Weather clustering for machine learning-based hourly building energy prediction models at design phase

• Machine learning is able to predict annual hourly building energy demands. • 82 million datasets were significantly reduced by weather clustering. • Ten weather patterns and 30 days were adequate to represent whole year. • Theoretical and implementation details of proposed clustering techniques provided for broad impacts. • ML-based hourly building energy predictions are real-time fast with acceptable accuracy. With global efforts aimed at reaching carbon neutrality by 2050, there is an increased emphasis on optimizing building energy management. Accurate hourly building energy predictions support crucial tasks such as predicting peak loads for equipment sizing, comparing energy systems, and optimization during the design phase. The main methods used to model building energy are physics-based and data-driven. The former method has been extensively studied, whereas the latter has not been thoroughly investigated. This paper investigates the advantages of using the machine learning (ML) model as a surrogate model in building engineering, specifically for predicting hourly building energy consumption during the design phase. Synthetic data is commonly used for training and testing ML models when real-life measured data is unavailable due to privacy concerns or pre-construction scenarios. However, the challenge arises from the vast dataset of synthetic data generated by combining long-term hourly meteorological data with building characteristics. Using an example building, 82 million data points were generated as a result of simulating 8,760 h when considering ten building performance parameters. To address this issue, a methodology utilizing weather clustering techniques is proposed in this work. This approach aims to reduce dataset size associated with day-by-day simulations by identifying representative weather patterns. Consequently, 7 million data points were generated by identifying ten weather patterns and selecting 30 days, with three days chosen from each cluster. The Extreme Gradient Boosting (XGBoost) algorithm is applied to develop the ML model using the condensed dataset. This model demonstrated commendable performance with testing results that are within the tolerances established by ASHRAE guideline 14. Although we used data from a residential building in Qatar, our application demonstrated that the approaches could be applied to other building types and climate zones. The developed ML model, utilizing easily accessible inputs, can predict hourly building energy consumption. It is user-friendly for non-experts, such as city developers and stakeholders, during the design and retrofit stage.