Improving the performance of machine learning models by integrating partly physical control response models in short-term forecasting of aggregated power system loads

Pekka Koponen, Harri Niska, Reino Huusko

    Research output: Chapter in Book/Report/Conference proceedingConference article in proceedingsScientificpeer-review

    65 Downloads (Pure)

    Abstract

    Combining the strengths of different modelling approaches and various information sources is studied in short-term forecasting of aggregated electrical loads that are controllable and include e.g. thermal storage capacity. Measurement data driven models tend to fail in forecasting power during rare situations such as dynamic control actions and extreme weather conditions. The thermal dynamics of the loads, large outdoor temperature variations, and changes in the technologies contribute to this challenge. Here we study a model integration approach using field trial data covering about 7000 houses and 27 months. Control responses and load saturation are forecast using a physically based structure. The residual is forecast with a machine learning model designed and tuned to learn also system dynamics. The load forecast is the sum of these component forecasts. The forecasting accuracy of this hybrid method is compared with using the machine learning alone. The results show improvement in the accuracy
    Original languageEnglish
    Title of host publicationProceedings ITISE 2017. Granada, 18-20, September, 2017.
    EditorsOlga Valenzuela, Ignacio Rojas
    PublisherUniversity of Granada
    Pages795-806
    Volume2
    ISBN (Print)978-84-17293-01-7
    Publication statusPublished - 2017
    MoE publication typeA4 Article in a conference publication
    EventInternational Work-Conference on TIme SEries Analysis, ITISE 2017 - Granada, Spain
    Duration: 18 Sept 201720 Sept 2017

    Conference

    ConferenceInternational Work-Conference on TIme SEries Analysis, ITISE 2017
    Abbreviated titleITISE 2017
    Country/TerritorySpain
    CityGranada
    Period18/09/1720/09/17

    Keywords

    • forecasting
    • machine learning
    • physically based models
    • smart grid

    Fingerprint

    Dive into the research topics of 'Improving the performance of machine learning models by integrating partly physical control response models in short-term forecasting of aggregated power system loads'. Together they form a unique fingerprint.

    Cite this