Bischof, Rafael ; Sprenger, Marius ; Riedel, Henrik ; Bumann, Matthias ; Walczok, Waldemar ; Drass, Michael ; Kraus, Michael A. (2023)
Temp-AI-Estimator: Interior Temperature Prediction using Domain-Informed Deep Learning.
In: Energy and Buildings
doi: 10.1016/j.enbuild.2023.113425
Artikel, Bibliographie

Kurzbeschreibung (Abstract)

Approximately 40 of total energy demand in the European Union is consumed by the residential buildings sector, thus also significantly contributing to carbon dioxide emissions. Circa 28 of this energy demand is attributed to space heating and cooling, primarily influenced by the building's envelope and need to ensure indoor thermal comfort. Given this significant energy consumption, there is an urgent imperative to explore energy-saving strategies and develop tools to assess the effects of various design alternatives, with a focus on wall and roof characteristics. While existing white and black-box predictive models lack generalisation capabilities, the goal of this study is to develop and train a domain-informed grey-box Deep Learning model called Temp-AI-Estimator to predict the indoor temperature of buildings and containers (acting as surrogates for residential or office buildings as well as intermodal containers) based on measurements of external meteorological conditions (e.g. exterior temperature, humidity, etc.), as well as physical properties due to building construction set-ups in questions (“ThermoProtect” (thin coating), “ThermoActive” (thick coating) and thermal insulation). The major difficulty lies in making the model generalise beyond the three geographical locations included in the dataset (Berlin, Abu Dhabi, Texel). Experiments with LSTMs and Transformers as baseline models showed overfitting on the particular conditions at the sites in the training set, while failing to generalise to a new out-of-sample location. We propose to pass the model the numerical derivatives of the time-sequences, as these are less location-specific. The estimation of the necessary initial/ final conditions is delegated to a very small network to minimise the risk of overfitting. Furthermore, we included a physically motivated module for modelling the latency and difference in amplitudes between exterior conditions and interior temperatures, informed by numerical approximation schemes of the differential equation of thermal conduction. Our experiments show that our domain-informed network achieves an increase in accuracy of almost 40 in addition to yielding results that can easily be inspected by human experts due to interpretability and explainability.

Frage zum Eintrag

Redaktionelle Details anzeigen