ITEN Intelligente Thermal Energie Netzwerke

• Forschungsstelle: METAS
• Projektnummer: F-5135.30169
• Projekttitel Englisch: ITEN Intelligente Thermal Energie Netzwerke

Schlüsselwörter (Englisch)

energy consumption, heating networks, electricity data, graph signal processing, graph neural networks (GNN), energy consumption forecasting

Kurzbeschreibung (Englisch)

The project has so far covered the evaluation of deep learning-based algorithms on synthetic data and simulations for a district heating network using open-source libraries along with real-world datasets for Internet of Things (IoT) air-pollution monitoring platforms. Two approaches were used: one purely data-driven and one using physics-enhanced deep learning.

The general methodology within the project is based on graph representation learning (graph signal processing and graph neural networks), alongside the integration of physics-based formulas. This selected methodology, focusing on graph representation learning, has been chosen for its effectiveness in comparison to traditional machine learning and deep learning methods, particularly in the context of sensor networks. The success control involves ongoing comparison and evaluation of the performance of graph representation learning against alternative methods, ensuring that the chosen approach consistently demonstrates its efficacy in addressing the unique challenges posed by sensor networks within the project.

The algorithms were tested on a simulated dataset of a district heating network serving four customers, and a real-world dataset from electricity meters in Bern, and their performance in forecasting energy consumption was evaluated. The algorithms are able to capture spatial correlations within the sensor network, and they are able to successfully forecast energy consumption for the next 24-hours, based on the previous week's data. 

Projektziele (Englisch)

• Establishment, testing, and evaluation of a digital model of a district heating network; 

The promising results observed in the initial analysis of simulated data highlight the substantial potential of our proposed method for enhancing district heating network sensing. Subsequent evaluation will involve assessing the model's performance on real-world data to validate its practical applicability and efficacy. Our initial findings indicate promising potential for the application of GNN in electricity networks.

• Anomaly Detection and Demand Forecasting:

We aim to forecast heat demand by leveraging both spatial and temporal correlations. Our forecasting method not only accurately predicts consumption but also has the capability to detect anomalies in consumption patterns.  

• Domain Adaptation:

A critical addition to the project's roadmap involves the exploration and implementation of domain adaptation strategies that will help us to transfer the developed methods to different operating and environmental conditions, application scenarios and datasets, ensuring its efficacy in real-world applications beyond its original training domain.

• Uncertainty Quantification:

Typically, deep learning models exhibit excessive confidence in their predictions, which is often undesirable. To address this issue, our objective is to integrate uncertainty and confidence intervals into our estimation and prediction processes allowing us to capture the inherent uncertainty associated with each prediction, ultimately enhancing the reliability of our model outputs.

• Theoretical Contributions in Graph Signal Processing:

The project will actively pursue theoretical contributions to the field of graph signal processing. This involves delving into novel concepts, algorithms, or methodologies that significantly contribute to the theoretical foundations of graph signal processing.