Abstracts Track 2024

Area 1 - Creating and Maintaining Energy Islands

Nr:	32
Title:	Segment Anything Model by Meta for PV Systems Detection: Impact of Image Resolution
Authors:	Aleksandre Kandelaki, Jessy Matar, Florian Kotthoff and Markus Duchon
Abstract:	Motivation and Research Question In the current energy landscape, integrating renewable sources seamlessly into power generation is crucial, particularly with the rise in solar power adoption. Accurate data on PV systems helps in addressing questions about socio-economic factors influencing PV adoption, regions lacking PV systems, and estimating the potential of PV power generation for specific regions. However, manual registration of photovoltaic installations by owners poses disadvantages, such as increased effort and errors happening upon data entry. To address this, automatic detection using aerial imagery, especially with deep learning methods, has proven successful. The success of PV Systems detection relies on image resolution. High resolution images promise high performance, but they are costly and often not available. In our work, we compare the performance of PV system segmentation on high-resolution (original dataset of 20cm spatial resolution covering Area of Munich) and low-resolution imagery (Fine-grained Downsampling) using Meta's SAM model, aiming to identify the minimum image quality for precise photovoltaic system detection. The research is valuable for government agencies, energy companies, and researchers, providing insights into the link between image resolution and detection performance. Findings and Future Work Other than gaining valuable insight into the resolution and detection quality trade-off, this work can also be thought of as a good starting point in popularising foundation models such as SAM in remote sensing downstream tasks (PV Systems Detection). We demonstrated that the model SAM can achieve attractive results and can be compared to state-of-the-art Deep Learning frameworks in PV detection. Downsampling images to resolutions similar to various satellite imagery revealed challenges in accurately identifying PV installations, particularly smaller residential PV systems, due to the limitations inherent in working with such low-resolution data. The key takeaway is the significance of a thoughtful selection process for resolution, considering factors such as cost, performance, and the specific needs of the application. Improvements could be addressed by meaningful extension of the used dataset with more diverse, heterogeneous samples, encompassing comprehensive features related to PV systems. This research is based on the Bachelor's Thesis of AK, which is included as an attachment.

Area 2 - Energy-Aware Systems and Technologies

Nr:	29
Title:	Smart Energy Forecasting: An in-Depth Study on Forecasting Methods for Electric-Thermal Storage Systems
Authors:	Cristian Sebastian Cubides Herrera, Jan Mayer, Jessy Matar and Markus Duchon
Abstract:	Motivation and Research Question: Smart grid developments have recently gained significant attention due to their potential to optimise energy consumption and reduce environmental impacts. For this reason, it is crucial to forecast future state conditions such as power, temperatures, heat, or SOC (state-of-charge) to make the most accurate and suitable control decision depending on the context and need. Since many processes are hard to model, the forecasting task can be executed by exploiting the advantages of machine learning algorithms such as LSTM, transformer, Autoformer, or CNN. In the context of the RESONANCE Horizon Europe project, in Fortiss EnergyLabs in Munich, we investigate a smart electric-thermal storage (SETS), which consists of a device that converts excess electricity into thermal energy. The system is equipped with 10 temperature sensors and a recently incorporated power measurement point. Implementing an optimal control module of a smart device, which is responsible for managing a resource, includes its profile and flexibility forecasts, hence the importance of achieving a high accuracy forecasting of the parameters. The central question from this scenario is: which forecasting method is the most effective for this use case, and to what extent can past data with missing features (such as power) be used to augment the current training data? Findings and Future Work: We have established a precise predictive model for the SETS, ensuring consistent and reliable temperature and power time series one-hour forecasts, establishing an optimal looking-back length of 10 minutes. In the process, we benchmarked state-of-the-art ML architectures, by which LSTM outperformed the other structures in terms of test error and computational cost. Additionally, we used already saved data without power information to perform data augmentation, attenuating the common oscillations present on the complete but small dataset even when the old dataset did not have any recorded power data. Finally, leveraging this enhanced predictive capability, we can now calculate the SOC of the SETS device, aligning its operation with the specific needs of users or central resource administrators. As a future step, we aim to implement an automatic control module that utilises these predictions to optimise energy management further.