Gleb V. Prokhorskii has followed the PhD Programme at Faculty of Engineering and Science, specialisation in Specialisation in Engineering Sciences, scientific field Renewable Energy.
- Trial lecture starts at 10:15
- Public defence starts at 12:15
Title of trial lecture: "Integrating First Principles and Data-Driven Methods for Thermal Modeling of Combined Heat and Power Plants: Status and Development Trends"
Read the thesis in the Norwegian Research Information Repository
Disputation chair: Paul Ragnar Svennevig
Assessment committee:
- First opponent: Professor Raphael Lechner, OTH Amberg-Weiden, Germany
- Second opponent: Professor Lars O. Nord, NTNU, Norge
- Chair of assessment committee: Associate Professor Muhammad Faisal Aftab, UiA
Supervisors in the doctoral work:
- Main supervisor: Professor Souman Rudra, UiA
- Co-supervisor: Professor Markus Preissinger, Vorarlberg University of Applied Sciences, Austria
Summary of thesis
Thermal and biomass power plants are complex systems in which components do notwork independently involving nonlinear processes. Traditional maintenance methods, such as corrective (run-to-failure) or preventive (calendar-based) approaches, are often inadequate for these environments. Predictive maintenance, which estimates the remaining useful life using data, expert knowledge, physical models, or a combination of these, has shown great promise. The main modeling approaches in this context are data-driven (DD), knowledge-based (KB), and physics-based (PB) models. Knowledge-based predictions rely on the experience of operational staff; however, this expertise is rarely centralized and can be lost with personnel changes and retirements. Physics-based approaches can be accurate but are usually difficult to apply on a full scale and often require simplifying assumptions. Data-driven machine learning (ML) models are well-suited for thermal power plant applications, as they can detect specific issues and identify the affected equipment directly from sensor data.
One major challenge is selecting the right prediction model from the numerous ML techniques available. To address this, we benchmarked four algorithms with varied complexity and learning mechanisms. We optimized hyperparameters for each model (such as regularization for Lasso, the number of trees for random forest, network architecture, or the number of neighbors for KNN) to achieve the best results. Models were built using two of the most general and simplified feature sets, representing core physical processes, and compared under two training strategies: static training, where the model is trained once on a historical dataset, and dynamic (rolling-window) training, where the model is regularly retrained on the latest data.
We also examined whether clustering could help the static training approach to reach the accuracy of dynamic training. By modeling both knowledge-based and ML-based clustering, we identified operational regimes where predictions differ and found that certain plant states are more challenging to model accurately. This insight helps target further analysis and improvement.
A second major challenge is the high dimensionality of modern sensor data. Thermal plants may have thousands of features, such as temperatures, pressures, and flow rates, which increase the computational load and the risk of overfitting. Careful feature selection is essential: reducing dimensionality improves both model accuracy and interpretability for engineers. Choosing the right features is not trivial and requires balancing expert knowledge with automated, data-driven techniques to ensure important signals are included. In this work, we developed models using a variety of feature selection strategies: simple physics-based sets, pure expert knowledge, fully data-driven selection, and hybrid approaches. This combination aimedto maximize accuracy, generalizability, and interpretability.
Despite advances in related areas, a comprehensive predictive modeling study for thermal power plant monitoring remains lacking. Most prior work addresses anomaly detection, forecasting, or fault diagnosis separately.
While the effectiveness of different feature selection methods, especially those incorporating expert knowledge, has only recently been evaluated on real datasets, this thesis advances the field by systematically benchmarking and comparing such approaches in operational thermal power plants. Furthermore, although research on models that can adapt to changing plant conditions and varying operational regimes remains limited, the methodologies developed in this work contribute to narrowing this gap. Specifically, the thesis demonstrates that combining expert knowledge and data-driven feature selection, along with a dynamic training approach and clustering, yields predictive maintenance models that are both adaptable and effective in managing the scale and complexity of modern power plants.
The effectiveness of different feature selection methods, particularly the role of expert knowledge, has only recently been evaluated on real datasets. Similarly, research on models that adapt to changing plant conditions and account for different operating regimes is still limited. This gap highlights the need for additional research into predictive maintenance models that can effectively address the scale and complexity of modern power plants.
To demonstrate the practical value of our approach, we developed a cloud-based live steam prediction system for one power plant and then adapted it to a second facility. In this setup, the model operates on a cloud platform that continuously receives sensor data in real-time. It generates ongoing predictions of key variables such as live steam output or efficiency. When actual sensor readings deviate from predicted values, the system flags potential anomalies. During such periods, we analyze model weights to determine which variables contributed most to the deviation, helping to identify root causes.
Thus, by integrating expert knowledge with advanced data-driven techniques, the solutions developed here provide a practical pathway to more intelligent and reliable predictive maintenance in complex industrial environments. The demonstrated adaptability and transparency of the models lay the groundwork for broader deployment in real-world plants, supporting not only improved performance and reduced downtime but also empowering operators to make informed decisions based on interpretable, actionable insights.
What to do as an online audience member
The disputation is open to the public. To follow the trial lecture and the public defence online, register on Zoom.
We ask online audience members to join no earlier than 10 minutes in advance. After these times, you can leave and rejoin the meeting at any time.
Opponent ex auditorio:
Deadline for the public to pose questions is during the break between the two opponents. Questions ex auditorio can be submitted to Kristine Reinfjord.
Contact person
- kristine.reinfjord@uia.no
- Phone
- +47 37 23 30 14