The impact of climate change on the environment has become increasingly visible today with the rise of sea level and the increased severity of hurricanes and heatwaves. Scientists are predicting a significant rise in temperature over the next century, and a severe increase of climate change effects over time [1, 2]. According to the Fourth National Climate Assessment [3], the United States will face long-term impacts of climate change, including higher temperatures, longer frost-free seasons, increased rainfall rates and storm intensity. There is a crucial need for communities to enhance their preparedness and capacity to absorb disasters, tackling all these challenges. The availability of local-scale long-term projections of climate variables provides communities with better insights to monitor climate change and mitigate the subsequent impacts. Wind in particular is a crucial climate field to predict, playing a critical part during adverse events, and is a widely leveraged source of renewable energy.
General Circulation Models (GCMs) are numerical-based physical models that are able to simulate the physical processes and dynamics taking place in the atmosphere. Simulated data from GCMs can be obtained at the daily scale for future periods, usually up to year 2100. A GCM delivers important information and insights around the physical processes and dynamics governing our atmospheric climate system. GCM simulations are providing vital assessments of climate change impacts on various lifeline entities, including public health, critical infrastructure, as well as the ecosystem, among others. It is known, however, that these simulations fall short of adequately tackling the inter-disciplinary climate questions, particularly at the local regional scale. For instance, the spatial scales represented by the GCM may be too coarse with respect to what stakeholders require. Moreover, GCM simulations are considered to include biases relative to the data used in developing the model. It is deemed though that over time, these impediments of GCMs will be reduced with the advances in model formulation. However, it will be unlikely that such improved models will be able to address the different scales of interest.
Currently, the resolution of GCM simulations is limited to 100 km or more, due to computationally expensive processes employed within GCMs. However, regional climate can fluctuate significantly, even within one GCM grid cell. Local research organizations and practitioners are interested in analyzing such regional fluctuation, especially above certain areas, such as watersheds, small islands, etc. As such, the focus on downscaling climate projections grew significantly in the past decades, especially with the increasing visible impacts of climate change. Downscaled projections of various climate variables, such as wind speed, lead to a better analysis of local climate variability [4], and enable better local efforts for climate change mitigation as well as disaster risk management.
Downscaling of GCM outputs is performed using dynamical [5-7] or statistical approaches [8-10]. Dynamical downscaling approaches mimic the methodology of GCMs by extracting numeral equations to model the relationship between the climate variables, but at a finer scale. These approaches are computationally expensive and require expertise in the physical interactions between the climate variables. On the other hand, statistical downscaling approaches presume a statistical stationarity relationship between local observations (e.g., wind speed over a local station) and global large-scale GCM outputs (e.g., simulated wind speed at a coarse-scale) [11]. Moreover, statistical approaches assume that local small-scale climate patterns are mainly influenced by global large-scale climate patterns [12, 13]. Such approaches are mainly used for downscaling in the climate field, and often include machine learning-based regression techniques [14-16].
In this paper, we present a novel deep learning framework for statistical downscaling, specifically for forecasting the daily average wind speed at the station level using GCM simulations. Our framework, named Wind Convolutional Neural Networks with Transformers, or WCT for short, consists multi-head convolutional neural networks (CNNs) prefixed to stacked transformers, and an uncertainty quantification [17, 18] component based on the Monte-Carlo dropout sampling approach [19-21]. We apply WCT to forecast daily wind speed over four stations in New Jersey and Pennsylvania, United States.
In [22], we introduced the downscaling daily wind speed problem. Here, we make new contributions and extend the work in [22] along five directions, listed below.
- We provide more background information related to GCMs, and the use of machine learning for statistical downscaling.
- We adopt GCM simulations as input to WCT instead of the NCEP dataset (which is a proxy of GCM simulations) used in [22].
- We leverage CNNs with attention to explicitly and better perform spatial feature embedding. Specifically, we adopt multi-head CNNs with self-attention here instead of the AIG component used in [22] and a stacked transformer with two transformer models here instead of the single transformer with one transformer model used in [22].
- We expand the comparative study to include more benchmarked methods and analyze the time and space complexities of WCT and the related machine learning (ML) methods.
- We report future projections of wind speed up to year 2100, which are absent in [22].
We organize the rest of this paper as follows. Section 2 presents background information and reviews some of the ML-based approaches for statistical downscaling. Section 3 formulates the wind downscaling problem as a multivariate time series forecasting problem, and describes the datasets used in our study. Section 4 details the WCT framework. Section 5 reports experimental results. Section 6 concludes the paper and points out some directions for future research.