Description

  • Participants of the course are allocated individual research topics to work over the seminar course. Groups of two are possible.
  • A kickoff session will be conducted to introduce the course and topics.
  • Attendance is only mandatory on the proposal and final presentation dates.
  • Regular follow-up meetings with the participants and supervisor are held to aid the participants with the seminar tasks.

The seminar tries to mimic the event flow of a conference publication. Therefore, the deliverables are the following:

  • a scientific paper (submitted as the seminar report).
  • a conference-style final presentation.

More details on ECTS, learning outcomes, etc is offered at the kickoff session and later on moodle.

Topic Allocation

To apply for topics, the top 3 topics (please provide the acronyms) and the current transcript of records must be sent to manuel.katholnigg@tum.de. Applications will only be accepted from 20.03.26. Information on deadlines can be found in the next section. Note that you can also apply for a topic as a group of two students!

Process and Time Line (Summer Semester 2026)

Topics are chosen in two blocks. Students who want to be sure of getting a fixed place in the seminar before the start of the semester must apply by 01.04. (12 noon). These students will be informed on 06.04. if they get a place in the seminar and which topic they will be working on.

Students who want to apply later can do so until 16.04. (7pm) and will then be informed on 20.04. whether they could still get a place in the seminar.

The most important dates for the seminar are summarized below:

Registration period: 01.03. - 23.04.

First deadline for submitting the top 3 topic requests: 01.04. (12 noon).

Assignment of topics block 1: 06.04.

Kick-off (introduction to the seminar, presentation of the topics still available): 15.04., 3:00 pm in N3815

For the kick-off, both in-person and online participation are possible. All presentation dates after the kick-off are strictly in-person, and online participation is not possible for these.

Kick-Off

Join: https://teams.microsoft.com/meet/343850788102102?p=pjIeS51wXecUlH4Enl

Meeting ID: 343 850 788 102 102

Passcode: Bi9iH97v

2. deadline for submission of the top 3 topic requests: 16.04. (7:00 pm).

Assignment of topics block 2: 20.04.

Proposal Presentations: 06.05 & 13.05, 3:00 pm in N3815

Final Presentations: 08.07. & 15.07., 3:00 pm in N3815

Topics

AcronymTopic descriptionRequired skillsGroup work possible (group of two)ContactTime Topic Added

FLEX_PROD 

Using Flexibility of Production Systems for Energy Management

Research method: Prototyping

Research questions:

  • How can the flexibility of production systems (e.g., running machines slower or faster) be used to improve energy management (e.g., reducing cost by using more locally generated solar power or adjusting consumption to dynamic tariffs)
  • How can corresponding optimization problems be defined and solved?
  • How can different optimization methods be benchmarked? 

Possible approach:

  • Understand state-of-the-art based on literature
  • Define optimization scenarios (e.g., based on production use cases described in the literature)
  • Develop efficient Python or Julia code to implement an optimization method
  • Benchmark optimization method using baseline scenarios


  • Python/ Julia programming skills
No

Christoph Goebel

christoph.goebel@tum.de

2025-09

DIST_GRID_GEN_MAPS

Map-Based Distribution Grid Model Generation Methods

Research method: Prototyping

Research questions:

  • How can realistic distribution grid models be synthesized based on map data?
  • How can synthesized distribution grid models be evaluated?

Possible approach

  • Understand state-of-the-art based on literature
  • Develop efficient Python or Julia code to implement a distribution grid generation method
  • Evaluate method using actual data


  • Python/ Julia programming skills
  • Understanding of distribution grid modelling
Yes

Christoph Goebel

christoph.goebel@tum.de



2025-09

LLM_EMS

LLM-Assisted Tariff Parsing for Local Energy Management Systems 

Research method: Prototyping

Research question:

  • How effectively can Large Language Models (LLMs) parse complex, multi-page industrial electricity tariff PDFs into functional Energy Management System (EMS) logic? 

  • How can this translation process be executed entirely locally to comply with legal constraints regarding cloud uploads?

Possible approach:

  • Investigate LLM-assisted methods for parsing and extracting rules from complex tariff documents.

  • Develop a fully local open-source pipeline that accepts a tariff PDF and generates a functional EMS configuration without data leaving the plant.

  • Benchmark the rule extraction accuracy and performance across various types of electricity tariffs.

  • Python programming skills
Yes

Manuel Katholnigg

manuel.katholnigg@tum.de


2026-03

ECON_FC

Decision-Aware Forecasting for Optimal Energy System Operation

Research method: Prototyping

Research question:

  • How can forecasting models for PV generation or electricity load be optimized for economic performance rather than purely statistical accuracy?

  • What loss functions or training frameworks can directly incorporate the economic consequences of forecast errors?

  • How do these economically-optimized forecasts perform in a simulated operational environment (e.g., battery scheduling) compared to standard forecasts?

Possible approach

  • Understand the state-of-the-art in decision-aware forecasting based on literature.

  • Define an economic objective for a specific application (e.g., minimizing electricity costs for a household with a battery and PV system).

  • Develop and implement a forecasting model (e.g., based on an ANN) using a custom, economically-derived loss function.

  • Evaluate and benchmark the model's resulting economic performance against models trained with standard statistical metrics (e.g., MSE, MAE).

 

  • Python programming skills
Yes

Manuel Katholnigg

manuel.katholnigg@tum.de


2026-03

MARL_COOP_OPT

Cooperative Control of Energy Systems using Multi-Agent Reinforcement Learning

Research method: Prototyping

Research question:

  • How can a cooperative multi-agent reinforcement learning (MARL) framework be designed to simultaneously optimize forecasting models and battery dispatch strategies?

  • What agent configurations and reward structures are most effective for achieving system-wide economic goals?

  • How does this cooperative MARL approach compare to decentralized or traditional, separate optimization methods?

Possible approach

  • Understand the state-of-the-art in MARL for energy systems and forecast-aware control based on literature.

  • Develop a simulation environment for a system including PV/load profiles, a battery, and market prices.

  • Implement and train a cooperative MARL algorithm where agents (e.g., a forecast agent and a battery control agent) work together.

  • Benchmark the MARL approach against relevant baseline strategies.

 

  • Python programming skills
Yes

Manuel Katholnigg

manuel.katholnigg@tum.de


2026-03

BENCH_WIND

State-of-the-Art Machine Learning and Deep Learning Models for Day-Ahead Wind Power Forecasting

Research method: Prototyping

Research questions:

  • How do advanced tree-based ensembles (e.g., XGBoost) compare against state-of-the-art deep learning architectures (e.g., Temporal Fusion Transformers) in predicting aggregate day-ahead wind farm output?
  • How does computational complexity compare?

Possible approach:

  • Establish an evaluation framework using generation data and Numerical Weather Prediction features.

  • Train hyperparameter-tuned tree-based models (such as XGBoost) as the machine learning baselines.

  • Implement and train multivariate deep learning sequence models (like a Temporal Fusion Transformer or an LSTM) on the exact same dataset.

  • Evaluate determinisitc forecasts for the day-ahead horizon

  • Evaluate computational complexity of the models
  • Python programming skills
Yes

Jonas Betscher

jonas.betscher@tum.de

2026-03

FOUND_WIND

Zero-Shot Time-Series Foundation Models for Day-Ahead Wind Power Forecasting

Research method: Prototyping

Research questions:

  • Can large-scale time-series foundation models accurately predict day-ahead wind power output at the farm level without any target-specific fine-tuning ?

  • How does the zero-shot performance of these foundation models compare to a fully supervised, site-specific XGBoost model?

Possible approach:

  • Select 1–2 recent open-source time-series or wind-specific foundation models (such as Chronos, Lag-Llama, or WindFM).

  • Generate zero-shot day-ahead forecasts.

  • Fine-tune the foundation model using a small subset of the wind park's data to measure the performance gain from domain adaptation.

  • Benchmark the zero-shot and fine-tuned results against a dedicated supervised baseline model to determine if the computational overhead of foundation models is justified.

  • Python programming skills
Yes

Jonas Betscher

jonas.betscher@tum.de

2026-03

AI_WIND

AI Weather Models for Day-Ahead Wind Power Forecasting

Research method: Prototyping

Research questions:

  • Do state-of-the-art AI weather models (e.g., GraphCast, FourCastNet, or Pangu-Weather) provide superior meteorological feature inputs for wind farm forecasting compared to traditional physics-based NWPs?

Possible approach:

  • Extract localized surface wind speed, wind direction, and temperature forecasts for the target wind park's coordinates using an open-source AI weather model such as GraphCast or FourCastNet.

  • Pair these AI-generated weather covariates with the existing historical farm-level power data.

  • Train a standard predictive model (such as your current XGBoost setup) using the AI weather model outputs, and train an identical model using traditional NWP data.

  • Compare the downstream day-ahead wind power forecasting errors to quantify the direct value added by using AI-driven meteorology.

  • Python programming skills
Yes

Jonas Betscher

jonas.betscher@tum.de

2026-03

ESPARX_SOLAR_FC

Using e-SparX for Developing Solar Power Forecasting Models 

Research method: Prototyping

Research questions:

  • How can e-SparX be used to develop solar power forecasting models and make them transparent and sharable?
  • How to extend and improve e-SparX to make it more useful and usable?

Possible approach:

  • Understand state-of-the-art of solar power forecasting based on literature
  • Understand current e-SparX implementation
  • Develop efficient Python or Julia code to implement several solar power forecasting models
  • Share the complete ML pipelines using e-SparX
  • Propose new e-SparX features to improve its usability
  • Optionally implement new e-SparX features


  • Python programming skills
Yes

Christoph Goebel

christoph.goebel@tum.de

2025-09

SOLAR_POWER_FC

Probabilistic Solar PV Power Forecasting

Research method: Prototyping

Research questions:

  • How can solar PV power output be forecasted probabilistically on different time horizons (intra-hour, intra-day, day_ahead)?
  • How can corresponding methods be implemented and benchmarked?

Possible approach

  • Understand state-of-the-art based on literature
  • Develop efficient Python or Julia code to implement probabilistic forecasting methods
  • Evaluate method using actual data


  • Python/ Julia programming skills
Yes

Christoph Goebel

christoph.goebel@tum.de


2025-03

DEEP_NILM

Deep Learning for Energy Disaggregation

Research method: Prototyping

Research questions:

  • How can Deep Learning be applied to Energy Disaggregation / NILM?
  • How does Deep Learning perform compared to other methods?

Possible approach

  • Understand state-of-the-art based on literature
  • Develop efficient Python code to train DL models on public NILM datasets
  • Evaluate performance by benchmarking with methods described in literature


  • Python programming skills
Yes

Christoph Goebel

christoph.goebel@tum.de

2023-09

DEEP_HEAT

Deep Reinforcement Learning for HVAC Control

Research method: Prototyping

Research question:

  • How can Deep Reinforcement Learning be applied to HVAC Control problems?
  • How does Deep Reinforcement Learning perform compared to other methods?

Possible approach

  • Understand state-of-the-art based on literature
  • Develop efficient Python code to train DRL models using HVAC model
  • Evaluate performance by benchmarking with methods described in literature


  • Python programming skills
Yes

Christoph Goebel

christoph.goebel@tum.de

2023-09

LOCAL_ENERGY_TRADING

Local Energy Trading Under Network Constraints

Research method: Prototyping

Research question:

  • How can prosumers trade energy locally without violating the physical constraints of the distribution system?
  • Which approaches already exist and which assumptions are they based on?
  • How can these approaches be benchmarked using realistic distribution grid models?

Possible approach:

  • Understand state-of-the-art based on literature
  • Develop efficient Python or Julia code implementing existing methods
  • Evaluate performance of methods using realistic models of distribution grids


  • Python/ Julia programming skills
No

Christoph Goebel

christoph.goebel@tum.de

2024-09

BAT_AGE_PRED

Machine Learning-Based Prediction of Battery Aging

Research method: Prototyping

Research question:

  • How to predict lithium-ion battery aging using ML models?
  • How performant are ML models compared to other methods?

Possible approach:

  • Understand state-of-the-art based on literature
  • Develop an efficient Python program that fits various ML models on battery aging data
  • Benchmarking of the developed models
  • Python/ Julia programming skills
  • First experience with deep learning
Yes

Christoph Goebel

christoph.goebel@tum.de

2023-09

RC_MODELS

R-C Models for Heat Transfer in Buildings

Research method: Prototyping

Research question:

  • How to design an easy-to-use Python library for fitting R-C-models of the heat transfer in buildings?
  • How to use fitted models to simulate buildings in control environments?

Possible approach

  • Understand state-of-the-art based on literature
  • Develop efficient Python code to fit R-C models
  • Evaluate implementation, e.g., by benchmarking results and evaluating computational complexity
  • Python/ Julia programming skills
Yes

Christoph Goebel

christoph.goebel@tum.de


2023-09

RC_MODEL_FITTING

Using EnergyPlus to fit R-C Models for Heat Transfer in Buildings

Research method: Prototyping

Research question

  • How can EnergyPlus models be approximated by R-C models?
  • How effective and efficient is the approach?   

Possible approach

  • Learn to work with EnergyPlus
  • Develop efficient Python code fit R-C models to output data
  • Benchmark method
  • Python/ Julia programming skills
Yes

Christoph Goebel

christoph.goebel@tum.de

2023-09

ANN_OPF

----This topic is only available in English----

Power Grids ANN-based Optimal Power Flow

Research method: Prototyping

Research question:

  • What model design is suitable for improving ANNs optimal power flow predictions?
  • What training approaches are suitable for maintain predictions feasibility and optimality?

Possible approach

  • Understand state-of-the-art based on literature
  • Implement ANN model for optimal power-flow prediction
  • Benchmark results against SOTA ANNs
  • Programming skills in Python
Yes

Arbel Yaniv

arbel.yaniv@tum.de

2025-09
 NILM_AL

 ----This topic is only available in English----

Non-Intrusive Load Monitoring with Active Learning

Research method: Prototyping

Research question:

  • What is the impact of different acquisition functions on the performance of appliances disaggregation?
  • What is the best fine-tuning approach, both layer-wise and sample-wise?

Possible approach:

  • Understand state-of-the-art based on literature
  • Leverage scikit-activeml to analyse different active-learning approaches for NILM
  • Try different fine-tuning experimental setups
  • Python programming skills
Yes

Arbel Yaniv

arbel.yaniv@tum.de

2025-09

PV_BAT_SZ

PV and battery sizing with limited or low-resolution data

Research method: Prototyping

Research question: 

  • How can the sizing of PV and battery storage systems be optimized when only limited or low-resolution data is available?

Possible approach

  • Understanding the state of the art from the literature
  • Implementation of possible approaches to improve dimensioning with limited data.
Pythonno

Elgin Kollnig

elgin.kollnig@tum.de 

2025-09

HOUSE_FLEX

Probabilistic household flexibility estimation

Research method: Literature review & Prototyping

Research question

  • What methods exist for determining household flexibility, particularly taking probabilistic approaches into account?

Possible approach

  • Literature review on methods for determining household flexibility
  • Analysis of methods for probabilistic determination of flexibility
  • Implementation and comparison of different approaches


Pythonno

Elgin Kollnig

elgin.kollnig@tum.de 

2025-09
DOPT_P2P

Distributed optimization methods for P2P energy trading

Reseach method:  Literature review, prototyping

Research question:  

  • How can distributed optimization methods (e.g., ADMM or consensus-based algorithms) be designed to ensure fast convergence, scalability, and network constraint satisfaction in P2P energy trading within distribution systems?

Possible approach

  • Compare different distributed optimization methods

  • Investigate trade-offs between convergence speed, privacy, and accuracy

  • Test on small IEEE distribution feeders with simulated P2P trades

Pythonno

Hippolyte Robin

hippolyte.robin@tum.de 

2025-09
 BLOCKCHAIN_P2P

Blockchain-based P2P energy trading

Research method:  Literature review, prototyping

Research question

  • What blockchain architectures, consensus mechanisms, and smart contract designs have been proposed for P2P energy trading, and how do they address challenges of scalability, security, and transaction costs?

Possible approach

  • Classify blockchain architecture, noting their trade-offs in decentralization, scalability, and governance
  • Compare consensus mechanisms, discuss how each handles energy efficiency, speed, security
  • Review smart contracts design (how trades are matched, settled, automated)
  • Identify challenges and research gaps


Pythonno

Hippolyte Robin

hippolyte.robin@tum.de 

2025-09
 MARL_P2P

Multi-Agent Reinforcement Learning Approaches for P2P Energy Trading

Research method: Literature review, prototyping

Research question:

  • How can multi-agent reinforcement learning approaches be designed and evaluated for peer-to-peer energy trading so as to achieve efficient decentralized coordination, scalability, and robust performance under realistic market and network assumptions?

Possible approach

  • Compare different MARL approaches for P2P energy trading

  • Investigate trade-offs between coordination quality, scalability, fairness, and training stability

  • Analyze assumptions regarding agent information, market mechanisms, and grid constraints

  • Optionally test a simple prototype on a small simulated P2P trading setup and compare against a baseline method

  • Python programming skills
yes

Hippolyte Robin

hippolyte.robin@tum.de 

2026-03
P2P_SETTL

On-Chain vs Off-Chain Settlement Architectures for P2P Energy Trading

Research method: Literature review, conceptual modelling

Research question:

  • How do on-chain, off-chain, and hybrid settlement architectures compare in terms of trust, verifiability, scalability, and cost for peer-to-peer local energy markets, and under what market and regulatory conditions does each approach offer the most appropriate trade-offs?

Possible approach

  • Survey existing settlement mechanisms in P2P energy trading platforms and adjacent domains (wholesale energy markets, decentralised finance) to build a taxonomy of architectural approaches
  • Define a structured comparison framework covering key properties: settlement finality, auditability, transaction throughput, latency, privacy, and cost as a function of community size
  • Analyse the measurement oracle problem — how off-chain meter data is anchored on-chain — and review cryptographic and governance-based approaches proposed in the literature
  • Examine how settlement design interacts with penalty and reward mechanisms, and what game-theoretic implications different architectures have on participant behaviour
  • Optionally map findings against an existing platform (e.g. GSY DEX) to ground the comparison in a concrete implementation reference
  • Python programming skills
yes

Hippolyte Robin

hippolyte.robin@tum.de 

2026-03
MOD_GEN

White-box building model generation for Germany using OCHRE

Research method: Prototyping

Research question:

  • How can the advanced python-based energy modeling tool OCHRE be adapted to generate realistic building models for Germany?

Possible approach

  • Analyse the OCHRE model framework, focusing on model set-up and required data formats

  • Review accessible data sources for german households
  • Develop a workflow to set-up the OCHRE model for german households based on publicly available data formats, making assumptions or simplifications where necessary 
  • Generate and analyze temperature and load profiles for an example model
  • Python programming skills

yes

Lennart Morlock

lennart.morlock@tum.de

2026-03
BAT_REV_OPT

Large-scale battery storage revenue optimization under multi-market participation

Research method: Prototyping

Research question:

  • What revenue can you expect for a large-scale battery storage system that participates in balancing and energy markets?

Possible approach

  • Analyse existing public repos for revenue estimation for large-scale battery storage systems (Battery Revenue Index by RWTH Aachen, bess-optimizer by FLEXPWR)

  • Determine potential improvements for a more realistic estimation considering existing literature (e.g. more complex market participation strategies, integration of uncertainty due to forecasts, ...)
  • Compare the outcomes of the new model with the previous implementation.
  • Python programming skills

yes

Lennart Morlock

lennart.morlock@tum.de

2026-03


  • Keine Stichwörter