Abstract
Solar energy generation, as part of the mix of energy generation sources, forms part of a national strategy for energy security in the current face of a changing climate. The projection of longer-term global solar radiation availability and its natural variability at local scales require the downscaling of the coarse resolution Global Climate Models (GCM) simulations to a finer scale resolution. Statistical downscaling methods, commonly applied to achieve such a localised mapping, require extensive feature engineering and do not generalise to the ground station very well geographically, often requiring an individual, site-specific model to be developed per each site. Deep learning methods have been shown to be effective in learning the non-linear relationships between GCM simulations and ground observations, and as such have been applied to the climate downscaling tasks, albeit, these have focused on mostly temperature and precipitation rather than on global solar radiation datasets.
This Master of Science thesis adopts an ensemble of three Coupled Model Intercomparison Phase CMIP5 GCM datasets (i.e., ACCESS1-0, HadGEM2-CC and MRI-CGCM3) in order to develop and apply the integrated Convolutional Neural Network and Long Short-Term Memory Network (CNN-LSTM) deep learning models to downscale monthly average global solar radiation at seventeen key solar energy sites in Queensland, Australia. Downscaling is performed both for a local modelling task, i.e., simulating a single grid point (defined by longitude, latitude) to construct a model denoted as CNN-LSTML and for a universal modelling task mapping a 3 grid point set (or a grid size 1.25°×1.875°) for each of the GCM variable and simulating a 9 set of high resolution grid point set for each site of interest (i.e., grid size 0.05°×0.05°) to construct another model denoted CNN-LSTMU.
The model’s architecture and network parameters of the CNN-LSTM models are determined by a search procedure to select the optimal configuration of the resulting network. Other than the objective model, three baseline models are also developed consisting of the Generalised Linear Model (GLM), Elastic-Net (GLMNET) and a Random Forest (RF) model. These baseline models are adopted to appraise the newly developed CNN-LSTM models against the performance of traditional methods applied in global solar radiation downscaling. Model evaluation metrics include the Coefficient of Determination (\(R^2\)), Kling-Gupta Efficiency (\(KGE\)), Willmott’s Index of Agreement (\(d\)), Nash-Sutcliffe Efficiency (\(NSE\)), Root Mean Square Error (\(RMSE\)), Mean Absolute Error (\(MAE\)) and the Relative Root Mean Squared Error (RRMSE), computed within the testing phase. The final results demonstrate that the newly developed CNN-LSTM models are able to exhibit good capability to generalise well across multiple study sites. However, there still remains some differences between historical (i.e., observed and simulated) global solar radiation, illustrating a plausible degree of local dependency between the study sites in both downscaling tasks.
The local modelling task, performed using the CNN-LSTM\(_L\) algorithm, seem to generate the best performance at the Springs site with a KGE value of 0.91, 0.84 (NSE), 0.85 (\(R^2\)), and a value of d = 0.96, RMSE = \(17.78Wm^{-2}\), \(MAE = 14.05Wm^{-2}\) and an \(RRMSE = 7.75%\), compared to all other study sites. In spite of its good performance, the universal model, denoted as CNN-LSTM\(_U\), appears to face greater challenges in terms an acceptable value of KGE, although this method does achieve a good score for the other performance metrics. The best performing site was Mount Larcom Post Office, with a value of 0.81 obtained for both the KGE and NSE, \(d = 0.94\), \(R^2 = 0.83\), \(RMSE = 19.43Wm^{-2}\), \(MAE = 18.86Wm^{-2}\) and an \(RRMSE\) value of \(8.43%\).
In terms of modelled errors, the uncertainty was reflected by the 95% Mean Prediction Interval (MPI) of the CNN-LSTM models assessed for the historical period (1989 – 2005) and also the two representative concentration pathways, RCP4.5 and RCP8.5 simulated from 2006 to 2020. The objective model (i.e., CNN-LSTM) demonstrated a lower uncertainty at the 95% prediction interval compared to the best performing baseline model (i.e., the GLM). For the case of the local simulations obtained by the proposed CNN-LSTM\(_L\) model, the 95% MPI, produced on a test set, was \(\pm101.69Wm^{-2}\) in comparison to \(\pm106.51Wm^{-2}\) produced for the baseline model. The CNN-LSTM\(_U\) model tested for the universal case, where the proposed model was developed to map a \(3\ \times3\) grid point set for each GCM variable in order to simulate global solar radiation in a \(9\ \times 9\) high resolution grid points set around the site of interest, produced a 95% MPI of \(\pm114.72Wm^{-2}\) on the test set. This contrasted a value for comparison models which generated a 95% MPI of \(\pm124.75Wm^{-2}\).
Accordingly, these results, demonstrate the effectiveness of the proposed deep learning approaches in automatically synthesizing large numbers of predictor variables to produce a sound mapping and downscaling of global solar radiation at the finer grid resolutions.
As a final modelling task, the CNN-LSTM models for both the local and universal modelling cases, are applied to project global solar radiation under two future climate warming scenarios, identifying the potential changes in solar energy availability up to 2100. In the results the average monthly projection for RCP8.5 scenario is shown to be marginally lower than the RCP4.5 scenario till the decade ending in 2090, and slightly higher than the RCP4.5 between 2090 and 2100.
These model projections are good estimates of the impact of a changing climate on solar energy availability within Queensland. The present projected results appear to have a marginal difference between the two warming scenarios at all study sites but the case for RCP4.5 warming scenario shows a slightly higher availability of the global solar radiation for a majority of the future period in comparison to the RCP8.5 scenario. The study ascertains that Deep learning methodologies can be adopted to provide resource estimations and uncertainty evaluations for renewable energy, mapping and modelling energy at local sites (e.g., solar and wind farms), mapping the relationship between numerical weather models and ground station observations and downscaling the atmospheric variables for other research and development.
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.