Self-adaptive differential evolutionary extreme learning machines for long-term solar radiation prediction with remotely-sensed MODIS satellite and Reanalysis atmospheric products in solar-rich cities

• Published in: Remote sensing of environment Vol. 212; pp. 176 - 198
• Main Authors: Ghimire, Sujan, Deo, Ravinesh C., Downs, Nathan J., Raj, Nawin
• Format: Journal Article
• Language: English
• Published: New York Elsevier Inc 01.06.2018; Elsevier BV
• Subjects: Algorithms; Ant colony optimization; Artificial intelligence; Artificial neural networks; Atmospheric models; Australia; BASIC (programming language); cities; climate; Diagnostic systems; Dipole moments; Dynamic stability; ECMWF; El Nino; Energy; Extreme learning machine; Feasibility studies; Genetic algorithm; Genetic algorithms; Genetic programming; Giovanni; Grid search; Hybridization; Indian Ocean; La Nina; Learning algorithms; Learning theory; Model testing; moderate resolution imaging spectroradiometer; MODIS; Neural network; Neural networks; Particle swarm optimization; Power plants; prediction; Prediction models; Radiation; Remote sensing; Renewable energy; Satellite solar prediction model; Satellites; Solar cycle; Solar energy; Solar radiation; Spectroradiometers; Support vector machine; Support vector machines; Swarm intelligence; system optimization; Weather forecasting; Indian Ocean; Australia; Satellite solar prediction model; Extreme learning machine; Genetic algorithm; ECMWF; Genetic programming; Grid search; Giovanni; Neural network; Support vector machine; Particle swarm optimization
• ISSN: 0034-4257; 1879-0704
• DOI: 10.1016/j.rse.2018.05.003

Designing predictive models of global solar radiation can be an effective renewable energy feasibility studies approach to resolve future problems associated with the supply, reliability and dynamical stability of consumable energy demands generated by solar-powered electrical plants. In this paper we design and present a new approach to predict the monthly mean daily solar radiation (GSR) by constructing an extreme learning machine (ELM) model integrated with the Moderate Resolution Imaging Spectroradiometer (MODIS)-based satellite and the European Center for Medium Range Weather Forecasting (ECMWF) Reanalysis data for solar rich cities: Brisbane and Townsville, Australia. A self-adaptive differential evolutionary ELM (i.e., SaDE-ELM) is proposed, utilizing a swarm-based ant colony optimization (ACO) feature selection to select the most important predictors for GSR, and the SaDE-ELM is then benchmarked with nine different data-driven models: a basic ELM, genetic programming (GP), online sequential ELM with fixed (OS-ELM) and varying (OSVARY-ELM) input sizes, and hybridized model including the particle swarm optimized-artificial neural network model (PSO-ANN), genetic algorithm optimized ANN (GA-ANN), PSO-support vector machine model (PSO-SVR), genetic algorithm optimized-SVR model (GA-SVR) and the SVR model optimized with grid search (GS-SVR). A comprehensive evaluation of the SaDE-ELM model is performed, considering key statistical metrics and diagnostic plots of measured and forecasted GSR. The results demonstrate excellent forecasting capability of the SaDE-ELM model in respect to the nine benchmark models. SaDE-ELM outperformed all comparative models for both tested study sites with a relative mean absolute and a root mean square error (RRMSE) of 2.6% and 2.3% (for Brisbane) and 0.8% and 0.7% (for Townsville), respectively. Majority of the forecasted errors are recorded in the lowest magnitude frequency band, to demonstrate the preciseness of the SaDE-ELM model. When tested for daily solar radiation forecasting using the ECMWF Reanalysis data for Brisbane study site, the performance resulted in an RRMSE ≈ 10.5%. Alternative models evaluated with the input data classified into El Niño, La Niña and the positive and negative phases of the Indian Ocean Dipole moment (considering the impacts of synoptic-scale climate phenomenon), confirms the superiority of the SaDE-ELM model (with RRMSE ≤ 13%). A seasonal analysis of all developed models depicts SaDE-ELM as the preferred tool over the basic ELM and the hybridized version of ANN, SVR and GP model. In accordance with the results obtained through MODIS satellite and ECMWF Reanalysis data products, this study ascertains that the proposed SaDE-ELM model applied with ACO feature selection, integrated with satellite-derived data is adoptable as a qualified tool for monthly and daily GSR predictions and long-term solar energy feasibility study especially in data sparse and regional sites where a satellite footprint can be identified. •SaDE-ELM is applied for solar radiation prediction for two cities in Australia.•Satellite (MODIS) and Reanalysis (ECMWF) data are used as model input.•SaDE-ELM is evaluated against ELM, hybrid-ANN, hybrid-SVR and GP.•Performance evaluation of SaDE-ELM outperforms the comparison models under study.•SaDE-ELM can accurately predict in an extreme climatic event (ENSO and IOD).