In-depth sensitivity analysis of heating demand and overheating in Dutch terraced houses using interpretable machine learning

• A novel sensitivity analysis method based on partial dependence plots is proposed. • It leverages dictionary learning and surrogate models to produce in-depth insights. • Surrogate models accurately predicted heating demand and overheating (MAPE = 3–4 %). • The impact of insulation on Dutch terraced houses was compared in various scenarios. • The value of our method for model diagnosis and design of experiments was shown. Sensitivity analyses are often performed to facilitate the design or modeling of complex building systems by focusing on the most influential parameters. However, the insights they produce are generally limited to a ranking of the parameters’ impact. Instead, many applications could benefit from more advanced insights, such as how these impacts vary across different parameter values and scenarios. In addition, sensitivity analysis results should be easily interpretable to facilitate subsequent decision making. With these goals in mind, this paper introduces a novel sensitivity analysis method based on partial dependence (PD) plots. PD plots are further combined with dictionary learning, advanced visualizations, and surrogate models to facilitate their analysis and reduce their computational cost. Using this method, the effect of 26 parameters on heating demand and overheating in Dutch terraced houses is investigated. Two surrogate models are trained on 66,000 EnergyPlus simulations to predict the annual heating demand and percentage of overheating hours with excellent precision ( ≈ 3 –4 % of percentage error). The benefits of our approach are demonstrated through 3 use cases: 1) a comparison of the impact and energy-overheating trade-off of insulation measures across various scenarios, 2) improving the design of parametric simulations by eliminating redundant parameter values, and 3) uncovering complex behaviors in simulation or surrogate models, to build trust in them and diagnose potential modeling or training errors. Finally, our results suggest that surrogate models can be trained on much less data (1000–3000) without compromising sensitivity analysis results.