Chapter 2: Literature Review
Overview
Chapter 2 provides further motivation for the downscaling of Global Climate Models and investigates the literature surrounding statistical downscaling as well as introducing deep learning and its application in forecasting and downscaling of climate data.
Solar Energy Generation and Climate Change
Solar energy generation systems are susceptible to changes in long term climate. Schaeffer, et al. indicate that increases in air temperature influence the efficiency of Photovoltaic (PV) cell operation and that changes in ambient temperature influence the efficiency of Concentrated Solar Power (CSP) [1]. The study of climate change impacts for solar generation has received less attention than for other sources of renewable energy [2]. There are a small number of studies that do focus on solar availability under different climate warming scenarios. Such studies leverage the projections of multiple Global Climate Models (GCM) for different climate scenarios in order to reduce the uncertainty inherent in such projections [2]. Changes in global solar radiation over the next 100 years are projected to vary between regions with increases projected for South Eastern Europe, and decreases expected in areas such as Canada, sub-Saharan Africa and the Middle East [1]. Solaun and Cerda report that the outlook of solar availability for the Australian region is expected to be less variable than in other regions [2].
Crook, et al.’s, analysis of two CMIP3 GCMs (HadGEM1 and HadCM3) indicated little expected change in projections of PV generation in Australia between 2000 and 2100 [3]. However, the authors described regional differences where increases are likely in Europe, China and Algeria and decreases likely in the USA and Saudi Arabia [3]. Such assessments are subject to uncertainties in regional differences of model projections for temperature, cloud amount and absorbed solar radiation (ASR) [3]. Projections for cloud cover and all-sky radiation have the most influence over levels of uncertainty as indicated by Wild, et al. in an assessment of 39 CMIP5 GCMs [4]. An ensemble of projections for the RCP8.5 scenario indicated that changes in all-sky radiation governed regional patterns for projected PV output [4]. The authors also indicated slight increases in PV generation within the Australian region [4].
As GCM output occurs at a coarse geographic scale, it is desirable to model climate variability at a finer regional scale. A method of translation between the geographic scale and regional scale is necessary in order to undertake such an analysis.
Downscaling Global Climate Models
Statistical downscaling (SD) methods develop statistical relationships between high resolution regional observations and coarse scale GCM outputs for different scenarios and produce projections at the regional level [5] [6]. However, such methods rely on assumed covariance, variability and distributional assumptions that may not be true for all states of the climate system under simulation [7]. Dynamical downscaling couples GCM outputs with a Regional Circulation Model (RCM) which provides a physical model for local processes [6, 21]. However, dynamical downscaling requires extensive computational resources and does not generalise well across different regions [6] [7]. Uncertainties produced from GCM outputs impact both downscaling methodologies [5].
Wilby and Wigley, 1997 [8], provide a further categorisation of downscaling methods as follows. Regression methods establish linear or non-linear mappings between observational sub-grid scale predictand and GCM grid scale predictor variables [8]. Weather pattern generators provide a probabilistic classification scheme of weather states, linking observed station or area-averaged meteorological data with GCM outputs [8]. Stochastic weather generators leverage stochastic models such as Markov processes to capture conditional time dependence of GCM output parameters [8]. High resolution limited area climate models (LAM) are embedded within a GCM, these are dynamical models which are a restricted form of RCM and exhibit similar computational cost [8].
In order to address the systemic “cascade of uncertainty” affecting downscaling models Mitchel and Hulme, 1999, propose three methodological approaches in acknowledgement of such inherent uncertainty; (1) the use of multiple climate warming scenarios; (2) the use of model ensembles and (3) consideration of the entire response of the system [5]. The latter indicates that the modelled variables are subject to feedback over time and that the distribution of all possible climate states needs to be explored as an outcome of the downscaling activity [5]. Wilby and Wigley, 1997, also indicate that the most uncertain components of the climate models are related to water vapour and cloud feedback effects which are non-stationary processes [8].
This review will focus on studies of statistical downscaling for GCM outputs. Studies of downscaling global solar radiation are of particular interest, although the literature does not include a large body of work in this area, whereas there is a broad availability of downscaling studies focused on other climate variables such as for precipitation or temperature. In the latter part of the review, attention will be drawn to the use of deep learning methodologies and the advantages of such methodologies in downscaling GCM model outputs to the regional scale.
Statistical and Machine Learning Techniques for Downscaling
There are a broad variety of statistical methods that have been applied to the downscaling of climate variables. The relative strength of each approach is challenged to capture the variability of climate at the regional scale. Vandal, et al. demonstrate this in the use of a reanalysis product as a proxy for their target GCM model [6]. The study compares multiple statistical and machine learning methods, including Bias Corrected Spatial Disaggregation (BCSD) and notably a Convolutional Neural Network (CNN) [6]. The authors found that the relatively simpler BCSD model performed better than the other methods having been evaluated in terms of (1) the ability to capture daily anomalies; (2) the ability to respond to large scale climate trends; and (3) the ability to capture extreme precipitation events [6]. While the CNN architecture was found to perform poorly, the authors acknowledge that experimentation with other CNN architectures may yield different results [6]. It is also interesting to note that projections on the daily scale demonstrated lower correlations, as opposed to those of aggregate monthly averages [6].
Reanalysis products are often adopted as a historical data source during the training phase of downscaling studies, as they produce outputs compatible with coarse scale GCM models and are derived from a variety of sources for observational data [9] [10]. Koukidis and Berg, 2009, demonstrated that the choice of reanalysis product influences the model bias [10]. The authors calibrated a downscaling model for temperature and precipitation with two common reanalysis products, the National Centre for Environmental Prediction/National Centre for Atmospheric Research (NCEP/NCAR) reanalysis data and the European Centre for Medium Range Weather Forecasts (ECMWF) reanalysis data [10]. Precipitation and temperature estimates for the Southern Great Lakes region in the United States, revealed differences between the two reanalysis products, with the NCEP/NCAR model being more suited for downscaling precipitation and the ERA40 model more suited to temperature [10]. Indicating that the source of global scale predictors will influence model predictions.
The synthesis of multiple data sets (such as an ensemble of data sources) may be a useful approach that could assist in addressing these types of uncertainties or may simply become a necessity where certain types of predictors are only available as disparate sources. As a result, increases in the number of predictors poses further challenges for statistical modelling, requiring additional data driven approaches to extract the most useful set of predictors when addressing the downscaling task.
Antonanzas-Torres, et al. synthesise multiple data sets in the modelling of aerosol loading in the atmosphere (AOD) and the Linke Turbidity factor (TL) throughout the region of Spain [11]. The Monitoring Atmospheric Composition and Climate (MACC) reanalysis data set is combined with observations from the Aerosol Robotic Network (AERONET) and downscaled against 213 meteorological station observations [11]. Due to the large number of covariates, the authors address the problem of feature selection through the use of a cross validated genetic algorithm (GA) in order to train a scalar vector machine for regression (SVR) [11]. The method was demonstrated to contribute to the ability of the SVR to generalise over the 213 sites [11]. Such genetic algorithms assist in the selection of pertinent features for use as inputs to statistical modelling processes. In a 2016 study, Aybar-Ruiz, et al., apply a grouping genetic algorithm (GGA) to the training of an Extreme Learning Machine (ELM) regressor, so as to select features in the task of solar radiation downscaling multiple model grid points against the site of observation in the region of Toledo, Spain [12]. The wrapper function provided by the GGA procedure resulted in a model requiring only 9 of the original features, in opposition to the baseline model which required 92 of the original features [12]. The ability to select pertinent features as part of the learning task is a key challenge in the modelling process. The use of tools such as genetic algorithms indicate a trend in the use of more general data driven techniques to address the high dimensionality of available data.
Interpolation between observation sites is also a desirable outcome for statistical downscaling. However, there are uncertainties where observational data is not available as distances increase from the observation point. Davy, Huang and Troccoli (2016), experiment with the synthesis of irradiance measures produced by a numeric weather model (the Conformal-Cubic Atmospheric model, CCAM) and satellite irradiance estimates in downscaling against ground station observations across Australia [13]. A set of Generalized Additive Models (GAMs) are constructed individually per site and measures are converted to clear sky indices for global horizontal irradiance (GHI) and direct normal irradiance (DNI) [13]. Uncertainty of the resulting models were assessed with estimates of conditional variance subject to the predictors, and the evaluation of predictions within a small region around the point of observation. It was determined that bias increased with the distance from the observation point, hence discouraging the use of such methods where observational data is unavailable [13]. The improvements produced via synthesis of both model outputs and satellite measurements provides evidence that the combination of multiple sources of input improves the accuracy of the estimate against site observations [13]. The study is also unique in that it is one of the few studies applying downscaling methods to estimate operational variables for solar power within Australia and is focused on improvement of satellite estimate quality rather than exploring future scenarios resulting from changing climate.
The evaluation and comparison of the resulting models is an important concern for any statistical model development. In the statistical downscaling literature multiple methods of evaluation are used, although some common metrics are present between papers, unless explicitly cited, comparison between downscaling models is difficult between different research efforts. Vandal, et al. include the Root Mean Square Error (RMSE), Pearson correlation coefficient, mean model error (mean bias) and a skill score measuring the similarity between observation and predicted distributions of precipitation [6]. Antonanzas-Torres, et al. (2014), compare models with the Root Mean Absolute Error (RMAE), Mean Absolute Error (MAE) and the coefficient of determination (R^2) for the downscaling of TL and AOD [11]. Aybar-Ruiz, et al. measure the Pearson correlation coefficient, the Root Mean Square Error (RMSE), and a summary skill metric defined as the ratio between the RMSE of the prediction and the variance of observations for GSR [12]. Davy, Huang and Troccoli compare models with the Relative RMSE and mean absolute bias in DNI and GHI averaged over all sites [13].
Studies apply statistical downscaling using a reanalysis product rather than directly to the GCM model outputs in order to address the bias and uncertainty of the GCM outputs, however the reanalysis products are also shown to exhibit bias [25, 26]. Observations drawn from meteorological stations are primarily leveraged for the purposes of ground-truthing the statistical model. Downscaling is applied either on the spatial or time dimension or in some studies both dimensions [7]. Much of the focus in applying statistical downscaling is in the pre-processing stages where seasonal trends are either removed and or model outputs are debiased [9]. The statistical modelling techniques are then applied to the pre-processed data and mapped to finer grained observations for the target sites. Often multiple models are constructed and evaluated in search parameter space or feature space for the best performing model. Additionally, models are often constructed per site, rather than a single model capable of generalising across the local variability evident across multiple locations. Once the mapping is complete the models are assessed for performative metrics, and the distributional qualities of predictions are also assessed against the observational distributions in order to determine whether the models capture longer term states of the climate system. Some works recognise the time dependence of climate information when downscaling per site and capture this as embedded within the modelling framework. Deep learning methods are of interest due to their potential to learn a representation of a spatial and time varying signal without the additional stages of feature selection. It would be of interest to determine whether such a method is capable of learning a representation that is sufficient to generalise across geographically diverse regions.
Deep Learning
Deep learning is a form of representation learning that composes simple but non-linear models in multiple layers which transform the representation of features at each layer into a higher, more abstract level [14]. This provides an advantage over traditional statistical and machine learning methods which are challenged in the area of feature representation [14]. As evident in the work surveyed so far, much effort of statistical downscaling is dedicated to feature selection, and methods such as genetic algorithms or domain specific knowledge are leveraged in order to define inputs appropriate to the modelling stage. The ability of deep learning techniques to learn representations has been demonstrated in a number of diverse areas such as image classification, speech recognition and language translation [14]. Key architectures that have been leveraged in representation learning tasks include Convolutional Neural Networks (CNN), Auto-Encoders (for distributed representations), Recurrent Neural Networks (RNNs) and Long Short-Term Memory Networks (LSTM) [14]. The Convolutional Network leverages the key ideas of local connections, shared weights, pooling and the use of many layers in order to take advantage of the properties of natural signals [14]. Convolutional layers detect local combinations of features from the previous layers while pooling merges similar features into one [14]. RNNs are dynamic systems, where outputs of hidden units at discrete time steps are also included as inputs to units in subsequent iterations, having the purpose of learning long term dependencies between sequences of input [14]. RNN architectures proved difficult to train to store information and hence an approach to explicit memory was defined through the use of Long Short-Term memory (LSTM) networks which include a gated hidden unit in order to memorise inputs over time [14].
Such architectures as provided by deep learning are capable of automatically extracting spatio-temporal features [15]. Hence there is opportunity for their application in the climate modelling domain, where previous time steps and grid cells contain hidden information on the state of the climate system [15]. Reichstein, et al. summarise commonalities between climatological data, problem domains and applications of deep learning, noting that traditional applications of deep learning such as to images and video are analogous to two-dimensional data fields containing multiple variables and time series of earth systems variables [15]. It is the complexity of the spatial, temporal and non-linear dependencies, that are found within climate data that prompt the use of hand crafted features for more traditional statistical and machine learning methods, in comparison, deep learning methods are highlighted by the authors for their potential to automatically extract such relationships between features [15]. They note that the analysis of model-observation mismatch for physical models may be augmented by deep learning methods in order to downscale simulations to finer resolutions [15]. Methods performing such tasks borrow from Super-Resolution (SR) techniques applied in computer imaging and apply Convolutional modules as the feature extraction component in order to learn the mappings between coarse scale outputs of physical models and finer observational scales [15]. The stacking of these modules enables the network to learn a representation which progresses from low level features to increasingly abstract features in the later layers of the network [14]. It is this internal representation which enables such an effective means of non-linear relationships through the composition of the network architecture. However, there is no prescriptive method for determining just how deep the network architecture needs to be composed, leading to the necessity for an empirical search algorithm in determining the depth and configuration for the feature extraction modules within the network.
Forecasting Climate Variables with Deep Learning
The spatial feature extraction of the CNN module has been combined with time varying signals by leveraging an LSTM module for the purposes of near-term forecasting in the climate and weather domain. Shi, et al. design a Convolutional-LSTM (Conv-LSTM) network for the task of now-casting precipitation based on a series of radar images from the local region of Hong Kong [16]. The LSTM is modified so that the state to state transition functions are implemented by Convolutional modules [16]. In their architecture two Conv-LSTM layers are applied where the first module acts as an encoder network while the second receives state and output from the first network as its input and acts as the forecasting module (a form of decoder network) [16]. Optical flow modelling has traditionally been used in the precipitation forecasting task for radar imagery, however the authors indicate that it is difficult to estimate parameters required for good prediction performance, while the deep learning approach enables an end to end means of learning both the spatio-temporal relationships and state to state transitions within the learnt representation [16]. The Conv-LSTM is evaluated against a similar model, the fully connected LSTM (FC-LSTM) as well as the optical-flow model (ROVER). The authors indicate that the Conv-LSTM out performs the other methods since the model is able to handle boundary conditions well (such as a sudden cloud mass on the boundary of the radar) and that the non-linear structure of the network is capable of learning complex spatio-temporal patterns [16]. While the focus of the study is on precipitation, solar is impacted by cloud cover and such a method shows potential for modelling such non-linear systems over time.
Ghimire, et al. apply a hybrid Convolutional LSTM (CLSTM) model in the task of forecasting global solar radiation (GSR) against observations for 30 minute intervals made at Alice Springs in Australia [17]. The study provides a novel application of a CNN-LSTM architecture to forecasting sub-hourly GSR within the Australian region [17]. GSR data was collected from the Bureau of Meteorology (BOM) observations for Alice Springs between 2006 and 2018 and the size of the lagged series was determined by autocorrelation in order to define the input to the model [17]. Comparison was performed against a standalone CNN, a standalone LSTM, a dense network (DNN), gated recurrent unit network (GRU), a multi-layer perceptron (MLP) and decision tree (DP) [17]. Multiple evaluation metrics were used including the correlation coefficient, root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), absolute percentage bias (APB), Kling-Gupta efficiency (KGE), Nash-Sutcliffe efficiency, Wilmot’s index and Legate-McCabe’s index [17]. In addition multiple forecast horizons for 30 minute timesteps were considered including 1 day, 1 week, 2 weeks and n-months [17]. The authors indicate that the CNN-LSTM hybrid out performed other models and in the case of longer term forecasting (where there is increased uncertainty) the model performs within the category of good performance having less than 20% error [17]. In addition, the distribution of the predicted GSR for the CNN-LSTM model was accurate especially around the noon-period in daily forecasts where other models exhibited higher uncertainty [17]. The authors note that the performance over the standalone CNN is relatively small, however performance differences with the other models is larger [17]. While the data is a univariate time series the CNN-LSTM architecture performs well in this setting, however application to multiple sites in Australia is not explored in this work, leading to further opportunity in applying such a hybrid model both at multiple sites and in the task of downscaling global solar radiation.
Downscaling GCM Outputs with Deep Learning
While the forecasting task benefits from the feedback of an auto-regressive setting for prediction the signal, downscaling differs in that it must establish links between model outputs and the observed high resolution signal, which may not guarantee a strong connection with the observed signal. Hence the challenge of learning the feature mapping depends upon the ability of the network to incorporate what is often a large number of inputs into a suitable representation to achieve the mapping. Vandal, Kodra and Ganguly 2019, had demonstrated that a three layer CNN network had not learnt such a mapping as effectively as the simpler bias corrected spatial disaggregation method (BCSD) in downscaling precipitation [6]. In an earlier work Vandal et, al., 2018, adopt the SR single image context in order to downscale precipitation data using an architecture of stacked modules for Super Resolution Convolution (SRCNN), termed the DeepSD model [18]. Traditional statistical downscaling methods are challenged by the spatio-temporal non-stationarity of the climate system and the authors propose that the DeepSD model provides an ability to learn the non-linearities within the data and to perform a generalised sparse encoding of data within the downscaling task [18]. The feature extraction layers in the DeepSD model comprises of several stacked SRCNN modules, each SRCNN module consists of three CNNs which are designed to map between an LR input to a HR input [18]. DeepSD stacks three SRCNN modules to progressively increase the spatial resolution training each layer independently and combining or “stacking” the group at inference time [18]. Observations for precipitation and topographical data (derived from the PRISM data set and Global 30 Arc-Second Elevation data set) are leveraged to generate the training and test data at each of the required scales for the study area over the Continental United States (CONUS) [18]. Several statistical methods are compared against the DeepSD method and evaluation indicates that DeepSD is shown to outperform the baseline methods on the precipitation forecasting task as well as subsequent tasks for seasonal variation and prediction of extremes [18]. An ensemble of CMIP5 models are used to downscale precipitation between 1950 and 2005, however this is more of a comparison between methods in terms of computational cost and resource demands, with DeepSD outperforming other methods in terms of computational cost and resource demands (evaluated within a super-computing environment rather than domestic hardware) [18]. Unfortunately, the characteristics of regional climate for the future scenarios derived from the CMIP5 model ensemble are not explored further in the paper.
While the majority of network architectures discussed so far have incorporated CNN modules, there are situations where CNN modules are unable to learn from input data, especially in cases where data may be missing as described by Li, et al. [19]. In their work, an auto-encoder network was constructed to downscale weekly 50km grid scale MERRA-2 reanalysis variables to a 1km grid scale over the region of California between 2000 and 2016 [19]. This allowed data fusion with satellite observations derived from the Multi-angle Implementation of Atmospheric Correction (MAIAC) dataset at a 1km resolution [19]. A second auto-encoder network was constructed, from the downscaled dataset, to perform spatio-temporal imputation over the region of interest, and was evaluated on observations of Aerosol Optical Depth (AOD) derived from the AERONET (Aerosol Robotic Network) monitoring stations within the region [19]. An auto-encoder network is a symmetrically layered architecture, where the middle-hidden layer provides a non-linear representation of the latent components of the inputs [19]. The authors also leverage residual connections between the symmetric pairs of dense layers in order to speed up training and reduce error propagation [19]. The first series of layers are known as the encoder network, the middle layer represents the latent space, and the outgoing series of layers are known as the decoder network [19]. Such an architecture learns to reconstruct the input signal, which consisted of covariates from both the MERRA-2 and MAIAC data sets [19]. While the output of the network produces all variables, the AOD was the subject of the modelling. In order to learn the temporal dependencies, the data was arranged in a sliding window of three weeks, and the central week was used as the target for prediction [19]. Depending upon the temporal correlation of the predictors with the target variable, it may be possible to adapt such a model to ingest longer time horizons, however, one limitation of the dense architecture is that it is not able to receive more than one frame at a time, an advantage of the CNN module is the ability to represent time as the second dimension of the input frame, permitting a larger window for lags of the input variable.
Li, et al., demonstrate that the auto-encoder approach produces imputation and downscaling more accurately than the GAM method [19]. The model was shown to perform well in predicting seasonal changes and extreme events, where, for example, records of wild fire significantly increased AOD in the region and were reflected in predictions produced by the ensemble [19]. The correction of missing data is similar to the work by Davy, et al. which employed the GAM method to perform data fusion for satellite and numeric weather model outputs in reference to ground site observations [13]. A key difference in that work was the requirement to construct a separate GAM per site, as opposed to the capability of the deep learning approach to generalise across multiple sites.
Both of the downscaling architectures presented so far do not implement recurrent networks (unlike the examples in the forecasting task). Misra, et al. combine the dimension reduction capability of an autoencoder network with the recurrent architecture of an LSTM in the downscaling of coarse resolution reanalysis products for the regions of the Campbell Basin in Canada and the Mahanadi Basin in India [20]. Their model performs the task of weather-typing as well as regression for precipitation in each region [20]. The auto-encoder network is leveraged in order to reduce the dimensionality of the predictor data and is trained independently for a subset of data available in each region [20]. Both the output of the auto-encoder and a set of precipitation states (obtained via k-means clustering) were fed as inputs to the LSTM component in order to perform weather typing (classification of the next precipitation state) for the area of study [20]. This model was stacked with a second deep network, such that the activations of the auto-encoder hidden layer (latent features) as well as the predicted weather typing states, were provided as input to a separate RNN-LSTM network which produced regression estimates for the next precipitation value for all sites within the area of study [20]. Each deep network was trained independently and composed during inference [20]. Grid scales for predictor data in both the Campbell and Mahandi basins were 2.5°×2.5° and were downscaled to 0.1°×0.1° [20].
The authors baseline their approach against two statistical methods and demonstrate that the RNN-LSTM approach outperformed the other models in most cases for the prediction task except at some sites where the DNN model had slightly better metric scores [20]. An uncertainty analysis was undertaken using bootstrap estimation for the average mean squared error against observational data, and the authors indicate that the proposed model performed well enough to be reliable for use in downscaling GCM predictions [20]. However, the authors do not perform a projection of future predictions for any of the climate projection scenarios, which could include measures of uncertainty in evaluation of each scenario. The authors assert that the model captures the spatio-temporal dependencies satisfactorily but that there is a slight difference in performance between the two regions [20]. This may have been explained by the differences in the predictands and reanalysis data sets available for each region. Separate training phases are necessary for the auto-encoder to initially learn the representation prior to being leveraged to train the subsequent LSTM model in the downscaling task, this process replaces the manual feature engineering stage.
Summary
There appears to be a variety of deep learning architectures applied to the tasks of forecasting and downscaling of climate data, not unlike the diverse range of statistical methods leveraged for the same tasks. CNN modules are commonly leveraged for the purposes of extracting spatial dependencies. The alternate approach has been to leverage an auto-encoder network to extract latent features of the input data, effectively producing a form of spatial feature reduction leveraging the non-linear representation of the network. Approaches to extracting temporal dependencies differ. Some methods expand short time horizons by leveraging multiple frames of the input series and feed those into a CNN or dense architecture, whereas other methods discussed leverage a recurrent module such as an LSTM in order to learn the long-term dependencies in the features. It is clear that deep learning provides the opportunity to assemble a larger model out of separate modules, where each layer performs a distinct representation learning task. However, choices of which components to apply appear to be based on a combination of heuristics and prior empirical evaluation against baseline models. There does not appear to be a recommendation as to how to best combine each respective module in an end to end architecture.
An additional challenge is the data format for source and target data, these are either a time series of 1-dimensional vectors, or alternately in the case of SR, time series of 2-dimensional matrices and necessarily inform the architecture of the network, at least in terms of the boundary layers.
GCM output is not commonly applied to the training of downscaling models, largely due to the uncertainty of such models, whereas, reanalysis data products are commonly leveraged as the data source during the training and evaluation process. The task of fusing data sources such as satellite data with other data products also poses a challenge, and the analysis of error present in such products is an entire field of study that has a relationship to downscaling for bias adjustment or error correction.
Evaluation methods that are employed most commonly include measures of correlation, the coefficient of determination and the mean squared error, and most studies do adopt multiple evaluation metrics. Distributional qualities of observations are also compared to model predictions in order to determine whether the variation of climate variables are captured in the model. However, only a few studies extrapolate and explore future climate scenarios after demonstrating the efficacy of the model.
In addition, studies concerned with downscaling solar variables are less common than the literature available on downscaling precipitation. Some variables impacting global solar radiation are studied, such as aerosol optical depth and clear sky indices. Studies of future climate scenarios for solar resources incorporating downscaling within the Australian region are less common than in other regions, while there are a number of studies exploring future projections of GCM models at a continental scale.
As such there are a number of opportunities presented for further work in order to contribute to the growing research in downscaling climate models.
Firstly, there is opportunity to examine the downscaling of solar variables from transformed GCM model outputs, such as values for GHI, DHI or GSR, since there is less availability of research focused on projected changes of solar resources for renewable energy in the coming century, especially in the Australian region. While coarse grained outputs indicate increases in the Australian region, there are bound to be regional differences. Given the investment in solar renewable energy projects from the Queensland Government, and the commitments to achieve 50% renewable energy targets within region, it will be of value to investigate multiple climate scenarios for potential changes in solar availability at a number of key sites in the Queensland region over the longer term.
Secondly, the use of GCM outputs, directly for the purposes of downscaling, is not common due to the uncertainty inherent in such coarse scale models, whereas reanalysis data products are preferred. Such data sources may be derived from a fusion of satellite data sources with a single dynamic model and have been shown to also exhibit a level of bias. There is opportunity to leverage an ensemble of GCM model outputs in the downscaling task. While an ensemble helps to address bias between models, there is still uncertainty inherent in the use of such models, which will impact the downscaling method. It will also be of interest to explore methods of quantifying such uncertainty resultant in the downscaling approach, as demonstrated in comparisons of distributions for projected values against observations or estimating conditional variance of the model against observations.
Thirdly, the standard statistical downscaling methods are reliant on distributional assumptions and leverage additional stages to perform feature engineering such as genetic algorithms. Methods such as super-resolution downscaling leverage Convolutional modules in order to map between low resolution and high resolution images. While such methods have been applied in forecasting and downscaling for precipitation, it is unclear as to whether the analogy operates in the setting of downscaling radiation. Of particular interest are the use of CNN modules to learn a spatio-temporal representation, within a variety of network architecture configurations presented so far. Fourth, the combination of CNN and LSTM modules has been demonstrated to perform well in the forecasting task, but neither modules appear to have been combined in the downscaling task. Opportunity exists to examine the performance of such a CNN-LSTM architecture in the downscaling of GCM ensembles against observational data for both 1-dimensional and 2-dimensional time series configurations. Given that most studies have compared proposed network architectures against traditional statistical models, there is opportunity for a comparative study for a series of network architectures in order to determine recommendations for the combination of modules in the GCM downscaling pipeline.
References
Downscaling Global Climate Models with Convolutional and Long-Short-Term Memory Networks for Solar Energy Applications by C.P. Davey is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.