In this thesis, a framework is presented for developing a surrogate model to approximate the atmospheric dispersion of pollutants released from a wind turbine. High-fidelity dispersion models, while accurate, are often too computationally expensive and time-consuming for real-time prediction and decision-making. Surrogate models are data-driven computationally-efficient tools that mimic the behaviour of complex engineering problems when fast and efficient results are required. In this work, firstly a high-fidelity model is proposed and developed that describes the problem under study, and is then approximated by the surrogate realization. The model, using a Finite Element approximation of the advection-diffusion equations, simulates the transport, diffusion, and deposition of an oil-based pollutant emitted from a source at the top of the tower of a wind turbine. Once defined, it is used to compute the probability of the ground deposition of the pollutant in the proximity of the wind turbine affected by the oil leak. From this numerical model, a synthetic dataset for training the surrogate model was generated, using a Design of Experiments (DoE) methodology, ensuring representative coverage of the input space despite the limited number of simulations. Finally, several Machine Learning surrogate model solutions are presented. Approaches ranged from linear regression to deep neural networks, combined with dimensionality reduction tools such as Principal Component Analysis (PCA) and Autoencoders, applied to handle high-dimensional output grids. The models are tested, and their accuracy metrics compared, especially focusing on their capability to predict the most probable spatial location of the deposition. Among the tested methods, a combined Autoencoder–Neural Network architecture demonstrated the best performance in reproducing the spatial patterns of deposition.
In questa tesi viene presentato una metodologia per lo sviluppo di un modello surrogato volto ad approssimare la dispersione atmosferica di inquinanti rilasciati da una turbina eolica. I modelli di dispersione computazionali ad alta accuratezza, risultano spesso troppo costosi in termini computazionali e di tempo per previsioni in tempo reale e processi decisionali rapidi. I modelli surrogati sono strumenti data-driven e computazionalmente efficienti che riproducono il comportamento di problemi ingegneristici complessi quando sono richiesti risultati rapidi ed efficienti. In questo lavoro, viene innanzitutto proposto e sviluppato un modello computazionale che descrive il problema in esame e che viene successivamente approssimato mediante il modello surrogato. Il modello, basato su un'approssimazione agli elementi finiti delle equazioni di diffusione-trasporto, simula lo spostamento, la diffusione e la deposizione di un inquinante di origine oleosa emesso da una sorgente situata sulla sommità della torre di una turbina eolica. Una volta definito, il modello viene utilizzato per calcolare la probabilità di deposizione a terra dell’inquinante nelle aree circostanti la turbina interessata dalla perdita accidentale. Dal modello numerico viene generato un datasbase sintetico per l’addestramento del modello surrogato, utilizzando una metodologia, parte dei Design of Experiments (DoE), che garantisce una copertura rappresentativa dello spazio degli input nonostante il numero limitato di simulazioni. Infine, vengono presentate diverse soluzioni di modelli surrogati basati su metodologie Machine Learning. Gli approcci considerati spaziano dalla regressione lineare a reti neurali, combinati con strumenti di riduzione della dimensionalità come l’Analisi delle Componenti Principali (PCA) e gli Autoencoder, applicati per gestire l’elevata dimensionalità delle griglie di deposizione di output. I modelli sono stati testati e confrontati tramite parametri di accuratezza, concentrandosi in particolare sulla loro capacità di prevedere la localizzazione spaziale più probabile della deposizione. Tra i metodi testati, un’architettura combinata Autoencoder–Rete Neurale ha mostrato le migliori prestazioni nel riprodurre i pattern spaziali della deposizione.
Numerical implementation of an advection-diffusion model and machine learning emulator for pollutant dispersion from wind turbines
PARINI, NICOLA
2024/2025
Abstract
In this thesis, a framework is presented for developing a surrogate model to approximate the atmospheric dispersion of pollutants released from a wind turbine. High-fidelity dispersion models, while accurate, are often too computationally expensive and time-consuming for real-time prediction and decision-making. Surrogate models are data-driven computationally-efficient tools that mimic the behaviour of complex engineering problems when fast and efficient results are required. In this work, firstly a high-fidelity model is proposed and developed that describes the problem under study, and is then approximated by the surrogate realization. The model, using a Finite Element approximation of the advection-diffusion equations, simulates the transport, diffusion, and deposition of an oil-based pollutant emitted from a source at the top of the tower of a wind turbine. Once defined, it is used to compute the probability of the ground deposition of the pollutant in the proximity of the wind turbine affected by the oil leak. From this numerical model, a synthetic dataset for training the surrogate model was generated, using a Design of Experiments (DoE) methodology, ensuring representative coverage of the input space despite the limited number of simulations. Finally, several Machine Learning surrogate model solutions are presented. Approaches ranged from linear regression to deep neural networks, combined with dimensionality reduction tools such as Principal Component Analysis (PCA) and Autoencoders, applied to handle high-dimensional output grids. The models are tested, and their accuracy metrics compared, especially focusing on their capability to predict the most probable spatial location of the deposition. Among the tested methods, a combined Autoencoder–Neural Network architecture demonstrated the best performance in reproducing the spatial patterns of deposition.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_10_Parini.pdf
accessibile in internet solo dagli utenti autorizzati
Dimensione
8.07 MB
Formato
Adobe PDF
|
8.07 MB | Adobe PDF | Visualizza/Apri |
|
2025_10_Parini_Executive_Summary.pdf
accessibile in internet solo dagli utenti autorizzati
Dimensione
1.11 MB
Formato
Adobe PDF
|
1.11 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/244017