% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@PHDTHESIS{GrsesTran:984188,
author = {Gürses-Tran, Gonca},
othercontributors = {Monti, Antonello and Ulbig, Andreas},
title = {{M}achine learning techniques for time series forecasting
in power systems operation; 1 {A}uflage},
volume = {127},
school = {Rheinisch-Westfälische Technische Hochschule Aachen},
type = {Dissertation},
address = {Aachen},
publisher = {E.ON Energy Research Center, RWTH Aachen University},
reportid = {RWTH-2024-03928},
isbn = {978-3-948234-41-6},
series = {E.On Energy Research Center},
pages = {1 Online-Ressource : Illustrationen},
year = {2024},
note = {Druckausgabe: 2024. - Auch veröffentlicht auf dem
Publikationsserver der RWTH Aachen University; Dissertation,
Rheinisch-Westfälische Technische Hochschule Aachen, 2023},
abstract = {Short-term forecasting of load and generation profiles is
one of the key enablers for coordinated power system
operation. This is because, in comparison to conventional
system operation, future system states are less predictable
due to intermittent generation. Consequently, given the
dependency on uncertain meteorological conditions, matching
the generation and load perfectly up-front becomes
impossible. Deviations between the predicted system state
and the actual generation and load cause costly
counter-balancing measures. This work deals with the
potential reduction of the necessary measures by means of
more accurate short-term forecasting under uncertainty. Most
challenging forecasting tasks in the operational planning
processes are intermittent wind generation and the resulting
volatile vertical grid load forecast, also known as residual
load. The reason for this is that in addition to the
significant uncertainty of wind generation, minor
fluctuations in the demand, as well as other generation
sources, will have an impact on the future net load. The
uncertainty impact of load and generation variability are
hardly separable and can only be modelled with accurate
knowledge of geographical and meteorological conditions and
often require physics based models. Large utilities and
advanced Transmission System Operators (TSOs), more often
than Distribution System Operators (DSOs) build their daily
operational processes on such forecasting tools.This work
develops advanced forecasting tools that are highly scalable
and applicable to forecasting tasks from different operator
perspectives. For this purpose, several challenges for
forecasting in future energy systems have been identified: i
A generalization of particularities of time series profiles
in the energy domain, ii an identification of scalable and
automatable techniques that are applicable for time series
forecasting to enable a fast uptake of digitalized power
system coordination, and iii linking prediction quality and
model usability in operational contexts. This work first
introduces common energy forecasting tasks and typical time
series properties. The introduced theoretical basis
facilitates the identification of potential data-based
techniques, which particularly suit forecasting issues,
i.e., demand and Renewable Energy Source (RES) power
forecasting. In a next step, a set of identified automated
machine learning techniques are compared in detail. A
special emphasis in the selection process is put on
scalability, development speed and ease of use. As a
benchmark a holistic machine learning development pipeline
is introduced theoretically and described in the practical
case of the ProLoaF project. The latter is the key
contribution of the present work. It serves as flexible,
highly configurable and tunable Machine Learning (ML) model
toolbox, which proves to be accurate over all major
forecasting tasks. Besides the evaluation based on
statistical metrics, economic aspects and user-centric
perspectives are discussed in this work. To give a complete
picture of black-box type of models in an operational
context, requirements are set to make ML models more
explainable. As an addition, ProLoaF includes explanatory
tools which can be used to analyze the correctness of model
configurations.In summary, this work achieves progress
towards removing practical obstacles, that power system
operators face today, to achieve a sound adoption of
advanced forecasting methods in an operational context. The
major contributions of this dissertation are: i the
identification of common patterns and criticalities of major
energy time series, ii a review of state of the art ML
forecasting techniques, iii study on quality of grid state
forecasts for operational load assess- ments based on
benchmarking of applicable autoML methods, andiv model
scalability, maintainability and trust through the Machine
Learning Operations (MLOps) pipeline of ProLoaF,v ProLoaF
forecasting as a toolbox to ease and improve forecasts in
system operation.},
cin = {616310 / 080052},
ddc = {621.3},
cid = {$I:(DE-82)616310_20140620$ / $I:(DE-82)080052_20160101$},
typ = {PUB:(DE-HGF)11 / PUB:(DE-HGF)3},
doi = {10.18154/RWTH-2024-03928},
url = {https://publications.rwth-aachen.de/record/984188},
}
Gast ::