Mario Beykirch

M.Sc. Mario Beykirch

Stochastic optimization

+49 (0) 6151 16-21716
fax +49 (0) 6151 16-21712

S3|10 303
Landgraf-Georg-Str. 4
64283 Darmstadt

Research Interest

  • Probabilistic energy forecasting
  • Stochastic optimization for energy scheduling problems
  • Machine learning and data analysis in the energy domain
  • The interface of probabilistic forecasting and stochastic optimization

 

Valuation of Probabilistic Forecasts in Energy Scheduling Problems

The standard approach for energy scheduling under uncertainty is first to predict these uncertain parameters, such as consumption and prices, and use them in a stochastic optimization. The forecast can be deterministic, univariate probabilistic, or even multivariate probabilistic. Our research shows that not all scheduling problems require multivariate probabilistic forecasts. But while more complex forecast types can offer more information, they might be computationally costly, complex in implementation, and cause sampling errors. Hence, they should only be employed if they lead to better results in the optimization task at hand. Our current research focuses on which type of forecast is sufficient to achieve the best mathematically possible result in a specific energy optimization problem and how specific forecast properties influence optimization results.

EnEff:Stadt Campus Lichtwiese 2 (Subproject: Digital Twin)

The subproject Digital Twin aims to monitor energy flows, predict consumptions and optimize the energy system's operation. To this end, we first implemented a building-level live monitoring system of electricity, heat, and cold consumption. Second, we developed models that use machine learning methods to forecast the building's consumption. Third, we developed a mixed integer linear program that uses these forecasts to compute the optimal operation schedules for controllable consumers and producers.

To support the university's plan to save energy, we applied our research on energy forecasting to quantify a building's weather-independent energy savings. We developed and now maintain an automated tool that allows university employees to assess each building's energy savings on a daily basis.

Open theses

Supervisor: Mario Beykirch
Earliest start: immediately
Type: Master Thesis


An established approach to obtain operation schedules for energy systems under uncertainty is to first predict the uncertain residual demands and energy prices with a probabilistic forecast model that minimizes a statistical score such as the logarithmic score or the continuous ranked probability score. Then, in a second, separate step, an optimization problem is solved to determine the optimal schedule given the demand and price forecasts. However, this method is imperfect, particularly for typically finite, small sample sizes, where errors cannot be eliminated, and the statistical score does not necessarily accurately reflect the real task-loss after the optimization step. Consequently, a forecast selected (or even trained) using this task-loss may outperform another forecast with a comparable or worse statistical score in the actual optimization task.

Approaches to use the optimization result’s performance directly as a loss function in the model training exist, e.g., stochastic end-to-end learning, but are computationally expensive because the optimization must be solved in every training iteration. Task-specific score functions, conversely, only approximate the forecast’s performance in optimization tasks. They are simpler to compute since they allow us to assess the forecast quality without solving the actual optimization task. A task-specific score function can be estimated with a machine learning model and then employed as the loss function in the forecast training.

The goal of this thesis shall be to obtain and employ a task-specific score for a simplified case of a market schedule optimization given an uncertain residual demand and a power plant without start-up costs and to quantify the task utility improvement compared to statistical scores.

Short Bio

  • Since 2018: PhD student at EINS
  • 2015 – 2018: M.Sc. Energy Science and Engineering at TU Darmstadt and École polytechnique fédérale de Lausanne
  • 2011– 2015: B.Sc. Physics at TU Darmstadt and University of Bristol

Publications

The Value of Probabilistic Forecasts for Electricity Market Bidding and Scheduling Under Uncertainty

[Journal]
Mario Beykirch; Andreas Bott; Tim Janke; Florian Steinke :
The Value of Probabilistic Forecasts for Electricity Market Bidding and Scheduling Under Uncertainty.
In: IEEE Transactions on Power Systems , 2024

Bidding and Scheduling in Energy Markets: Which Probabilistic Forecast Do We Need?

[Conference]
Mario Beykirch; Tim Janke; Florian Steinke :
Bidding and Scheduling in Energy Markets: Which Probabilistic Forecast Do We Need?.
In: 17th International Conference on Probabilistic Methods Applied to Power Systems (PMAPS 2022), virtual Conference, 2022

Probabilistic Forecast Combination for Anomaly Detection in Building Heat Load Time Series

[Conference]
Mario Beykirch; Tim Janke; Imed Tayeche; Florian Steinke :
Probabilistic Forecast Combination for Anomaly Detection in Building Heat Load Time Series.
In: 2021 IEEE PES Innovative Smart Grid Technologies Conference Europe (ISGT-Europe), virtual Conference, 2021

Multi-modales Echtzeit Energiemonitoring - Multi-modales echtzeit Energiemonitoring als Basis eines digitalen Zwillings des Energiesystems des Campus Lichtwiese der Technischen Universität Darmstadt

[Conference]
Christopher Ripp; Mario Beykirch; Johannes Oltmanns; Florian Steinke :
Multi-modales Echtzeit Energiemonitoring - Multi-modales echtzeit Energiemonitoring als Basis eines digitalen Zwillings des Energiesystems des Campus Lichtwiese der Technischen Universität Darmstadt.
In: Digitalisieren - Sektoren koppeln - Flexibilisieren. Systemische Integration der Bioenergie und weiterer erneuerbarer Energien in Gebäuden & Quartieren, virtual Conference, 2020

Evaluation of Day-Ahead Electricity Price Predictions with Multi-Stage Stochastic Programs

[Conference]
Mario Beykirch; Tim Janke; Florian Steinke :
Evaluation of Day-Ahead Electricity Price Predictions with Multi-Stage Stochastic Programs.
In: 8th International Ruhr Energy Conference (INREC 2019), Essen, Germany, 2019

Bayesian Inference with MILP Dispatch Models for the Probabilistic Prediction of Power Plant Dispatch

[Conference]
Mario Beykirch; Tim Janke; Florian Steinke :
Bayesian Inference with MILP Dispatch Models for the Probabilistic Prediction of Power Plant Dispatch.
In: 16th International Conference on the European Energy Market (EEM), Ljubljana, Slovenia, 2019

Learning Dispatch Parameters of Thermal Power Plants from Observations

[Conference]
Mario Beykirch; Tim Janke; Florian Steinke :
Learning Dispatch Parameters of Thermal Power Plants from Observations.
In: 8th IEEE PES Innovative Smart Grid Technologies Conference Europe, Sarajevo, Bosnia-Herzegovina, 2018