The Role of Computational Science in Wind and Solar Energy: A Critical Review

Abstract

This paper concerns technology challenges for the wind and solar sectors and the role of computational science in addressing the above. Wind energy challenges include understanding the atmospheric flow physics, complex wakes and their interaction with wind turbines, aeroelastic effects and the associated impact on materials, and optimisation of wind farms. Concentrated solar power technologies require an optimal configuration of solar dish technology and porous absorber in the volumetric solar receiver for efficiency and durability and to minimise the convective heat losses in the receiver. Computational fluid dynamics and heat transfer have advanced in terms of numerical methods and physics-based models and their implementation in high-performance computing facilities. Despite this progress, computational science requires further advancement to address the technological challenges of designing complex systems accurately and efficiently, as well as forecasting the system’s performance. Machine Learning models and optimisation techniques can maximise the performance of simulations and quantify uncertainties in the wind and solar energy technologies. However, in a similar vein, these methods require further development to reduce their computational uncertainties. The need to address the global energy challenges requires further investment in developing and validating computational science methods and physics-based models for accurate and numerically efficient predictions at different scales.

1. Introduction

It is estimated that by 2050 the world population will approach 9.8 billion, and the global electricity demand will exceed 38,000 terawatt-hours annually [1]. Moreover, governments globally push for zero-carbon technologies to address the impact on climate and public health. However, the above trends and aspiring targets pose significant technological and economic challenges.

Renewable energy can provide a percentage of the global energy resource required. This includes wind, solar and hydropower. The present study does not cover hydropower. However, we note that hydropower is a vital energy source and has always been considered a means of political power in both western [2] and eastern economies [3,4]. A review of computational science and its contribution to the advancement of hydropower deserves a separate review study.

Complete reliance on renewable energy sources would be practically complex, if not impossible, but the discussion on this subject is beyond the scope of this paper. Nevertheless, wind and solar energy will increasingly play a significant role in meeting energy requirements. These energy resources can help countries achieve a more sustainable environment. There is still room for advancing these technologies, but this needs to be accelerated. It is worth mentioning that renewables have received a significant boost from European Union funds, national funds and foreign direct investments. This is particularly important in areas where social risk and energy poverty, including former mining areas and other fossil fuels, are at stake [5]. Renewable energies and AI can potentially contribute to reducing the poverty risk in some regions. Furthermore, renewable energies could be a source of alternative development in developing countries, with significant socio-economic consequences [6].

Harvesting wind energy is a multi-physics and multi-disciplinary engineering problem. It depends on the atmospheric flow, particularly around the wind turbine’s or farm’s operation, the interaction of the flow with several wind turbines and their surrounding environment, particularly complex wake flows. Moreover, aeroelasticity has become a significant challenge, particularly for large offshore wind turbines. Thus, materials, vibration, control and system dynamics are critical issues—and, finally, wind farm optimisation for energy harvesting under different wind and terrain conditions.

Solar energy is one of the safest and cleanest renewable energy resources [7]. Harnessing solar power (solar radiation) can be achieved through three major technologies: photovoltaic electrical (PVE), photovoltaic thermal (PVT) and solar thermal collectors (STC) technologies [8]. The applications of these technologies can vary between electricity generation and direct water/air heating or steam production at different scales, e.g., domestic, industrial, and urban. The various applications and, more importantly, the intermittent behaviour of solar radiation between day and night or weather seasonality usually require different storage systems to ensure the continuous solar energy production for electric generation or heating. For example, in PVE technologies, electric power storage and management of hybrid systems are needed, such as advanced life-long-standing batteries and automatic control and integration into multi-scale electricity grids. Enhancing the overall performance and efficiency of PVE panels also requires research to advance the cooling technologies applied in these systems. In PVT technologies, such as water heating by solar collectors, thermal storage systems are needed, such as innovative reservoirs with integrated intelligent materials, e.g., phase change materials (PCM) [9].

Solar radiation absorbing materials (SRAM), thermal storage (TS), and electric power storage (EPS) systems have been the subject of continuous research and development during the last decade. Computational science methods, such as computational fluid dynamics (CFD) and artificial intelligence (AI), were employed at different scales to improve the overall efficiency of the absorbing materials and the storage systems. The objective is to optimise the designs of solar collectors and PV cells to reduce the electric-power losses and maximise the materials’ absorptive capacity, thus maximising the overall performance of the solar energy technology employed.

CFD plays a crucial role in designing wind turbines and wind farms. However, despite the progress in CFD, many areas have yet to advance significantly. The cost of numerical simulations remains high, and turbulence modelling has marginally progressed in the last three decades. Performing high-resolution direct numerical simulations relies on expensive multi-processor clusters containing thousands of processors. These high-performance computing systems cannot be purchased by small and medium size companies and are primarily maintained by a few universities and nationally funded research centres. Therefore, fine-grain simulations, such as direct numerical simulation (DNS), remain an advanced academic tool for understanding the complex turbulent flow physics, but not for using it in the design phase of a wind turbine or a wind farm [10,11]. Large Eddy Simulation (LES) and its variants, including detached eddy simulation (DES), implicit large eddy simulation (ILES), and hybrid LES-Reynolds-Averaged-Navier-Stokes (LES-RANS), provide an alternative compromise between accuracy and computational cost. However, the computational cost could still be high for computing the flow or the aeroelastic effects around a wind turbine and for modelling the atmospheric flow around a wind farm at a reasonable micro-scale resolution. Moreover, determining noise effects from wind turbines poses even more stringent computational restrictions, particularly in the near field around the turbine.

Computational modelling of innovative solar energy conversion systems can be classified into three major scales: one-dimensional or reduced order models applied at the scale of a solar power plant for operation performance optimisation and monitoring, physics-based and CFD models used at the scale of local components for the optimal design and conjugated heat and mass transfer analysis combined with physical models and ML techniques applied at intermediate and large scales for solar radiation forecasting under unsteady weather and environmental conditions.

CFD can be used to investigate the conjugated heat transfer in PVT and STC hybrid systems. Most importantly, researchers perform parametrically shaped optimisations of new-generation solar plates and collectors for efficient conjugated heat transfer and fluid flow. Three-dimensional computations and analysis allow the investigation of local phenomena of radiation, conduction and convection in the collectors, taking into account the dynamic effects of boundary and operating conditions and the thermophysical properties of the different operating fluids and materials. CFD enables the creation of new designs of plates that distribute the local fluid flow to better homogenise the local temperature field and avoid maximum peaks of the latter that can damage the components. Moreover, CFD reduces the cost of finding optimal operating conditions.

Research regarding Machine Learning (ML), the pathway to Artificial Intelligence, grows rapidly [12]. In this paper, ML incorporates all the variants of deep learning and neural network (NN) methods. ML methods could be trained to predict wind turbines and wind farms’ performance and assist in designing more advanced wind energy systems, subject to data availability. Computational science such as CFD and aeroelastic methods can help generate fine-grain data, which is difficult (and expensive) to obtain by deploying several sensors on a wind turbine and across a wind farm. Furthermore, CFD can be combined with sensors to assimilate field data and improve the CFD simulations’ accuracy, and thus, the ML models’ training accuracy. Accurately trained ML models could accelerate the design of wind turbines and the installation of wind farms. However, ML is not free of predictability uncertainties. Thus, there are computational challenges as much as there are technological challenges.

ML methods have been rapidly developing in the last few years for solar radiation emission estimation [13]. These include ARIMA (Auto Regressive Integrated Moving Average), NN, SVM (Support Vector Machines), SVR (Support Vector Regression), k-mean, boosting method, regression tree and random forest methods. Of course, no single ML method is a panacea. Still, NN methods are generally the most widely used because they currently offer nonlinear approximations and are more flexible than the other ML methods. Generally speaking, the overall accuracy of ML models depends on the training data quality fed into the model.

Optimisation combined with CFD is a multi-disciplinary field of computational science and engineering. Shape and topology optimisation incorporating adjoint methods [14,15,16,17] (gradient-based optimisation algorithms) are up-and-coming computational tools that can be applied to find optimal designs of wind turbines and solar collectors. These advanced multi-physics-based optimisation methods within CFD allow the creation of multiple engineering components in the wind and solar energy sectors. For example, they can enable the optimal design of future aerodynamic blades and the embedded electrical motor in a wind turbine for optimal performance [18]. They can also provide the optimal distribution of semiconductor materials used in PV cells and the optimal design of solar collectors [19]. The adjoint gradient-based constrained optimisation techniques are less computationally expensive than gradient-free optimisation techniques [20].

Thanks to HPC and data centres, multi-objective CFD optimisation can be applied with nonlinear inequality constraints to solve complex problems such as the dynamic cooling of wind turbine motors and PV plates; the aerodynamic drag and noise reduction of wind turbine blades; the optimisation of the radiation absorption capacity of PV cells; and the reduction in the overall cost, volume and weight of the different components employed in wind and solar energy conversion systems.

Different challenges in the wind and solar renewable energy technologies arise due to the growing number of applications (Figure 1). We discuss the challenges and opportunities regarding these technologies and the role of computational science in the following sections.

2. Challenges in Wind Energy

Governments, consultancy organisations and the International Energy Agency (IEA) anticipate that wind and solar will be up to two-thirds of the total electricity demand, with wind share being 25% to 30% globally [21,22,23,24,25]. To achieve these very ambitious goals, we need to make wind energy economically viable and reliable for consumers and investors [26,27].

We know that wind power varies with the cube of the air speed and the rotor’s swept area multiplied by a power coefficient, which is a measure of the aerodynamic and electro-mechanical performance. However, in practice, we know that the energy captured scales with the area of the rotor.

2.1. Technology Challenges

Wind turbine technology has progressed via optimised wind turbine blades, larger wind turbines and higher hubs. Wind turbine blades are quite advanced [28,29]. Nowadays, we can achieve higher tip speeds, reduce torque and drive-train weight, achieve higher lift using more slender and lighter blades, and reduce aerodynamic noise through optimally designed tip shapes. Furthermore, aeroelastic optimisation and associated materials have led to an aerodynamic-loading reduction (coupling of bending and twist) and improvement of aerodynamic performance in the proximity of the load-bearing near the hub. Several other enhancements emerged from the manufacturing [30,31]. Major challenges concerning wind turbines are illustrated in Figure 2.

Wind technology experts via the IEA Wind Technology Collaboration Programme have defined the critical innovation areas required for meeting the aspiration targets set by governments [1]. These comprise:

• An improved understanding of flow physics at an atmospheric and wind farms scale.
• Improvements regarding the aerodynamics of onshore and offshore turbines and hydrodynamics of offshore turbines. Furthermore, a better understanding of large wind turbines’ structural dynamics (aeroelastic effects in particular).
• Improved integration of wind power plants into the electricity grid.

2.2. Flow Physics

Flow physics around a wind turbine and a wind plant is a problem of physical scales. Mesoscale processes are usually considered above 5 km up to hundreds of kilometres in size. Microscale processes are below 1 km. Wind turbines are currently placed below 300 m within the atmospheric boundary layer. Within this region, the turbulence will be caused by the landscape’s features, such as artificial structures, trees, mountains and hills, and uneven terrain. The area from 1.5 to 0.5 km lies approximately in the interface of mesoscale and microscale processes. It is also known as “terra incognita” [32], as the scales and atmospheric processes are intertwined, and we have even less understanding of the flow physics in this region [33,34].

Wind turbine technology reaches rotor sizes up to 200 m at the tip. However, the need for energy harvesting at a lower cost points to bigger wind turbines. Therefore, mesoscale phenomena that contain many unknown flow physics parameters become more relevant as we aim at larger wind turbines. The lack of understanding of the flow physics at the interface of microscale and mesoscale regions poses uncertainties for the wind farms’ design. The design will depend on the details of the terrain and the transitional and turbulent flow characteristics onshore and offshore. We use simplified predictions that allow for predictions over short times, but conditions can change over a more extended period, thus making predictability uncertain or impossible [35,36]. This challenge is similar to trying to forecast the weather for several months. Offshore wind adds to the complexity, as here we have the interaction of the wind with random waves. Modelling free surface effects under the interaction of highly turbulent winds is a remarkably complex problem.

Wake flows influence the onshore and offshore wind farms’ performance. The wake’s behaviour depends on the inflow and boundary conditions, the terrain, and the turbine size [11,34,37]. Several methods are used to understand flow wake physics, including CFD, laser and radar measurements [38,39,40,41]. The challenge, however, lies in the variability of the flow phenomena depending on the local atmospheric flow conditions, turbine types and wind farm design.

2.3. Aeroelastic Behaviour

For smaller wind turbines, years of computational and experimental research have shed light on aerodynamic and aeroelastic behaviour. However, larger wind turbines push the boundaries of the design because the aerodynamic loads are higher, and turbulence plays an even more dominant role. In addition, wind turbines’ flow wake occupies larger spaces, and the interaction of these phenomena imposes significant forces on the turbine’s structure. The aeroelasticity phenomena for smaller turbines were less potent than for large wind turbines. The average hub height for offshore turbines can be up to 150 m, with a mean rotor diameter of 127.5 m. Interestingly, MingYang Smart Energy, a Chinese wind turbine manufacturer, has designed an offshore wind turbine with a diameter of 242 m, 118 m of blades and 46,000 m² swept area.

The above dimensions push the limits of aeroelastic response and materials. They require a significantly more advanced understanding of turbulence, fluid-structure interactions and materials properties under extreme conditions over the years, considering that these turbines will not be easy to repair and maintain, particularly in an offshore environment.

For offshore structures, the challenges are even more significant due to the interaction of the rotor with its vorticity, especially in offshore structures [42,43]. In addition, the aerodynamic is intertwined with hydrodynamic, requiring careful stability analysis of the aero-hydro-servo elastic systems [44,45]. Theoretical approaches are insufficient to understand the coupled problem because of their assumptions. Wind tunnel experiments on such large scales are also challenging despite efforts to understand the interaction of turbulence around large blades [46,47].

2.4. Wind Farms

The overall aim of wind energy technology is to supply reliable and stable energy to the grid. Therefore, we must optimise and maximise the wind farms’ input to the overall grid [10]. Further research is required to integrate the system’s components, including all the wind turbines, wind farms, and the electric system. The collection and processing of data become crucial. Thus computational models that can provide predictions while minimising the numerical, physics and engineering uncertainties would be required. We discuss the role of such models in the subsequent sections.

Wind farms’ control in real-time with embedded predictive capabilities could be an area in which further research and development are required. Research should aim to develop a digital twin for wind farms where the energy provider could adjust parameters and demonstrate the impact of these parameters on the near-term energy supply. Digital twins can benefit from models for the power control of wind farms [48,49,50].

2.5. Other Challenges

The aerodynamic and aeroelastic challenges pose challenges in the materials used for manufacturing the wind turbine blades and the wind hubs. These challenges concern the material properties required for withstanding the excessive forces and the manufacturing integration of them and the associated costs. Moreover, the endurance of the wind turbine structures over several years is mandatory. Regarding materials science and engineering, the scales vary from the turbine to the small materials scales, i.e., down to millimetres. Other challenges in large turbines include the advanced engineering of sensors, load-bearing supports, electrical drive-train, bearings and lubricants, health monitoring, and solutions for extreme weather events such as typhoons.

Despite the wind energy benefits, wind turbines also encompass challenges beyond the technology. For example, Ellis and Ferraro [51] reported different factors that define wind energy social acceptance levels. Their report focused on significant impacts from onshore and offshore wind turbines, such as the landscape, health and well-being, noise, bio-diversity and property values. Eneveoldsen and Sovacool [52] investigated ways that may increase social acceptance in France of onshore wind projects. More recently, Taylor and Klenk [53] presented a review of probable health effects from wind turbines. They claimed that the health effects of wind turbines and farms emerge as a new topic of debate within social and political communities. They also discussed the vital role of science in providing more accurate value-based information to address the adverse impact of wind energy systems and provide future directives, e.g., enhanced policies for emerging renewable energy projects.

3. Challenges in Solar Energy

According to the IRENA (International Renewable Energy Agency), more than 25% of the world’s electricity requirements could be covered by solar power resources. However, solar conversion technologies encompass several technical challenges that must be addressed (Figure 3).

3.1. Intermittency

The intermittency of solar radiation poses a significant challenge. Research on silicon solar cells showed that light degrades the solar cells, while higher temperatures may facilitate recovery [54]. However, degradation and recovery also occur during exposure to sunlight and in darkness; degradation stops while recovery continues [54].

Understanding photo-injected electrons’ diffusion in different materials is an important topic, and intermittent light studies quantified the time-dependent photo-current growth in other semiconductors’ films [55]. Furthermore, integrating supercapacitors with solar panels can reduce intermittency [56].

Prasad et al. [57] coupled WRF (Weather Research Forecasting) models to different components of solar irradiance. Such modelling allows deeper analysis and enhanced forecasting of intermittency influenced by varied weather conditions. Manohar et al. [58] employed ML models to develop new protection schemes for PVE-integrated microgrids under different solar irradiance intermittency scenarios.

3.2. Air Pollution

Air pollution and dust accumulation significantly reduce solar power. Airborne dust particles constitute a significant challenge and an essential barrier to solar rays. Mani et al. [59] published a review on the impact of dust on PVE performance. Through experiments, Sulaiman et al. [60] and Hussain et al. [61] analysed the effects of dust and dirt particle accumulation on the performance of solar panels. Roumpakias et al. [62] investigated the effects of dust and aerosol particles on the efficiency of photovoltaic systems. They demonstrated that the aerosol scattering of different wavelengths could affect the PVE and PVT systems’ efficiency. For further details, the readers may refer to the detailed review by Salamah et al. [63] on the cleaning methods for improving the solar PV cells’ performance in different climate conditions.

Other studies include Bergin et al. [64], who studied the effect of atmospheric particulate matter (PM) on the diminishing solar energy production, focusing on large areas of India, China, and the Arabian Peninsula. In addition, Li et al. [65] shed light on aerosols in different regions of China, responsible for decreasing solar PV’s performance up to 35%. Finally, Son et al. 2020 [66] studied the particles’ effects on solar PV in the Republic of Korea. For poor air quality, they found a reduction rate exceeding 20%.

3.3. Storage

Intermittency and grid connectivity imply that the solar energy absorbed requires efficient storage. This ensures continuous energy conversion, production, and total energy use without interruption in different applications from building to industry.

PVE systems require electrical energy storage technologies such as lithium batteries, while PVT systems require thermal storage, such as reservoirs with phase change materials (PCM). Storage technologies based on electrochemical, sensible heat, latent heat, and thermochemical storage are still essential topics of research and development to overcome different persisting challenges. Recent research efforts aim to maximise storage capacity, period and compactness and reduce weight and cooling technologies [67,68].

3.4. Sustainability

Solar cells are usually made of processed crystalline materials such as silicon that have an average lifetime between 20 and 25 years. Before their use, they undergo complex engineering processes, which can have an environmental impact. Therefore, present and future efforts aim to optimise these processes and explore alternative materials to silicon to overcome future challenges, reduce ecological footprints, and simultaneously increase efficiency and lifetime. Reducing the environmental pollution footprint is a significant challenge to meet the decarbonisation goals on a large scale. Therefore, essential areas of future research should include greenhouse gas emissions analyses from PV technologies, materials and manufacturing; life cycle analyses of PV technologies; optimisation of the PV designs and their manufacturing processes, for example, to reduce the mass or material and increase the efficiency and lifetime period.

Research on PV materials’ recovery strategies is fundamental in terms of the long-term-added societal values and recycling. Tao et al. [69] presented a detailed review of feasible recycling technologies for solar PV modules, shedding light on their advantages and disadvantages. Researchers conducted a life cycle assessment (LCA) analysis of the recycling process of crystalline silicon (c-Si) and cadmium telluride (CdTe) PV technologies [70,71,72]. They provided information on the end-of-life stage of these PV modules, which facilitates the assessment of PV technologies in the future. Majewski et al. [73] provided a detailed description and analysis of the existing policies for recycling c-Si-based PV panels. Chen et al. 2021 [74] investigated the transparent conductors and recycling lead (Pb) from perovskite PV modules to reduce the PV modules’ recycling cost. They demonstrated that devices fabricated from the recycled lead iodide and transparent conductors exhibit similar performances to those devices manufactured from raw materials. Maria et al. 2021 [75] reviewed the solar PV Chain regarding the circular economy.

3.5. Grid Connectivity and Management

Smart grid connectivity and intelligent management of renewable energy resources are essential to fulfil the decarbonisation needs of the 21st century [76]. Nwaigwe et al. [77] discussed the solar-grid integration technologies and highlighted their benefits. David et al. [78] investigated solar energy management by employing a bibliometric analysis of published papers on this topic between 2000 and 2019. Kharrazi et al. [79] reviewed the assessment methods to quantify the impact of grid-connected PV modules installed on rooftops on the quality of power in low voltage distribution networks (LVDN) [80]. PV and EV (electric vehicles) systems integrated into local distribution systems are also an essential topic of research [81]. In the case of residential zones, Fachrizal et al. [81] showed that PV and EV exhibit a positive correlation for hosting capacity. Finally, Esteban et al. [82] discussed grid disturbances associated with solar energy generation. They showed California’s and Southwest’s USA net (hourly) energy generation and demand and the energy imbalance taking place in these states with the largest solar energy production. Computational modelling in optimising networks and grid distribution systems will thus play a significant role in overcoming all of the above challenges.

4. Modelling and Simulation

Computational modelling and simulations play a critical role in the wind and solar technology development. Figure 4 and Figure 5 illustrate the vital role of computational science in the multi-disciplinary design and optimisation of wind and solar technology components, equipment and systems. We review computational models and simulation techniques in the following sections.

4.1. Wind Energy

Atmospheric flow simulations are obtained by solving the Navier-Stokes equations (NSE). They incorporate the equations for mass, momentum and energy conservation. The equations can be used to model incompressible and compressible flows. For memory efficiency reasons, the solution of NSE is typically obtained through an iterative process. Block solutions, i.e., solving the equations in one iteration, lead to numerical convergence improvements [83]; however, the method is memory intensive. Various computational frameworks can be used to solve the equations, including Finite Volume (FV) [84] and Finite Element (FEM) methods [85,86]. These methods have been implemented in open-source and commercial CFD codes [87,88].

One of the biggest challenges in fluid flow simulations is capturing transition and turbulence. These phenomena exhibit coherent eddies at different spatial and temporal scales ranging from large ones to the Kolmogorov length scale. Transition and turbulence have significant flow physics and computational consequences. They strongly depend on the initial conditions, thus predicting the flow around a wind turbine and a wind farm stochastic. The most accurate computational approach would be to solve the discretised NSE on fine grids that capture all turbulent scales, i.e., Direct Numerical Simulation (DNS). DNS is a computationally expensive approach, scaling with the Reynolds number as $\mathcal{R}{e}^{9/4}$ [89]). Other methods separate scales. The large scales can be explicitly simulated via a subgrid-scale model (SGS), an approach that is used in the classical (or explicit) Large Eddy Simulation (LES). Another method is called the Implicit LES (ILES) [84,90,91,92,93,94,95,96,97,98]. ILES originates from the observations made in [90], i.e., the numerical dissipation of a class of methods called high-resolution [84,92] provide the same (or even better) results with SGS models used in classical LES.

For accurate LES, the grid resolution must resolve a significant part of the turbulent kinetic energy. As we refine the grid, the equations turn to DNS. ILES can be considered a coarse-grain DNS in simple terms, since no explicit turbulence model is used. Both explicit and implicit LES encompass challenges in near-wall flows at high Reynolds number flows [99]. However, ILES is overall a simpler and more robust approach in the implementation. Finally, LES and Reynolds Averaged Navier-Stokes (RANS) can be combined by using RANS in the near-wall region and LES away from this region [100,101].

Many wind energy simulations are still performed with RANS methods. Upon applying RANS modelling, the equations yield additional terms representing correlations between the velocity fluctuations, and similarly, there are correlations between temperature, pressure and velocity fluctuations. RANS assume turbulence to be isotropic and dissipative, so the Boussinesq approximation can be used to model the flow. Several models have been developed for modelling the fluctuations and obtaining the turbulent kinetic energy and turbulent flow dissipation [102,103,104,105]. All RANS models separate the mean steady flow and the stochastic turbulent component. However, this assumption breaks down in practice.

Modelling microscale physics in atmospheric flows also needs to address multiphase flow effects [106,107,108,109,110]). The complete atmospheric flow modelling should address the dispersed multiphase flow, where two or more components are present. The domain may contain discrete particles (solid particles), liquid droplets or gas bubbles. The modelling methodology will depend on the mixture characteristics and the desired accuracy of numerical predictions. At low phase or volume fractions (e.g., a small number of solid particles), kinetic approaches can be employed at an affordable computational cost. These methods allow for evaluating the fluid forces acting on each particle individually, e.g., Discrete Element Method (DEM) codes [111]. At high volume fraction values, the dispersed phase(s) are usually treated as continual, where phase-averaging techniques derive separate sets of NSE for each phase [112]. Finally, more complex models can incorporate particle–particle kinetics solving at the sub-grid scale [113,114,115].

4.2. Solar Energy

Computational modelling and simulation tools of PVE and PVT systems are classified into three categories:

• Reduced order models based on the integration of PV modules in operating cycles for different power types generation, e.g., thermodynamic cycles simulation software, Matlab/Simulink.
• Statistical [116], ML and AI models for an increased prediction accuracy (forecasting) and thus higher confidence in the PV system continuous production and performance, e.g., enhanced solar energy management and conversion.
• CFD and optimisation tools for enhanced efficiency of PV systems at the components level, i.e., materials, cooling technologies for energy storage, thermal isolation and regulation for thermal energy storage.

Bourdoucen et al. [117] presented the analytical modelling and simulation of PV panels considering solar cells’ different physical and connectivity parameters, such as temperature, irradiance, and array configuration. Campana et al. [118] developed a dynamic modelling tool for PV water pumping systems (PVWPS). The tool allows dynamic simulations for an optimal design of PVWPS depending on operating conditions and dynamic water demand. Vogt et al. [119] developed models to simulate the c-Si PV modules’ performance. They coupled the semiconductor modelling via FEM to account for thermal conduction, convection and radiation. Their physical model allows the prediction of the percentage of the sun’s intensity that becomes parasitically absorbed within different solar modules under a specific spectrum. Neelagam et al. [120] developed ML models for predicting the solar radiation for solar systems. Their numerical techniques were based on NN models with different backpropagation algorithms. Vinod et al. [121] developed a solar PV modelling and simulation tool on top of Matlab/Simulink that considers the performance of PV modules at actual metrological data such as irradiance and weather temperature. Finally, Lee et al. [122] developed a hybrid agent-based modelling of rooftop PV systems. They employed geographic information and data mining techniques for enhancing the solar energy conversion and management.

5. Artificial Intelligence and Optimisation

5.1. Machine Learning

Coupling numerical methods with data-driven algorithms using ML can help accelerate computational predictions [12]. For ML, we adopt the definition of Mitchell et al. [123], i.e., (paraphrased slightly from the above paper), a computer program is considered to learn from experience E for some tasks T and performance measured as P if its performance for T, as measured by P, improves with experience E. Thus, ML allows one to solve problems through learning from data.

ML can speed up predictions subject to training ML models with sufficient data. However, ML cannot replace the flow, acoustic and aeroelastic simulations, as these simulations, in conjunction with field data and wind tunnel measurements, are needed to generate the data for developing the ML models. Nevertheless, securing sufficient and reliable data can develop ML prediction models for engineering predictions that could assist the design and implementation phases. Furthermore, ML can help extract patterns from large data sets (e.g., obtained by sensors or high-performance computing simulations). Researchers have applied ML in image, text and signal processing, autonomous driving, mechanical damage detection and natural language processing. Thus, several industrial sectors invest in incorporating ML techniques into their practices.

Wind energy technologies require real-time predictions, whether for safety-critical reasons or testing many variants of wind turbines or wind farms. Similarly, the wind energy sector can significantly benefit from such developments.

ML has been used in several engineering applications. Some examples are provided below to demonstrate ML’s potential. For example, NN has been used in predicting phase configuration and volume fractions of multiphase flows [124,125]; multiscale modelling of flows and heat transfer [126]; and flow control [127].

Similarly, ML was used in solid mechanics for a long time. Examples include understanding the stress–strain relationship of materials [128,129,130,131]. In addition, ML can be beneficial in aeroelasticity to calculate the vibrational properties of wind turbines. There are already examples of using NN in structural dynamics [132,133], and the design of materials [134]. All the above are areas that are hugely crucial for wind energy technologies.

Applications of ML in the manufacturing and health sectors, are rapidly growing [12,135]. Moreover, ML can be used for the damage detection of different structures and systems [136,137,138,139]. For example, Long-Short Term Memory (LSTM) models were employed in real-time aircraft anomalies detection [140]. Similarly, these methods can be applied to detecting damages on wind turbines.

Another area in wind technology that could benefit from ML is the optimisation of wind turbines and wind farms, which requires hundreds of simulations. For example, in [141], the researchers employed NN to predict the fluid–structure interaction in non-well-resolved surfaces accurately.

The future of engineering designs and systems’ operation will require accurate performance forecasting and real-time predictions. Thus, future developments will concern digital twins for engineering systems. This would require precise performance at all computational levels, e.g., from the performance of an aerodynamic blade to the complete wind farm. Designing ML algorithms for such dynamical behaviour is essential to unsteady wind turbine and wind farm simulations. The most basic form of using ML for transient problems requires accurate predictions at consecutive times. In turn, ML will output the variable’s value for a future time. Furthermore, the time-series data can be compressed using some dimensionality reduction algorithm. This approach was used successfully to predict the behaviour of various physical systems such as the dampened harmonic oscillators [142] and the 2D Navier-Stokes [143,144].

For predictions over a more extended period, ML can redirect its output as a new input for longer time forecasts. Simulation of wind energy systems always depends on the initial and boundary conditions. Thus, setting up the simulations accurately is crucial [145]. The error for highly chaotic systems increases in time [146,147]. More recently, regression forests [148] and LSTM [149] were used for transient flow physics.

Sparse Identification of Nonlinear Dynamics (SINDy) [150] and its variants [151,152] are alternative approaches. SINDy reconstructs governing equations using data obtained across different times. The user empirically selects a few non-linear functions for the input features in conjunction with sparse regression. In theory, the methods can predict outside the data’s space boundaries. However, in practice, the sparsity of the data may make the above implementation difficult. The idea is to develop a method unrestricted by the training data space. The challenge is to make such a method work for non-perfect (polynomial) systems Pan and Duraisamy [152], which increase the complexity and cannot be decomposed into a sparse number of functions.

Other approaches use Bayesian inference for numerical homogenisation [153], i.e., finding solutions to partial differential equations. Subsequent studies have used Gaussian Process (GP) regression to discover the underlying linear, and non-linear [154] differential equations, such as the Navier-Stokes equations. GP-based methods can provide solutions to differential equations for various initial conditions. Although GP allows the incorporation of information through covariance operators, developing the physics-informed NN is a more challenging task considering NN’s black-box nature and the physics challenges associated with complex phenomena such as turbulence.

NN can be used for automatic differentiation to calculate derivatives for a broad range of functions [155]. In addition, one can apply automatic differentiation with supervised learning. An interesting example is the lift calculation around a bluff body, using sparse flow field data [154].

5.2. Optimisation

Physics-based optimisation is a rapidly growing field of computational science thanks to the progress in computational power resources (developing data centres and high-performance cloud computing). Optimisation methods coupled with CFD, and other physics-based techniques, are usually highly non-linear constrained-based optimisation problems. The solution algorithms to these complex numerical optimisation problems can be classified into gradient-based and gradient-free optimisation algorithms. They are continuous or discrete adjoint-based optimisation algorithms. Both fall within the class of gradient-based methods that are very promising due to their high accuracy at a less expensive computational time than gradient-free algorithms, e.g., genetic algorithms. This is in addition to the ease of mathematical constraints formulation possibility in multi-objective optimisation problems [14,15,16,17,20]. For example, Anderson et al. [18] developed an adjoint-based multidisciplinary optimisation platform for optimising the structural properties of a 13 m scaled wind farm technology facility. They coupled the optimisation problem to the RANS fluid dynamics solver within a structural finite element method framework. Xiao et al. [19] developed an advanced adjoint-based electromagnetic optimisation to design an optimal diffracting optical element (spectral-splitting) that is then fabricated by a 3D printing technology.

6. Conclusions

Wind and solar energy are essential alternatives for producing clean energy. However, they encompass several challenges, as we outlined in this review. These include challenges in the design and manufacturing, implementation of the technologies, storage, and sustainable energy provision on a large scale, considering the population growth. Undoubtedly, wind and solar will play an important role in the energy transition. At the same time, nuclear energy, which is not covered by this review, is a vital energy sector that will also play an important role in achieving this transition and reducing dependence on fossil fuels. However, solar and wind will increase in use considering the closure of many nuclear plants in the past few decades and the time required for building new nuclear plants.

Solar and wind power technologies (solar cells, wind turbines and batteries) depend extensively on a wide range of rare earth minerals and materials (e.g., c-Si, CdTe, Pb, graphite, cobalt, etc.). Therefore, the success of these technologies in providing continuous energy power on a large scale depends on the sourcing and costs of rare earth minerals and materials. Moreover, areas (and states) hosting these materials are at a high risk of geopolitical conflict [156,157,158]. Thus, global energy security and supply chains are critical for the sustainable implementation of renewable energy technologies.

Computational science can accelerate the improvement of wind and solar technologies. For example, CFD and computational multi-physics models can be combined with ML and optimisation algorithms to improve the design of wind turbines and solar panels and guide the optimal implementation. With the advancement of large onshore and offshore wind engineering farms, there is also a large-scale deployment of sensors to monitor the systems’ performance. ML can be used for processing large datasets produced by sensors. A few examples of ML future applications for the wind sector include massive (fine-grain) fluid flow simulations around wind farms and geographical terrains that uncover complex flow features, which are difficult to discover through traditional visualisation. In addition, ML can classify physical patterns and identify correlations.

Furthermore, ML could also be used as a surrogate model. Surrogate models can be used both for present and future time predictions. In addition, Recurrent Neural Networks (RNN) are suited for dynamic models. Although they are complex in terms of training, they will play a more critical role in the future. Finally, ML can aid pre-existing computational models by augmenting some of the model’s features with a data-driven component.

The current trend in ML algorithms is to develop the supervised learning trained on high-fidelity data. Supervised learning uses parts of the data for training and another for testing and validating ML and its hyperparameters. Therefore, it is essential to identify under or over-fitting, the uncertainty associated with different ML algorithms, and select the most optimal method. Ultimately, the algorithm choice will depend on the training data availability and the feature vector’s size. This approach can provide potentially essential results for wind energy technologies, because performing large-scale experiments and fine-grain numerical simulations are expensive.

Computational science can also advance wind and solar designs and optimise grid performance efficiency. However, implementing computational models and optimisation techniques on a large scale, e.g., wind farms, remains a computational challenge. Furthermore, physical modelling coupled with optimisation algorithms requires a further development to optimise the solar PV components considering the multiscale level of their integration in more extensive systems for different types of power generation (electrical; thermal). ML and, subsequently, AI, must show their potential with diverse data from wind, solar, meteorological, and grid network sources.