Project Description

Basic Research

The project will mainly focus on basic research in machine learning models for time series and graphs. One of the main goal will be to push the boundaries in the field of relational deep learning by:

• Creating innovative tools (novel architectures, training strategies, etc,…) for processing spatio-temporal data, e.g., multiple time series whose relationships are described by a graph.
• Enhance the capabilities of existing deep-learning models by gaining theoretical and practical insights.

The research activities are divided into four Work Packages.

1. Randomized architectures to handle big data

The research will focus on developing randomized architectures, such as those from Reservoir Computing, to generate informative spatio-temporal representations without the need for traditional training or supervision.

By using suitable randomized techniques, it will be possible to improve the scalability of large spatio-temporal models without compromising their performance.

2. Multi-scale representations with graph coarsening

The objective is to create multi-scale representations with graph pooling (a procedure to generate smaller graphs that carry the original information) to manage the complexity of spatio-temporal models. New graph pooling techniques suitable for spatio-temporal data will be developed and used to enhance the performance on tasks of interest (e.g., forecasting), to identify underlying factors in the system, handle missing data, and integrate multi-resolution data from various sources.

Possible applications include adjusting the granularity of data from the detailed low-voltage Distribution System Operator (DSO) grid to the broader Transmission System Operator (TSO) grid.

3. Uncertainty quantification

The goal is to model the uncertainty in deep learning models for spatio-temporal data by means of Bayesian and frequentist approaches. This involves modifying existing deterministic models to include probabilistic components and extending techniques for generating confidence intervals to spatio-temporal data by addressing the challenge of capturing both spatial and temporal dependencies.

4. Interpretability

Develop new techniques that allow for a human-understandable explanation of the model's output, thus aiding in systematic pattern discovery within the data. Due to the complex and irregular structure of spatio-temporal data, existing interpretability tools for GNNs are not suitable. The goal is to extend current approaches to spatio-temporal models and to develop new ones based on probabilistic frameworks.

Applications

The methodologies developed in the basic research WPs will be utilized to tackle the challenges of energy systems. These systems present complexities that traditional models frequently struggle to address, and the application of advanced relational deep learning techniques aims to provide more effective solutions. The work will be done in collaboration with industrial partners from the energy sector located in Northern Norway and focuses on three main tasks.

1. Enhanced Load Forecasting: RELAY will integrate grid topology into forecasting models, enabling precise predictions of energy demand and production from renewable sources, which can improve energy distribution and resource planning.

2. Dynamic Power Flow Optimization: The project will develop flexible power solvers that can account for historical data and make predictions about future energy flows. This will be crucial for real-time grid adjustments, ensuring balance and optimal resource allocation.

3. Effective Outage Localization: By representing power grids as dynamic graphs, RELAY will identify risk areas, enabling faster response to outages, a vital aspect in rural regions with weaker infrastructures.