In the context of increasing global energy demand, fossil power plants remain the main contributor to atmospheric carbon emissions. In the last 15 years, renewable energy generation has increased drastically, with solar photovoltaics (PV) being the fastest growing, expanding from 56 TWh in 2011 to 2,780 TWh in 2025. Among these technologies, Building-Integrated Photovoltaics (BIPV) offer an effective strategy for urban decarbonization. By generating electricity directly at the point of consumption, BIPV systems eliminate conventional grid transmission losses. However, evaluating BIPV systems during early-stage architectural design remains challenging due to the computational complexity of predicting visual appearance, modeling spatial context, and optimizing panel placement under localized shading. Current simulation tools rely heavily on manual processes, while traditional evolutionary algorithms treat panel placement purely as a numerical sorting problem. To resolve these limitations, this thesis presents an automated, Artificial Intelligence (AI)-driven computational pipeline designed to predict BIPV appearance, optimize spatial distribution, and forecast energy yields. The developed framework extracts localized urban context via the OpenStreetMap Overpass API and dynamically parses this payload into the Rhino simulation environment. Both the geometric alteration of buildings and the aggregation of baseline environmental inputs are programmatically controlled by an orchestrating AI agent. To address the NP-hard (Non-deterministic Polynomial-time hard) problem of combinatorial panel placement optimization, a Machine Learning (ML) approach offers a computationally efficient alternative. A Convolutional Neural Network (CNN) architecture was developed to map complex spatial feature tensors and rapidly isolate near-optimal layout arrays. The optimization framework evaluates spatial configurations against the temporal load profile of a building, balancing self-consumption (SC) and self-sufficiency (SS) metrics. The results demonstrate that the neural network drastically reduces optimization time compared to traditional brute-force evolutionary algorithms while maintaining high spatial accuracy, bridging automated GIS workflows with deep learning for sustainable cities.

In the context of increasing global energy demand, fossil power plants remain the main contributor to atmospheric carbon emissions. In the last 15 years, renewable energy generation has increased drastically, with solar photovoltaics (PV) being the fastest growing, expanding from 56 TWh in 2011 to 2,780 TWh in 2025. Among these technologies, Building-Integrated Photovoltaics (BIPV) offer an effective strategy for urban decarbonization. By generating electricity directly at the point of consumption, BIPV systems eliminate conventional grid transmission losses. However, evaluating BIPV systems during early-stage architectural design remains challenging due to the computational complexity of predicting visual appearance, modeling spatial context, and optimizing panel placement under localized shading. Current simulation tools rely heavily on manual processes, while traditional evolutionary algorithms treat panel placement purely as a numerical sorting problem. To resolve these limitations, this thesis presents an automated, Artificial Intelligence (AI)-driven computational pipeline designed to predict BIPV appearance, optimize spatial distribution, and forecast energy yields. The developed framework extracts localized urban context via the OpenStreetMap Overpass API and dynamically parses this payload into the Rhino simulation environment. Both the geometric alteration of buildings and the aggregation of baseline environmental inputs are programmatically controlled by an orchestrating AI agent. To address the NP-hard (Non-deterministic Polynomial-time hard) problem of combinatorial panel placement optimization, a Machine Learning (ML) approach offers a computationally efficient alternative. A Convolutional Neural Network (CNN) architecture was developed to map complex spatial feature tensors and rapidly isolate near-optimal layout arrays. The optimization framework evaluates spatial configurations against the temporal load profile of a building, balancing self-consumption (SC) and self-sufficiency (SS) metrics. The results demonstrate that the neural network drastically reduces optimization time compared to traditional brute-force evolutionary algorithms while maintaining high spatial accuracy, bridging automated GIS workflows with deep learning for sustainable cities.

Implementation ML and AI solutions for BIPV modeling simulation

LALAIAN, ALBERT
2025/2026

Abstract

In the context of increasing global energy demand, fossil power plants remain the main contributor to atmospheric carbon emissions. In the last 15 years, renewable energy generation has increased drastically, with solar photovoltaics (PV) being the fastest growing, expanding from 56 TWh in 2011 to 2,780 TWh in 2025. Among these technologies, Building-Integrated Photovoltaics (BIPV) offer an effective strategy for urban decarbonization. By generating electricity directly at the point of consumption, BIPV systems eliminate conventional grid transmission losses. However, evaluating BIPV systems during early-stage architectural design remains challenging due to the computational complexity of predicting visual appearance, modeling spatial context, and optimizing panel placement under localized shading. Current simulation tools rely heavily on manual processes, while traditional evolutionary algorithms treat panel placement purely as a numerical sorting problem. To resolve these limitations, this thesis presents an automated, Artificial Intelligence (AI)-driven computational pipeline designed to predict BIPV appearance, optimize spatial distribution, and forecast energy yields. The developed framework extracts localized urban context via the OpenStreetMap Overpass API and dynamically parses this payload into the Rhino simulation environment. Both the geometric alteration of buildings and the aggregation of baseline environmental inputs are programmatically controlled by an orchestrating AI agent. To address the NP-hard (Non-deterministic Polynomial-time hard) problem of combinatorial panel placement optimization, a Machine Learning (ML) approach offers a computationally efficient alternative. A Convolutional Neural Network (CNN) architecture was developed to map complex spatial feature tensors and rapidly isolate near-optimal layout arrays. The optimization framework evaluates spatial configurations against the temporal load profile of a building, balancing self-consumption (SC) and self-sufficiency (SS) metrics. The results demonstrate that the neural network drastically reduces optimization time compared to traditional brute-force evolutionary algorithms while maintaining high spatial accuracy, bridging automated GIS workflows with deep learning for sustainable cities.
2025
Implementation ML and AI solutions for BIPV modeling simulation
In the context of increasing global energy demand, fossil power plants remain the main contributor to atmospheric carbon emissions. In the last 15 years, renewable energy generation has increased drastically, with solar photovoltaics (PV) being the fastest growing, expanding from 56 TWh in 2011 to 2,780 TWh in 2025. Among these technologies, Building-Integrated Photovoltaics (BIPV) offer an effective strategy for urban decarbonization. By generating electricity directly at the point of consumption, BIPV systems eliminate conventional grid transmission losses. However, evaluating BIPV systems during early-stage architectural design remains challenging due to the computational complexity of predicting visual appearance, modeling spatial context, and optimizing panel placement under localized shading. Current simulation tools rely heavily on manual processes, while traditional evolutionary algorithms treat panel placement purely as a numerical sorting problem. To resolve these limitations, this thesis presents an automated, Artificial Intelligence (AI)-driven computational pipeline designed to predict BIPV appearance, optimize spatial distribution, and forecast energy yields. The developed framework extracts localized urban context via the OpenStreetMap Overpass API and dynamically parses this payload into the Rhino simulation environment. Both the geometric alteration of buildings and the aggregation of baseline environmental inputs are programmatically controlled by an orchestrating AI agent. To address the NP-hard (Non-deterministic Polynomial-time hard) problem of combinatorial panel placement optimization, a Machine Learning (ML) approach offers a computationally efficient alternative. A Convolutional Neural Network (CNN) architecture was developed to map complex spatial feature tensors and rapidly isolate near-optimal layout arrays. The optimization framework evaluates spatial configurations against the temporal load profile of a building, balancing self-consumption (SC) and self-sufficiency (SS) metrics. The results demonstrate that the neural network drastically reduces optimization time compared to traditional brute-force evolutionary algorithms while maintaining high spatial accuracy, bridging automated GIS workflows with deep learning for sustainable cities.
Machine Learning
AI
BIPV
CNN
Photovoltaic
File in questo prodotto:
File Dimensione Formato  
Lalaian_Albert.pdf

Accesso riservato

Dimensione 9.02 MB
Formato Adobe PDF
9.02 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

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.12608/116323