Predictive modelling of global solar radiation with artificial intelligence approaches using MODIS satellites and atmospheric reanalysis data for Australia

Predictive modelling of global solar radiation with artificial intelligence approaches using MODIS satellites and atmospheric reanalysis data for Australia

Global solar radiation (GSR) prediction is a prerequisite task for agricultural management and agronomic decisions, including photovoltaic (PV) power generation, biofuel exploration and several other bio-physical applications. Since short-term variabilities in the GSR incorporate stochastic and intermittent behaviours (such as periodic fluctuations, jumps and trends) due to the dynamicity of atmospheric variables, GSR predictions, as required for solar energy generation, is a challenging endeavour to satisfactorily predict the solar generated electricity in a PV system. Additionally, the solar radiation data, as required for solar energy monitoring purposes, are not available in all geographic locations due to the absence of meteorological stations and this is especially true for remote and regional solar powered sites. To surmount these challenges, the universally (and freely available) atmospheric gridded datasets (e.g., reanalysis and satellite variables) integrated into solar radiation predictive models to generate reliable GSR predictions can be considered as a viable medium for future solar energy exploration, utilisation and management. Hence, this doctoral thesis aims to design and evaluate novel Artificial Intelligence (AI; Machine Learning and Deep Learning) based predictive models for GSR predictions, using the European Centre for Medium Range Weather Forecasting (ECMWF) Interim-ERA reanalysis and Moderate Resolution Imaging Spectroradiometer (MODIS) Satellite variables enriched with ground-based weather station datasets for the prediction of both long-term (i.e., monthly averaged daily) as well as the short-term (i.e., daily and half-hourly) GSR. The focus of the study region is Queensland, the sunshine state, as well as a number of major solar cities in Australia where solar energy utilisation is actively being promoted by the Australian State and Federal Government agencies.

Firstly, the Artificial Neural Networks (ANN), a widely used Machine Learning model is implemented to predict daily GSR at five different cities in Australia using ECMWF Reanalysis fields obtained from the European Centre for Medium Range Weather Forecasting repository. Secondly, the Self-Adaptive Differential Evolutionary Extreme Learning Machine (i.e., SaDE-ELM) is also proposed for monthly averaged daily GSR prediction trained with ECMWF reanalysis and MODIS satellite data from the Moderate Resolution Imaging Spectroradiometer. Thirdly, a three-phase Support Vector Regression (SVR; Machine Learning) model is developed to predict monthly averaged daily GSR prediction where the MODIS data are used to train and evaluate the model and the Particle Swarm Algorithm (PSO) is used as an input selection algorithm. The PSO selected inputs are further transformed into wavelet subseries via non-decimated Discrete Wavelet Transform to unveil the embedded features leading to a hybrid PSO-W-SVR model, seen to outperform the comparative hybrid models. Fourthly, to improve the accuracy of conventional techniques adopted for GSR prediction, Deep Learning (DL) approach based on Deep Belief Network (DBN) and Deep Neural Network (DNN) algorithms are developed to predict the monthly averaged daily GSR prediction using MODIS-based dataset. Finally, the Convolutional Neural Network (CNN) integrated with a Long Short-Term Memory Network (LSTM) model is used to construct a hybrid CLSTM model which is tested to predict the half-hourly GSR values over multiple time-step horizons (i.e., 1-Day, 1-Week, 2-Week, and 1-Month periods). Here, several statistical, Machine Learning and Deep Learning models are adopted to benchmark the proposed DNN and CLSTM models against conventional models (ANN, SaDE-ELM, SVR, DBN).

In this doctoral research thesis, a Global Sensitivity Analysis method that attempts to utilise the Gaussian Emulation Machine (GEM-SA) algorithm is employed for a sensitivity analysis of the model predictors. Sensitivity analysis of selected predictors ascertains that the variables: aerosol, cloud, and water vapour parameters used as input parameters for GSR prediction play a significant role and the most important predictors are seen to vary with the geographic location of the tested study site. A suite of alternative models are also developed to evaluate the input datasets classified into El Niño, La Niña and the positive and negative phases of the Indian Ocean Dipole moment. This considers the impact of synoptic-scale climate phenomenon on long-term GSR predictions.

A seasonal analysis of models applied at the tested study sites showed that proposed predictive models are an ideal tool over several other comparative models used for GSR prediction. This study also ascertains that an Artificial Intelligence based predictive model integrated with ECMWF reanalysis and MODIS satellite data incorporating physical interactions of the GSR (and its variability) with the other important atmospheric variables can be considered to be an efficient method to predict GSR. In terms of their practical use, the models developed can be used to assist with solar energy modelling and monitoring in solar-rich sites that have diverse climatic conditions, to further support cleaner energy utilization.

The outcomes of this doctoral research program are expected to lead to new applications of Artificial Intelligence based predictive tools for GSR prediction, as these tools are able to capture the non-linear relationships between the predictor and the target variable (GSR). The Artificial Intelligence models can therefore assist climate adaptation and energy policymakers to devise new energy management devices not only for Australia but also globally, to enable optimal management of solar energy resources and promote renewable energy to combat current issues of climate change. Additionally, the proposed predictive models may also be applied to other renewable energy areas such as wind, drought, streamflow, flood and electricity demand for prediction.