A systematic approach in constructing typical meteorological year weather files using machine learning

Abstract

A typical meteorological year weather file, supposedly representing the climatic conditions, is commonly used as an input for building energy simulation to predict the long-term performance of the buildings. Depending on the building type and climatic locations, the deviation between the actual long-term average energy performance and the predicted one may not be insignificant. Weather files are synthetically constructed on historical weather data over a long period of time for an array of weather parameters, such as solar radiation, temperature, wind speed and others. The statistical procedure to construct the weather files depends on the weights assigned to these weather parameters. Under current practice, these weighting factors are universally assigned regardless of climatic locations nor building applications based on experts’ judgment. The objective of this paper is to propose a novel approach in assigning weighting factors systematically, demonstrate its applicability, and quantify the potential reduction in deviation.Instead of basing on experts’ judgement, Random Forest algorithm was deployed to extract the feature importance of the weather parameters in order to assign weighting factors straightly proportional to their impacts on energy performance of buildings. To ensure the assigned weighting factorsfully reflect the climatic characteristics, five different training approaches were proposed, and the resulting climatic location dependent weather files were constructed and cross-validated.The newly constructed typical meteorological year weather files were applied to two different climatic locations to investigate the representativeness of these new weather files as compared to existing weather files and historical weather data of actual years. The representativeness was indicated in terms of the deviation in predicted energy performance of buildings between using the typical meteorological year weather file and actual historical weather data. The results indicated that typical meteorological year weather file based on the novel approach offers better prediction (with statistical significance) on energy performance for climatic locations with wider temperature range. As a result, the suggested approach avoids potential under/oversizing of equipment and promotes energy conservation.