Paper Conference

Proceedings of eSim 2020: 11th Conference of IBPSA-Canada


Analyzing building envelope retrofit strategies using a neural network model based on sensor data

Chirag Deb, Zhonghao Dai, Arno Schlueter
ETH Zurich, Switzerland

Abstract: This study presents a black-box model to analyze retrofit strategies on the building envelope of a singlefamily residence in Switzerland. The model is a Long-short term memory-based Recurrent neural network (LSTM-RNN) developed on data from a wireless sensor network (WSN). The 39 measured variables are filtered using a two-step feature selection process. The selected features are fed as inputs to the LSTM-RNN model that predicts the hourly heating energy consumption as an output with an accuracy of 81.24 %. A detailed significance analysis ranks the various inputs according to their order of significance. Based on this significance, we analyse the cost-optimal retrofit solution and compare it with the conventional process of exhaustive search. The cost calculations are based on the annuity method using data from a retrofit matrix of various insulation materials. We see that the proposed methodology with machine learning-based significance analysis results in a better retrofit strategy selection.