TY - GEN
T1 - Evaluating Machine Learning and Heuristic Optimization Based Surrogates as a Replacement for a Complex Building Simulation Model
AU - Kefer, Kathrin
AU - Haijes, Samuel
AU - Mörth, Michael
AU - Heimrath, Richard
AU - Mach, Thomas
AU - Kaisermayer, Valentin
AU - Zemann, Christopher
AU - Muschick, Daniel
AU - Burlacu, Bogdan
AU - Winkler, Stephan
AU - Affenzeller, Michael
N1 - Publisher Copyright:
© 2023 The Authors.
PY - 2023
Y1 - 2023
N2 - Intelligent energy management systems can play a vital role in supporting the much needed energy transition. However, in order to train machine learning models for this task, often very complex and detailed simulation models are needed. This can make the overall training process very slow or even impossible, which is why using resource efficient surrogates of the original simulation model during the training can be a possible solution. This work therefore focuses on the training of surrogates of a very detailed building simulation model using three different algorithms (k-Nearest Neighbour, Random Forest and Genetic Algorithm) and evaluates and compares them for their prediction capabilities, learned behaviours as well as execution time. Results show that the Random Forest algorithm achieves the best overall performance for 28 of the 35 surrogates, can learn the expected behavior and improves the execution speed by a factor of up to 664 compared to the original IDA ICE simulation model.
AB - Intelligent energy management systems can play a vital role in supporting the much needed energy transition. However, in order to train machine learning models for this task, often very complex and detailed simulation models are needed. This can make the overall training process very slow or even impossible, which is why using resource efficient surrogates of the original simulation model during the training can be a possible solution. This work therefore focuses on the training of surrogates of a very detailed building simulation model using three different algorithms (k-Nearest Neighbour, Random Forest and Genetic Algorithm) and evaluates and compares them for their prediction capabilities, learned behaviours as well as execution time. Results show that the Random Forest algorithm achieves the best overall performance for 28 of the 35 surrogates, can learn the expected behavior and improves the execution speed by a factor of up to 664 compared to the original IDA ICE simulation model.
KW - Building Simulation Model Surrogates
KW - Energy Management System
KW - Heuristic Optimization
KW - Machine Learning
UR - http://www.scopus.com/inward/record.url?scp=85179134118&partnerID=8YFLogxK
U2 - 10.46354/i3m.2023.sesde.004
DO - 10.46354/i3m.2023.sesde.004
M3 - Conference contribution
AN - SCOPUS:85179134118
T3 - Proceedings of the International Workshop on Simulation for Energy, Sustainable Development and Environment, SESDE
BT - 11th International Workshop on Simulation for Energy, Sustainable Development and Environment, SESDE 2023
A2 - Bruzzone, Agostino G.
A2 - Janosy, Janos Sebestyen
A2 - Nicoletti, Letizia
A2 - Zacharewicz, Gregory
PB - Cal-Tek srl
T2 - 11th International Workshop on Simulation for Energy, Sustainable Development and Environment, SESDE 2023
Y2 - 18 September 2023 through 20 September 2023
ER -