Work

Rakennusten lämpöenergian erittely syväoppimismenetelmin

Year: 2023
Type: dissertation
Abstract: Energy retrofit projects for existing buildings aim to improve energy acquisition costs and overall energy efficiency. Informed energy investment decisions require extensive data to achieve optimal results. Overall energy consumption provides only a partial picture of the information required and consumption data segmented by energy end-use is needed to make informed investment decisions. However, detailed consumption data is rarely available and only overall energy consumption time series are available. Engineering consultancies receive technical information, architectural drawings and building parameters as part of project initial data. Based on this information, a simulation model can be constructed to obtain the disaggregated consumption data. However, developing a simulation model is time and resource intensive. In this study, a data-driven model for building heat disaggregation task was developed. A literature review of data-driven models in energy disaggregation and energy prediction and forecasting was conducted. Machine learning algorithms were found to achieve good performance in disaggregation and prediction tasks in the energy domain with neural networks exhibiting best results. A dataset was constructed from the simulations of the building design tool IDA ICE. The dataset was analysed and engineered into a training dataset for data-driven model development. A variety of neural network architectures were trained on the heat disaggregation task. Feedforward network with four hidden layers and 127 004 trainable parameters achieved the best validation performance with MAE of 0.0814, RMSE of 0.1388 and r2-score of 0.704. The performance of the model was evaluated with respect to energy industry model validation methodologies as well as with respect to the end-use case of the model. The model was assessed to perform the heat energy disaggregation task sufficiently. High maximum errors as well as a relatively low r2-score are key factor to reconsider if the model is to be deployed in other than preliminary assessment tasks.
Author: Forsman, Kaarlo
Institution: Sweco (Sweden)
Language: English
Open Access status: green
OpenAlex ID: W7061475037
Cites: 44
Cited by:
Related to: 10