TUTCRIS - Tampereen teknillinen yliopisto

TUTCRIS

Mitigating the Weaknesses of Machine Learning in Short–Term Forecasting of Aggregated Power System Active Loads

Tutkimustuotosvertaisarvioitu

Yksityiskohdat

AlkuperäiskieliEnglanti
OtsikkoThe 17th IEEE International Conference on Industrial Informatics (INDIN 2019)
KustantajaIEEE
Sivut303-310
Sivumäärä8
ISBN (elektroninen)9781728129273
DOI - pysyväislinkit
TilaJulkaistu - 23 heinäkuuta 2019
OKM-julkaisutyyppiA4 Artikkeli konferenssijulkaisussa
TapahtumaIEEE International Conference on Industrial Informatics - Helsinki, Helsinki, Suomi
Kesto: 22 heinäkuuta 201925 heinäkuuta 2019
https://www.indin2019.org/

Julkaisusarja

NimiIEEE International Conference on Industrial Informatics (INDIN)
ISSN (elektroninen)1935-4576

Conference

ConferenceIEEE International Conference on Industrial Informatics
LyhennettäINDIN '19
MaaSuomi
KaupunkiHelsinki
Ajanjakso22/07/1925/07/19
www-osoite

Tiivistelmä

Machine learning methods predict accurately in situations that are adequately included in the learning data and do not require detailed domain knowledge based model development. They have their weaknesses compared with other forecasting methods, however. For example, they completely fail in many new situations not experienced before. Hybrid models are increasingly popular as they are capable of combining the strengths of several modelling methods and mitigate the weaknesses. We study short–term forecasting of aggregated electricity demand that includes dynamically controlled thermal storage. Purely measurement data driven models tend to fail in forecasting power in rarely occurring situations, such as dynamic load control actions and extreme weather. The thermal dynamics of the loads, large outdoor temperature variations, and changes in the energy technologies contribute to this challenge. Combining various information sources and the strengths of different modelling approaches is needed. We study the following approach using field trial data covering over 7500 houses and 27 months. We forecast control responses and load saturation using models that have physically based model structures. Then we forecast the residual using data driven models, such as machine learning models designed and tuned to learn also system dynamics. The load forecast is the sum of these component forecasts. We further improve the forecast by using ensemble forecasting and physically based range forecasts. We find that the hybrid methods are more accurate than their component methods alone and combining several hybridization approaches can improve the performance and reliability.

Tutkimusalat

Julkaisufoorumi-taso