Abstract
In recent years, there has been increasing focus on analyzing and predicting waterresource management to optimize urban water supply. Advances in machine learning
and AI have significantly inspired the development of smart water systems. This thesis explores the effectiveness of learning independent graphs in a federated learning
environment, specifically comparing the performance of a federated model with a centralized model in predicting water demand. The study aims to address the challenges of
managing distributed data using modern AI tools within a federated framework and to
improve water demand networks by leveraging advanced machine learning techniques.
The research employs a FedAVG setup with Spatio-Temporal Graph Neural Networks,
implemented on a test bench utilizing Apache Airflow, MlFlow and MinIO, allowing for
extensive hyperparameter tuning and uniform experiment testing. Tests were realized
to compare the performance of centralized clients against one federated learning setup
comprising multiple clients. Results indicate that the centralized clients outperform the
federated learning method so far. A significant observation was the slow learning rate
of the graph convolutional model in centralized and federated learning. Analyzing the
given data suggests that the model is not powerful enough for the complexity of data
given. The findings underscore the challenges and potential improvements required for
federated learning models, specifically in terms of computational efficiency and scalability. Enhancing the complexity of the model is crucial for improving learning rates and
overall performance in federated environments, ultimately benefiting the accuracy and
reliability of water demand prediction networks.
| Date of Award | 2024 |
|---|---|
| Original language | English (American) |
| Supervisor | Viktoria Dorfer (Supervisor) & Alexander Valentinitsch (Supervisor) |