A machine learning-based framework for cost-optimal building retrofit

The current process of analysing building retrofit strategies relies on physics-based, thermal balance models. However, these models are oblivious to the significance of the input variables for devising the retrofit strategies. This leads to the process of exhaustive search for obtaining the cost-optimal retrofit strategy. On the contrary, this work presents a framework for a data-driven, cost-optimal retrofit analysis based on machine learning (ML) techniques which capitalizes on the importance of the input variables. The framework involves four steps, which are feature selection, model development, feature significance and cost-optimal retrofit analysis. The developed framework is applied on a case study involving a single-family residence. The time series data on building variables at 5-minute interval is gathered using a wireless sensor network (WSN) for February and March 2019. A two-step feature selection using relevancy and redundancy filters provide the input variables for developing the prediction model. We test four ML models and find that the recurrent neural network (RNN) is the most accurate in predicting the target variable, which is, the space heating demand. This RNN model is used to test the various retrofit options to derive the cost-optimal retrofit solution. Further, we use SHAP (SHapley Additive exPlanations) values to determine the most significant features. We see that these significant features can produce the cost-optimal retrofit strategy in an optimized way without undergoing the exhaustive search process. The best retrofit strategy is found to be the replacement of the oil-based heating system with a brine-to-water heat pump while retrofitting the cellar ceiling with a U-value of 0.388 W/(m(2)*K). This strategy leads to a primary energy consumption of 61.75 kWh/m(2) at a cost of 23.86 CHF/m(2) per year.

Automated summary

A data-driven, cost-optimal retrofit analysis framework based on machine learning techniques is proposed in this work to capitalize on the importance of input variables. Through feature selection and model development, the recurrent neural network is found to be the most accurate in predicting the target variable (space heating demand). The best retrofit strategy involves replacing the oil-based heating system with a brine-to-water heat pump and retrofitting the cellar ceiling with a U-value of 0.388 W/(m(2)*K).