Modelling changing systems

Many machine learning methods rely on the assumption that future data will follow the same patterns as historical data. While this assumption may hold for some systems, many real-world systems evolve over time due to technological, operational, or societal changes. As a result, data distributions can drift, raising fundamental questions about when historical data can still be considered representative and when models should be updated.

This thesis explores how machine learning systems can remain reliable when the underlying data-generating processes change over time.

Several directions are possible:

1. Structural Changes in the Norwegian Power Market

Power companies rely heavily on forecasting prices and volumes to operate across different electricity markets, from day-ahead markets to balancing services such as mFRR. However, the Norwegian and European power markets have undergone several structural changes in recent years, including:

• Introduction of the FFR market (2022) 
• Introduction of 15-minute trading intervals (2025) 
• Temporary financial support mechanisms during high-price periods (2024–2025) 
• Introduction of mFRR EAM (2025) 

At the same time, the energy system itself is evolving. Renewable generation (especially wind and solar) is increasing, while consumption patterns are changing due to factors such as electric vehicle adoption and greater price awareness among consumers.

These developments raise an important question: Are historical datasets still representative of current and future market behavior?

In this project, the student will analyze the statistical properties of power market data to and evaluate which time scales are statistically stable for forecasting models. Methods for detecting distribution shifts and dataset representativeness will be explored (e.g., approaches discussed in https://arxiv.org/abs/2402.09134). Depending on the student’s interests, the project may also include forecasting experiments and uncertainty quantification.

2. When Should Machine Learning Models Be Updated?

Machine learning models used for condition monitoring of physical systems—such as process plants, ship thrusters, wind turbines, or power cables—are typically trained on data collected during stable operation shortly after installation.

However, physical systems gradually change due to wear, aging, maintenance, or environmental conditions. Over time, the signatures of “normal operation” evolve, meaning that models trained on early data may eventually become outdated.

This raises a key practical challenge:

When should models be retrained or updated to maintain accuracy without introducing instability or model drift?

In this project, the student will investigate strategies for updating machine learning models in evolving systems. The work may include analysis of operational datasets, evaluation of model degradation over time, and exploration of update policies in collaboration with an industrial partner.

Requirements

• Background in machine learning, statistics, or data science 
• Experience with Python and common ML libraries (e.g., PyTorch, TensorFlow, or scikit-learn) 
• Interest in time series analysis or statistical modeling 

Experience with energy systems, industrial data, or uncertainty quantification is an advantage but not required.