Estimación del potencial fotovoltaico en azoteas utilizando Redes Neuronales = Estimation of rooftop photovoltaic potential using Neural Networks

Photovoltaics have seen an increase in popularity and a decrease in associated costs over the last years. Each year more countries introduce policies that allow the exchange of energy from individuals to the grid with feed-in tariffs for renewable energies, making household production more profitable. A future in which households in certain regions generate most of their electricity demand seems closer every year, with all the associated benefits of increasing renewable energy shares and increased independence from the grid. However, this raises the question: how can we make informed decisions about which areas represent the highest economic potential? In this master thesis, a methodology for estimating PV potential using satellite imagery and open data is developed. This has the benefits of being free and worldwide available, over other approaches that use LiDAR data to perform the simulations. To achieve this, a Deep Learning model is used to segment the input satellite image of the region of interest, and image processing techniques are implemented to learn more about the morphology of each predicted building. Finally, micro simulations are performed in every detected rooftop, which is afterward aggregated and presented in the form of energy produced per year and its translation into economic terms. This approach has the opportunity to provide local users with an assessment of their region, how profitable the panels would be, and which regions have the highest ROI first. Results show reasonable estimations for German regions, where images are more similar to the data sources. In Aachen, a potential equivalent to 53:45% of its current electricity demand was estimated. Other regions with highly different building morphologies underperform, showing overestimations in Spanish and Indian regions