From Waves to Shadows

Abstract

Solar photovoltaic (PV) energy has the potential to become a major source of electricity production worldwide. However, the deployment of this renewable energy source comes with several challenges. Besides societal, political, and grid-limiting, several technical limitations decrease the trust in this technology. Within this context arises the Trust-PV project, whose objective is to improve the performance and reliability of photovoltaic systems.

This work contributes to this European project by exploring different power prediction models for several types of PV systems. Considering the broadness of the topic, four parts or blocks are identified. The first part deals with machine learning models to forecast the yield of residential PV systems. The second block focuses on analytical models used during the design phase. The third part is dedicated to systems floating on water. Lastly, a metric to assess the tolerance towards shading of different modules is developed in the fourth block.

Starting with the first block of machine learning techniques for PV power forecasting, Chapter 2 introduces the topic by reviewing a large number of manuscripts. The chapter performs a broad classification of the reviewed literature with the objective to identify trends and gaps in the field. Among the identified trends, one can highlight the high percentage of predictions for the day ahead, the generally low number of systems employed to train the models, and the concentration of systems in mild climates.

The latter points may stem from researchers primarily using the systems available within their institutions. To promote collaboration, Chapter 3 presents a developed website that lists PV power open source databases. The website aims to encourage researchers to train and test their models with different data sources.

One consequence of the concentration of systems in mild climates is that the effect of climate on machine learning models remains underexplored in the literature. Chapter 4 addresses this gap by studying how machine learning models behave for systems located in different climatic zones. The results show that weather homogeneity affects the accuracy of the models. Models developed for systems located in uniform climates - like desert areas - achieve in general higher accuracy than the models developed for systems in highly varying climates - like tropical areas.

Chapter 5 addresses another challenge: creating a single machine learning model able to monitor the performance of a large fleet of residential PV systems. The developed model surpasses in accuracy an analytical reference model but is limited by a fundamental characteristic of machine learning methods: the focus on large errors which resulted in the overlooking of smaller systems. Consequently, Chapter 6 develops a different approach based on the peer-to-peer methodology. In this approach, the power output of similar neighboring systems is compared to identify any malfunctions. The method is tested for the residential fleet of PV systems and proves effective for detecting faults.

Moving on to the second block of analytical power predictions, Chapter 7 presents the PVMD toolbox, a state-of-the-art analytical simulation framework that can predict the power of systems that do not exist yet. The abilities of the toolbox are tested for residential systems in the same chapter and the results show the negative influence that inaccurate input irradiance data has on the predictions.

This importance of accurate irradiance data affects all kinds of PV systems, but especially large-scale ones. Therefore, to monitor them, a proper allocation of irradiance sensors is essential. Hence, in Chapter 8, a software tool is developed to identify the optimal number of irradiance sensors and their position in a PV farm. The tool’s strengths are more prominent in plants located on terrains with significant elevation changes.

The third block focuses on PV systems that are installed on floating platforms rather than on land. Chapter 9 introduces the topic by examining three factors influenced by proximity to water that can impact the production of a floating PV system in a French quarry lake: movement fluctuations, dust accumulation, and module temperature. The results reveal a limited influence of all factors on the production for the period of study therefore facilitating the deployment of floating systems.

The block continues by studying the effect of sea waves for a system located in the North Sea. The simulation results from Chapter 10 reveal that wave fluctuations can have a negative yet limited effect on the DC and AC yield of floating PV systems. These results are further elaborated in Chapter 11, where the model is improved by considering the fluid-structure interaction. This advanced model enables to study the effect of various platform characteristics on the power mismatch losses. The results reveal a trade-off between mechanical stability and mismatch losses.

Finally, the last part deals with the power lost when a PV module is partially shaded. Chapter 12 develops a simulation tool to efficiently calculate the shading tolerability of a PV module given its datasheet. The shading tolerability is a metric that quantifies the resilience towards shading of a PV module, that is how much power is lost when the module is partially shaded. The developed tool is used to create a database of shading tolerability of commercial PV modules, to compare the resilience of different modules towards shading.