Master's Thesis Dawid Nickel


Analysis and development of thermal demand forecast models for residential buildings

The ambitious aims of the German government for reducing the greenhouse gas emissions and increasing
the share of Renewable Energy Sources paved theway to the transition of the energy system
to a decentralized power generation. Because of the volatility of the decentralized power generation
diverse energy management concepts i.e. demand side management and demand side response,
are investigated. The main feature of these concepts is the coordination of Distributed Energy Resources
i.e. heat pumps and combined heat and power units, according to different constraints to
achieve a certain target. The coordination typically includes ahead scheduling of the DERs to anticipate
load peaks resulting from consumption or generation of renewable energies. To ensure the
thermal comfort of the residents by fulfilling the constraint of the impending thermal demand this
ahead scheduling strategy is depended on a sufficiently good forecast.
In the context of this thesis, three models for the prediction of the heat demand in residential buildings
were studied. These are an advanced regression model, an artificial neural network and a
Grey-Box-Model. For all three models a past time series was used for the parameter identification,
consisting of the weather conditions, the indoor temperature and the heat consumption or heat generation.
There are two different approaches considered for the prediction of the heat demand. On
the one hand, under the condition of a constant indoor temperature, the direct heat demand will
be determined to compensate heat losses and to keep the temperature constant. On the other hand
the behavior of the indoor temperature is represented to determine the heat generation, which is
necessary for obtaining a comfortable temperature.
With the help of the regression model, a polynomial relationship between the environmental conditions
and the heat requirement was created. By an integration of a low-pass filter, the thermal inertia
of a building was depicted. In the prediction of the heat losses a relative error of 28% and a coefficient
of determination of 0.95 could be achieved. The indoor temperature could not be depicted.
In the case of the artificial neural network various structures and also the quantity of neurons and
layers were analyzed. Thereby the autoregressive network, with the inclusion of exogenous input
data reached the best results. The deviation in the heat loss prediction was 29% and the coefficient
of determination was 0.84.
Concerning the Grey-Box-Model a simplified physical behavior of a residential building was reproduced.
In spite of the case that the determination of the heat loss revealed an error of 37% (coefficient
of determination of 0.94) it exhibits the greatest potential of the models examined because it
can be created a prediction on the condition of both a constant and dynamic indoor temperature.
In the prediction of the indoor temperature the difference in temperature averaged 0.57 K.