Skip to content
Technische Universität Berlin Deutsch
Course catalog and module transfer system

Project Neural Time Series ModellingModule description #51211 / Version 1

← Module page Deutsche Fassung
Module / Version#51211 / #1
ValiditySince Winter semester 2026/27
Available language(s)English, German
Credit points6
Person responsible for the moduleStender, Merten
GradingGraded
Exam formatPortfolio examination
Language(s) of instructionEnglish

Affiliation

FacultyFakultät V
InstituteInstitut für Maschinenkonstruktion und Systemtechnik (IMS)
Academic chair35352200 FG Cyber-Physical-Systems in Mechanical Engineering

Contact

OfficeH 66
Contact personStender, Merten
E-mail addressmerten.stender@tu-berlin.de
Websitehttps://www.tu.berlin/en/cpsme

Learning outcomes

After successfully completing the module, students are able to:
- Formulate time-series problems from engineering applications as data-driven learning tasks.
- Identify temporal structures such as autocorrelation, non-stationarity, and seasonality and account for them in modelling and evaluation.
- Explain and apply the mathematical foundations of common time-series models.
- Select and justify suitable methods ranging from classical statistical approaches to neural architectures.
- Assess models in terms of performance, generalisation, assumptions, limitations, and method-induced biases.
- Identify and interpret relevant scientific publications and use them to justify methodological choices.
- Independently implement, train, and evaluate time-series models in Python and PyTorch.
- Integrate insights and subject-specific knowledge into a well-founded modelling strategy.

Content

Fundamental modelling approaches:
- Time series regression (sequence-to-sequence modelling)
- Time series forecasting (single-step, multi-step, autoregressive)
- Time series classification, anomaly detection, feature engineering
- Learning time series representations (embeddings)

Families of methods:
- classical statistical models (ARIMA-type)
- stateless and attention-based models (convolutional, transformer)
- stateful models (recurrent, Mamba)
- alternative approaches (reservoir computing, neuromorphic computing)
- physics-informed models, kernel methods
- time series foundation models

Domains:
- Univariate and multivariate time series
- Deterministic (regular, irregular) and stochastic (NARMA-type) processes
- Temporal multi-scale dynamics and spatio-temporal processes
- Small-data limits

Applications:
- Energy systems, manufacturing, production systems
- Structural dynamics of (electro-) mechanical structures
- Aerodynamics, robotics and control

Module components

Compulsory area

The following courses are mandatory for this module:

CoursesTypeNumberFrequencyLanguageHours per weekCourse catalog
Neural Time Series ModellingProject Winter semester[en]4 ISIS

Courses in the course catalog

Courses created from the teaching units of this module.

Winter semester 2026/27 (1)

Neural Time Series Modelling (Project) 1

All groups of this course in the catalog

Workload and credit points

Neural Time Series Modelling (Project)

Workload descriptionMultiplierHoursTotal
Attendance152 h30 h
Pre/post processing101 h10 h
Independent Project Work190 h90 h
Concept Paper120 h20 h
Prototype Development120 h20 h
Final Presentation Preparation110 h10 h
180 h

The workload of the module adds up to 180 hours. The module therefore comprises 6 credit points.

Description of teaching and learning methods

The course combines lectures with project-based learning. Lectures introduce the fundamental and advanced approaches to time series modelling that form the methodical basis for the projects. Students then work individually or in small groups on individual projects, supervised through weekly meetings with the teaching staff and structured around clear milestones, including requirements specification (concept paper), implementation, evaluation, and documentation. As part of their project work, students identify a suitable time series modelling problem, conduct data analysis, select and train an appropriate model, and evaluate its performance. Building on this, they develop a deployable prototype, including a user interface, and conclude the course with a final presentation of their methodical approach and results.

Requirements for participation and examination

Desirable prerequisites for participation in the courses:

Students are expected to have completed an introductory course in machine learning or deep learning and to be able to implement and train a basic neural network in Python. Prior experience with PyTorch, JAX, Keras, or TensorFlow is strongly recommended.

Mandatory requirements for the module examination registration:

This module has no examination requirements.

Completion of the module

GradingGraded
Exam formatPortfolio examination
Type of portfolio examination100 points in total
LanguageEnglish, German

Test elements

NamePoints/WeightCategoryDuration/Extent
Concept Paper30written2 Pages
Software Prototype and Documentation40practicalExecutable web application including source code
Final Presentation30oral15min

Exam description

Portfolio assessment, comprising three components carried out throughout the semester in different formats:
- (1) Concept paper (written report, 2 pages: description of the modelling task, baseline results, choice of metrics, 30%),
- (2) Prototype (documented practical work: a working web application for inference on a self-trained PyTorch model, 40%),
- (3) Final presentation (oral presentation, 15 minutes, 30%).

The module mark is calculated from the total number of points using the specified grading scale. All examination components are completed individually.

Duration of the module

The following number of semesters is estimated for taking and completing the module: 1 semester(s).

This module can be started in the following semesters: Winter semester.

Maximum number of participants

The maximum capacity of participants is 30.

Registration procedures

Registration via the ISIS course; admission to the module examination via Moses. If the number of registrations exceeds the number of available examination places, selection will be made according to merit categories in accordance with Section 48(3) of the AllgStuPO. It is possible to attend the course without sitting the examination.

Recommended reading, lecture notes

Lecture notes

unavailable

Electronic lecture notes

available – Presentation slides, Jupyter Notebooks

Recommended literature

  • Goodfellow, Bengio, Courville: Deep Learning (MIT Press, free online access)
  • Hyndman, Athanasopoulos: Forecasting: Principles and Practice (3rd ed., free online access)
  • Lim, Zohren: Time Series Forecasting With Deep Learning: A Survey (Phil. Trans. R. Soc. A, open access)
  • Prince: Understanding Deep Learning (MIT Press, free PDF + print)

Assigned programs

This module is not used in any program.

Miscellaneous

Not specified