Towards Resilient Energy Systems - Probabilistic Deep Learning for Forecasting and Control under Uncertainty

by Marcel Arpogaus
Doctoral thesis
Date of Examination:2025-09-29
Date of issue:2025-10-09
Advisor:Prof. Dr. Gunnar Schubert
Referee:Prof. Dr. Thomas Kneib
Referee:Prof. Dr. Marcus Baum
Persistent Address: http://dx.doi.org/10.53846/goediss-11536

 

 

Files in this item

Name:20251002--Dissertation_Marcel_Arpogaus-online.pdf
Size:23.2Mb
Format:PDF

The following license files are associated with this item:

Abstract

English

The integration of variable renewable energy source (RES) and the electrification of end-use sectors introduce significant uncertainty into energy systems, challenging traditional deterministic modeling and operational paradigms. This cumulative dissertation addresses these challenges by exploring, developing and applying probabilistic deep learning methods to improve uncertainty quantification and management in power grids and related domains. This research advances novel probabilistic deep learning techniques, focusing on deep conditional transformation models (CTMs) and their extensions. Key contributions include: (i) Demonstrating that deep CTMs, particularly Bernstein normalizing flows (BNFs), deliver accurate and robust probabilistic short-term load forecasts at the low voltage (LV) level, outperforming conventional approaches, especially under data scarcity. (ii) Proposing a real-time grid operation framework that combines probabilistic forecasts with a graph neural network (GNN) based control model, enabling proactive management of distribution grids under uncertainty. (iii) Extending autoregressive deep CTMs to probabilistic indoor temperature forecasting, providing uncertainty estimates for heating, ventilation and air conditioning (HVAC) control through stochastic optimization strategies. (iv) Introducing hybrid Bernstein normalizing flows (HBNFs), a novel hybrid model class that integrates interpretable marginal modeling through CTMs with flexible multivariate modeling using neural network (NN) based autoregressive normalizing flows (NFs), addressing the trade-off between interpretability and expressiveness in multivariate density estimation. Together, these contributions offer advanced methods and tools for uncertainty quantification and control in energy systems. By advancing probabilistic forecasting and decision-making, this work supports the development of more resilient, efficient and adaptable energy grids. The findings underscore the practical value of distributional modeling and the potential of hybrid statistical and deep learning approaches for addressing the limited interpretability arising from the black-box nature of pure deep learning methods.
Keywords: Normalizing Flows; Probabilistic Regression; Deep Learning; Low-Voltage; Probabilistic Load Forecasting; Probabilistic Deep Learning; Distributional Regression; Transformation Models; Bernstein Polynomials; Conditional density estimation; Indoor temperature; Autoregressive Transformation Models; Machine learning; Multivariate Density Estimation; Multivariate Conditional Transformation Models; Resilient Energy Systems; Grid Operation Management; Renewable Energy Sources; Electrical Vehicles
 
Statistik