Machine Learning-assisted Numerical Prediction of Dust Deposition in Solar Fields

Abstract

Solar energy is considered as the major contributor for United Arab Emirates towards achieving the National Energy Strategy goals by 2050. One quarter of the total energy will be obtained by solar energy by 2030. Meanwhile, levelized cost of energy (LCOE) for concentrated solar power (CSP) and photovoltaic (PV) solar plants has shown a 30% decrease since last decade due to the technological advancements. With wide deployment of large-scale solar power plants, maintenance of solar fields becomes a pressing issue. Conventional mirror and panel cleaning strategies for solar power plants are water-intensive and inducing high operational and maintenance costs (O&M). A better approach is required to identify the dust-concentrated solar collectors and panels for targeted cleaning. This thesis aims at developing a hybrid approach: two-phase computational fluid dynamics (CFD) simulation is integrated with machine learning technique to study the spatial dust prediction over solar field in CSP plants. The proposed hybrid approach involves the identification of meteorological parameters that affect reflectance loss, then generation of the required data using CFD technique and finally applying the clustering algorithm for prediction. As the benchmark case study, the solar field of SHAMS CSP power plant, located in the western region of Abu Dhabi is considered. A qualitative analysis of meteorological and reflectance parameters is performed using the SHAMS database. This study indicates that wind speed, wind direction, humidity, and cleaning pattern are the main factors that significantly affect the dust deposition rate and as a result reflectivity decreases. A 3D Euler Lagrange CFD approach is implemented and validated for dust deposition in SHAMS-1 CSP plant. Meteorological parameters (wind speed, wind direction and relative humidity) and geometrical parameters (plant size, plant shape, number and size of solar collectors) are taken from the actual solar field to build a CFD model, in which air is the primary phase and dust particle is the secondary phase deposited in solar field. Two extreme cleaning cycles of solar collectors with and without sandstorms are considered for validation. Representative spatial validation is performed by comparing the obtained reflectivity values with the measured database from SHAMS-1 plant. For both regular and irregular cleaning cycles (without/with sandstorm), the CFD model can identify and predict the spatial reflectivity of solar field, and applicable to similar solar power plants in desert environments. Spatiotemporal dust data is generated for different weather conditions for year 2021 by using the two-phase CFD approach, then a machine learning approach, particularly unsupervised clustering, is successfully utilized to capture the inherent relationship between the meteorological conditions and reflectance loss. Instead of time-consuming experiments, the proposed hybrid approach can successfully predict the dust-concentrated mirrors and solar panels for on-demand cleaning, which can significantly reduce the water usage and O&M costs.

Date of Award	Aug 2023
Original language	American English
Supervisor	TJ Zhang (Supervisor)