceptual level; this part includes a brief discussion on the implications of treating numerical model output as "data," certain features of the spatio-temporal model used, and the authors' handling of non-Gaussianity. In the second part I propose some alternative computational strategies that could be used to facilitate inference.…”

“…In the second part I propose some alternative computational strategies that could be used to facilitate inference. Tagle et al (2020) use high-resolution data from the Weather Research and Forecasting (WRF) model to construct a surrogate statistical model that can, in effect, replace WRF over the region and time horizon of interest. Such a strategy of replacing the deterministic model with a statistical one serves several purposes.…”