4th Hagenberg Prescriptive Analytics Days 2025

10th of December 2025,  1  – 5 p.m., on-site Hagenberg / online

11th  of December 2025, 9 – 12 a.m., on-site Hagenberg / online

We are pleased to announce the 4th Hagenberg Prescriptive Analytics Days 2025 (HPAD25), which will take place on 10th - 11th of December 2025 at the Software Competence Center Hagenberg (SCCH).

SCOPE

Workshop HPAD25 focuses on the AI of time series: current trends in data and software scientific methods for analyzing and predicting time series data. The event will feature 10 presentations by leading researchers and practitioners in the field, addressing a wide range of topics at the intersection of AI, data science, and time series analysis, including:
• State-of-the-art forecasting models (Selective State Space Models, Foundation Models, ...)
• Explainable AI on time series methods (feature importance, local surrogate modeling)
• Physics-informed approaches (integration of domain-specific knowledge)
• Time series algorithms (anomaly detection and prediction, prescriptive maintenance)
• Novel approaches for processing time series data

REGISTRATION:

until 1st of December to servicecenter@scch.at<

AFTERNOON AGENDA, 10th of December 2025

until 12:45 p.m.
ARRIVAL & WELCOME COFFEE (Community Deck, 4th floor)

1 – 1:15 p.m.
Introduction HPAD and Time Series Research at SCCH
Dr. Florian Sobieczky, Senior Researcher Data Science (SCCH)
DI DI Dr. Bernhard Heinzl, Senior Researcher Data Science (SCCH)

1:15 – 1:55 p.m.
Time series forecasting for electrical demand
Valerio Piersimoni, MSc, Cellforce

2 – 2:40 p.m.
From Molecules to Minds: The Evolution of Health AI
Prof. Wolfgang Maaß, Scientific Director, DFKI

2:40 – 3 p.m.
Coffeebreak – Cookies and Demonstrators

3 – 3:40 p.m.
Adaptive Analytic Prescriptions: Integrating Models, Data, and Optimization in Dynamic Environments
Bernhard Werth, MSc, Hagenberg University of Applied Sciences FHOÖ

3:45 – 4:25 p.m.
Time and space combined: Hybrid deep learning models for predicting solar energy production
Moritz Sondheimer, Ph.D., NTUST IMLab

4:30 – 5:10 p.m.
Anomaly detection and Feature Extraction of Robotic Motion via Learning Entropy of Explainable IPLNA
doc. Ivo Bukovský, research associate professor at the Faculty of Science, University of South Bohemia in Ceske Budejovice

 

MORNING AGENDA, 11th of December 2025

9 – 9:40 a.m.
SaC, an alternative to Python for the performance critical part of novel ML approaches?
Prof. Sven-Bodo Scholz, Director of education, Radboud University

9:45 – 10:25 a.m.
TiRex: Zero-Shot Forecasting across Long and Short Horizons
Levente Zólyomi, MSc, Researcher, NXAI

10:25 – 10:45 a.m.
Coffeebreak – Cookies and Demonstrators

10:45 – 11:00 a.m.
TBA
Alfredo Lopez, Ph.D, SCCH

11:05 – 11:20 a.m.
Towards Data-Driven Discovery of Interpretable Differential Equations
David Jödicke, MSc, SCCH

11:25 – 11:55 a.m.
TBA
Janneke Verbeek, MSc, Data Science Department of Radboud University
Anna-Christina Glock, MSc, SCCH

11:55 – 12:00 p.m.
Recap
Dr. Florian Sobieczky, SCCH

ORGANIZATION:

Florian Sobieczky, florian.sobieczky@scch.at

 

HPAD25 – Book of Abstracts

Prof. Wolfgang Maaß

Title: From Molecules to Minds: The Evolution of Health AI

The integration of artificial intelligence into health research is  transforming how we understand, predict, and intervene in human biology. This talk traces a translational pathway from molecular analysis to  cognitive modeling by spanning metabolic systems, genomics, and medical reasoning. Beginning with AI-based detection of hormonal manipulation in  anti-doping research, we demonstrate how machine learning can capture  the dynamics of metabolic regulation and identify subtle deviations in  biomarkers such as erythropoietin (EPO) and testosterone. Extending this  approach to genomics, we present pipelines for polygenic risk modeling  (PGS) that transform genome-wide variation into clinically meaningful  predictors of disease susceptibility. Building on these foundations, we  explore the use of large language models (LLMs) to synthesize and  interpret the rapidly growing body of medical literature. A method is  presented that mitigate hallucinations by scoring hallucinations and  faithfulness. These efforts define a framework for trustworthy Health  AI: systems that reason across molecular, genomic, and textual evidence  while maintaining transparency and verifiability. The talk concludes with a forward-looking view toward Cognitive Health AI, in which artificial mental models integrate genomics, biomarkers, and  biomechanical simulations to create self-explaining patient avatars.

Wolfgang Maaß, with degrees from RWTH Aachen University and Saarland University, including a Ph.D. in Computer Science, is a distinguished professor at Saarland University specializing in Business Informatics and Computer Science. He also serves as a Senior Data Science Advisor at the National Cancer Institute (NCI) in the United States. As the scientific director at the German Research Center for Artificial Intelligence (DFKI), Maaß focuses on research in data economics and applying artificial intelligence across various industries such as manufacturing and health.

 

Prof. Sven-Bodo Scholz

Title: SaC, an alternative to Python for the performance critical part of novel ML approaches?

Python has taken center stage when it comes to implementing ML algorithms.The languages convinces through its flexibility and the availability of many libraries.Libraries such as NumPy or CuPy ultimately enable very good performance which helps bridging the gap between Python's flexibility and the sophisticated parallel code needed for performance. In this talk, we make the case for Single Assignment C (SaC) as an alternative forcomputational kernels where performance across a wide range of parallel architectures is needed but no highly tuned libraries are available yet. In contrast to Python, SaC focusses on arrays as main data structure, builds on a type system that is aware of array shapes, and generates high-performance codes for multi-core machine, GPUs and clusters alike. As it turns out, rank-polymorphism, a key feature of SaC, does not only allow for more generic programs but, surprisingly, it also helps when optimising across different parallel platforms.

Sven-Bodo Scholz' research is driven by the desire to bridge the gap between high-productivity programming tools and high-performance computing on heterogeneous many-core systems by means of compilation or code generation technology. His contributions are typically validated through readily available software, e.g. the software eco-system around the functional array programming language SaC (https://www.sac-home.org).

 

Levente Zólyomi, MSc

Title: TiRex: Zero-Shot Forecasting Across Long and Short Horizons

TiRex is a lightweight forecasting model based on the xLSTM architecture, enabling zero-shot prediction across short and long time horizons. It employs Contiguous Patch Masking and Masked Inference to efficiently capture temporal dependencies and uncertainty without autoregressive feedback. Trained on a diverse mix of real and synthetic data, TiRex combines strong generalization with exceptional computational efficiency. With only 35 million parameters, it achieves state-of-the-art zero-shot performance while offering fast inference and low energy use. Its internal representations also transfer effectively to downstream tasks such as time-series classification, highlighting xLSTM’s potential for scalable and sustainable time-series modeling.

Levente is a machine learning researcher with an international academic background, holding degrees from the Netherlands and Finland. He is currently a PhD researcher at NXAI while pursuing his doctoral studies at Johannes Kepler University. Before joining NXAI, he worked at Microsoft Gaming’s King AI Labs, where he focused on self-supervised learning and the analysis of neural  architectures. His research centers on advanced recurrent models – particularly xLSTM – and their application to time-series foundation models such as TiRex.

 

doc. Ivo Bukovský

Title: Anomaly detection and Feature Extraction of Robotic Motion via Learning Entropy of Explainable IPLNA

This talk presents an explainable approach to robotic motion analysis using Learning Entropy (LE) applied to in-parameter-linear neural architectures (IPLNAs), such as Higher-Order Neural Units (HONUs). LE tracks dynamic changes in neural weights to detect anomalies and extract informative features in real time, making the method particularly suitable for nonstationary environments. Suppressing data redundancy through input decorrelation enhances the uniqueness and interpretability of the learned weights. Combined with the convexity of IPLNAs, this results in a unique, structurally transparent, and post hoc physically explainable solution.

Ivo Bukovsky (Senior Member, IEEE) received his Ph.D. degree in control and systems engineering from Czech Technical University in Prague, Prague, Czechia Republic, in 2007. He is with the Department of Computer Science, Faculty of Science, University of South Bohemia České Budějovice, České Budějovice, Czech Republic. He was a short-term Visiting Researcher with the University of Saskatchewan, Saskatoon, SK, Canada, in 2003, and with the University of Manitoba, Winnipeg, MB, Canada, in 2010. Since 2009, he has been cooperating with Tohoku University, Sendai, Japan. His research interests include novelty detection, multiscale analysis approaches, dynamical systems and data analysis, adaptive control with in-parameter-linear nonlinear neural architectures, and novel information theory through machine learning. Dr. Bukovsky has served as an Associate Editor for IEEE Transactions on Neural Networks and Learning Systems, IEEE Transactions on Neural Networks and Learning Systems, and IEEE Transactions on Cybernetics, as well as PC for numerous IEEE conferences. Very recently, Ivo has initiated the Centre of Explainable Learning Systems (CELS) at USB.

 

Valerio Piersimoni, MSc

Title: Time series forecasting for electrical demand

Cellforce Group GmbH, Porsche Group, Germany and Czech Technical University in Prague, Faculty of Mechanical Engineering

Rising electrification and variable renewable generation place high demands on accurate and reliable load forecasts. This talk presents four case studies of electricity demand prediction on a real 5-minute dataset (CAISO/NREL) with exogenous variables (weather, PV/wind output) and a focus on uncertainty quantification. We combine LSTM architectures with Bayesian learning (SVI with Trace ELBO) to deliver not only point forecasts but also confidence intervals to support system operators and resource planners. We demonstrate one-step and 12-step-ahead forecasts (including a 15-minute aggregation variant) and summarize achieved metrics. We also discuss data requirements and practices that prevent information leakage between training and test sets.

Valerio Piersimoni works in Cellforce Group GmbH, Porsche Group, Germany, as process engineer.

 

Moritz Sontheimer, Ph.D.

CTO bei ecoOops and currently on leave from NTUST IMLab

Title: Time and space combined: Hybrid deep learning models for predicting solar energy production  

Accurate prediction of photovoltaic performance is a key challenge in the integration of renewable energies. While classic time series models such as ARIMA or Prophet are based on statistical assumptions, modern deep learning architectures allow adaptive modeling of nonlinear temporal and spatial dependencies. This article presents a hybrid approach that combines three-dimensional convolutional neural networks (3D-CNNs) for extracting spatiotemporal features with long short-term memory (LSTM) networks for capturing long-term dynamics. Using real sensor data from a photovoltaic system, both the forecast quality and the computational complexity of the models are compared. Finally, we discuss how model architecture and feature selection affect the generalizability and interpretability of the predictions, as well as the perspectives opened up by current research in the field of explainable AI and efficiency metrics for time series.

Moritz Sontheimer is an AI engineer and PhD student specializing in machine learning and cloud-based solutions. His work focuses on computer vision, natural language processing, and optimization methods, with experience in projects such as license plate detection and anomaly detection. At Frontier.cool, he develops AI-driven software using AWS and Docker, and his academic research explores topics like photovoltaic power prediction and text mining.

 

Bernhard Werth, MSc

Titel: Adaptive Analytic Prescriptions: Integrating Models, Data, and Optimization in Dynamic Environments

Prescriptive analytics systems increasingly operate in environments where both the optimization problem and its underlying models evolve over time. This talk examines how recommended actions can remain effective as new information arrives, data distributions shift, and models are retrained. The focus lies on the interplay between learning components and optimizers, and how their interaction enables adaptive decision-making in dynamic settings. Examples from dynamic production optimization and warehouse operations illustrate how streaming observations can influence both the optimizer and the models guiding it. While such feedback can enhance responsiveness and robustness, it can also lead to self-reinforcing dynamics where optimizers and models amplify their own misconceptions. Understanding these interactions is essential for developing prescriptive analytics that remain trustworthy in changing environments.

Bernhard Werth is researcher from Austria with the University of Applied Sciences (FHOÖ) and the Heuristic and Evolutionary Algorithms Laboratory. His experience is in heuristic and dynamic combinatoric optimization, symbolic regression and simulation-based optimization."