The objective of this project is to develop a model for estimating the power production of grid-connected photovoltaic (PV) plants. The work was conducted during an internship at ORTUS Italy S.r.l, a company operating in the development of PV plants, wind farms and battery energy storage systems (BESS). The project focuses on a ground-mounted PV plant with a rated capacity of 4.165 MWp, featuring a single-axis tracking system and bifacial modules, developed by ORTUS Italy in Gaibanella (FE, Italy). In the first phase, the plant was modelled using PVSyst, a widely adopted commercial software in the PV industry. This simulation produced hourly time series of the power output and the independent variables influencing it. These data served as the basis for developing a Machine Learning model using the Matlab Statistics and Machine Learning Toolbox. The model was trained to predict the output power of the subarrays that make up the plant, on the basis of predictors values. After evaluating the performance of various models on training data, they were applied to data coming from plant measurements. Specifically, after developing a model that included the module temperature as a feature, a second model was created for scenarios where this measurement is unavailable. This required selecting alternative predictors to compensate for the missing temperature data. In the second part of the project, a physical model of the plant was developed based on equations describing the entire process of energy conversion from solar radiation to electrical energy injected into the grid. In this phase, a reverse transposition model was implemented to reconstruct the global horizontal irradiance (GHI) from the measured tilted irradiance (GTI). This GHI data was then fed into a 2D model of the plant to estimate the available irradiance on both the front and rear surfaces of the modules, accounting for reflections from the ground, module surfaces, and obstruction of the sky from neighbouring rows. Subsequently, the Single Diode Model was used to calculate the power output of the sub-arrays. The irradiance evaluation using the 2D model was implemented in Python with the pvfactors library, developed by SunPower and Sandia National Laboratories. Similarly, the power output calculation was performed using the pvlib library, originally developed by Sandia National Laboratories and now open-source. Finally, an analytical evaluation of AC-side losses was carried out for both model types, based on datasheet information for the inverters, cables, and transformer.
L’obiettivo di questo progetto è la realizzazione di un modello per stimare la produzione fotovoltaica in grandi impianti connessi alla rete nazionale. Il lavoro è stato svolto durante uno stage presso ORTUS Italy S.r.l., una società specializzata nello sviluppo di impianti fotovoltaici, eolici e sistemi di accumulo (BESS). Il progetto prende in esame un impianto fotovoltaico a terra da 4,165 MWp, dotato di un sistema di inseguimento monoassiale e di moduli bifacciali, sviluppato da ORTUS Italy a Gaibanella (FE). Nella prima fase, l’impianto è stato modellizzato utilizzando PVSyst, un software commerciale ampiamente adottato nell’industria fotovoltaica. La simulazione ha prodotto serie temporali orarie della potenza generata e delle variabili indipendenti da cui essa dipende. Questi dati sono stati utilizzati per sviluppare un modello di Machine Learning nell’ambiente Matlab Statistics and Machine Learning Toolbox. Il modello è stato addestrato per prevedere la potenza in uscita dai sub-array dell’impianto sulla base delle variabili predittive. Dopo aver valutato le prestazioni di diversi tipi di modelli sui dati di addestramento, questi sono stati applicati ai dati misurati in impianto. In particolare, dopo aver sviluppato un modello che includeva la misura della temperatura dei moduli, è stato realizzato un secondo modello per i casi in cui tale misura non sia disponibile, identificando le variabili predittive necessarie per compensare questa mancanza. Nella seconda parte del progetto è stato sviluppato un modello fisico dell’impianto, basato sulle equazioni che descrivono l’intero processo di conversione dall’energia solare all’energia elettrica immessa in rete. In questa fase è stato implementato un modello di reverse transposition per ricostruire l’irradianza globale sul piano orizzontale (GHI) a partire dall’irradianza misurata sul piano inclinato (GTI). Questa variabile è stata poi utilizzata in un modello bidimensionale dell’impianto per stimare l’irradianza disponibile sulla superficie frontale e posteriore dei moduli, tenendo conto degli effetti di riflessione del terreno, delle superfici dei moduli e della ridotta vista del cielo a causa delle file di moduli adiacenti. Successivamente, è stato implementato il Single Diode Model per calcolare la potenza in uscita dai sub-array dell’impianto. La valutazione dell’irradianza con il modello bidimensionale è stata realizzata in Python utilizzando la libreria pvfactors, sviluppata da SunPower e Sandia National Laboratories. Analogamente, il calcolo della potenza è stato effettuato con la libreria pvlib, inizialmente sviluppata da Sandia National Laboratories e ora open-source. Infine, per entrambe le tipologie di modelli è stata effettuata una valutazione analitica delle perdite nella parte AC dell’impianto, basata sui dati dei datasheet di inverter, cavi e trasformatore.
Comparative analysis of Machine-Learning and physical models for PV power plant production estimation
PINATO, MICHELE
2024/2025
Abstract
The objective of this project is to develop a model for estimating the power production of grid-connected photovoltaic (PV) plants. The work was conducted during an internship at ORTUS Italy S.r.l, a company operating in the development of PV plants, wind farms and battery energy storage systems (BESS). The project focuses on a ground-mounted PV plant with a rated capacity of 4.165 MWp, featuring a single-axis tracking system and bifacial modules, developed by ORTUS Italy in Gaibanella (FE, Italy). In the first phase, the plant was modelled using PVSyst, a widely adopted commercial software in the PV industry. This simulation produced hourly time series of the power output and the independent variables influencing it. These data served as the basis for developing a Machine Learning model using the Matlab Statistics and Machine Learning Toolbox. The model was trained to predict the output power of the subarrays that make up the plant, on the basis of predictors values. After evaluating the performance of various models on training data, they were applied to data coming from plant measurements. Specifically, after developing a model that included the module temperature as a feature, a second model was created for scenarios where this measurement is unavailable. This required selecting alternative predictors to compensate for the missing temperature data. In the second part of the project, a physical model of the plant was developed based on equations describing the entire process of energy conversion from solar radiation to electrical energy injected into the grid. In this phase, a reverse transposition model was implemented to reconstruct the global horizontal irradiance (GHI) from the measured tilted irradiance (GTI). This GHI data was then fed into a 2D model of the plant to estimate the available irradiance on both the front and rear surfaces of the modules, accounting for reflections from the ground, module surfaces, and obstruction of the sky from neighbouring rows. Subsequently, the Single Diode Model was used to calculate the power output of the sub-arrays. The irradiance evaluation using the 2D model was implemented in Python with the pvfactors library, developed by SunPower and Sandia National Laboratories. Similarly, the power output calculation was performed using the pvlib library, originally developed by Sandia National Laboratories and now open-source. Finally, an analytical evaluation of AC-side losses was carried out for both model types, based on datasheet information for the inverters, cables, and transformer.| File | Dimensione | Formato | |
|---|---|---|---|
|
PINATO_MICHELE.pdf
Accesso riservato
Dimensione
21.32 MB
Formato
Adobe PDF
|
21.32 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/82343