Procedures for Assessing Summer Heat Protection Using AI Methods

Abstract Climate change leads to rising surface temperatures in Europe, making the built environment particularly vulnerable to heat events. For the task of evaluating summer heat protection of buildings, the use of a machine learning method was investigated. This paper first describes the limitations of the procedures according to DIN 4108-2 and then the development of a method based on the Random Forest Regressor (RFR). Subsequently, the results are compared with those from a thermal building simulation. The RFR was trained with randomized simulation data to predict the operative temperature and indoor air temperature. The evaluation shows that while the RFR provides qualitatively similar time series, significant quantitative differences exist. The overheating degree hours are systematically underestimated compared to simulation, especially in extreme weather scenarios. The study discusses challenges of unevenly distributed training data and the sensitivity of results towards the test reference years used. The approach offers valuable insights, highlights the limits of the chosen method for assessing summer heat protection, and underscores the need for further optimizations.