Prediction of cooling load in the office building: a comparison of various machine learning methods

Prognozowanie obciążenia chłodzenia w budynku biurowym: porównanie różnych metod uczenia maszynowego
Jakub Banaszak, Arkadiusz Szczęśniak, Wojciech Bujalski, Kamil Futyma, Andrzej Grzebielec

    Streszczenie
    Accurate prediction of consumer’s cooling demand is critical to the optimal control strategy of a hybrid district
    heating substation, which uses system heat to produce cold for individual demand by means of an adsorption chiller. This
    study focuses on the investigation of different machine learning methods for cooling load prediction for the office building
    located in Poland, where the temperate climate zone with four distinct seasons exists. The study includes the comparison of a
    wide range of simpler and more complex algorithms with varying accuracy-complexity trade-off potential, including Linear
    Model Trees. The correlation matrix for different characteristics (such as air temperature, irradiance, etc.) has been determined.
    The most promising methods are examined based on the simulated cooling consumption data of the office building. Several
    types of input features combinations were taken into consideration. The best performing methods were MLP, Linear Tree and
    SVR. In real operation, a prediction of the cooling load for the whole upcoming working day is required for the optimal
    planning of the work of heat and cold storage. Instead of using historical data, weather forecasts need to be used where needed.
    Dostęp do pełnej wersji artykułu po zalogowaniu się na stronie.