AI-Powered Solar Mapping: A Multimodal Approach for Scalable and Automated Photovoltaic Site Assessment

In recent years, Europe has faced an increasing need to address climate change and reduce its dependence on fossil fuels. As a result, several policy strategies have been implemented to promote the adoption of renewable energy systems, such as photovoltaic (PV) installations. The REPowerEU plan, introduced in May 2022, mandates the installation of solar panels on commercial and public buildings starting in 2025 and on new residential buildings from 2029. Identifying suitable roofs for the installation of photovoltaic systems is therefore a fundamental task. Although useful simulation tools such as PVIGIS assist in estimating the power output of photovoltaic systems under different hypotheses, this thesis explores how Artificial Intelligence (AI) can support this process, making it more efficient and scalable. In the conducted study, a multimodal neural network that combines images and climate data was developed to forecast the annual energy production. By merging these modalities, the model’s robustness is improved, leading to more reliable predictions on unseen data. The training material integrates OpenStreetMap data, high-resolution aerial orthophotos of Switzerland, and Typical Meteorological Year (TMY) data obtained through the PVGIS API. The predictions obtained are consistent with those achieved with traditional simulation and forecasting tools. However, unlike conventional methods, this machine learning-based approach offers greater flexibility in integrating new features such as roof characteristics extracted from aerial or satellite images. Although the model has been trained on simulated energy production data, the methodology can be extended to real-world data, potentially enhancing generalization by including additional features. A key aspect of this study is the automatic analysis of satellite or aerial images, as it significantly reduces the time needed to assess structural roof parameters that directly affect the energy production. To achieve this, a promptable segmentation model isolates the target roof pitch. Subsequently, a color-based analysis of the segmentation masks, automatically retrieved from the segmented roof, detects and remove obstacles. This leads to a more accurate estimate of the suitable roof area for photovoltaic installation. After removing obstacles from the segmented roof pitch, this study explores how to leverage the output of a Monocular Depth Estimation model to determine the roof tilt and orientation. Furthermore, an auxiliary graphic interface has been developed to support the estimation of roof parameters by connecting different processing steps and facilitating the evaluation and interpretation of the results. This research work demonstrates the potential of a multimodal approach for energy forecasting. By leveraging artificial intelligence to extract key roof parameters from a single image, the proposed methodology simplifies the assessment of roof suitability for photovoltaic deployment. The findings of this study support the widespread adoption of photovoltaic systems by enabling more efficient, scalable and automated site evaluations.