Development and validation of a data quality assessment framework for machine learning-based building load prediction

Building energy consumption accounts for a significant proportion of global energy use, making smart control and efficient management critical for sustainability. Machine learning (ML) models have shown promise in building energy control systems, but their performance is inherently tied to data quality characteristics, a factor often overlooked in traditional ML modelling. This study proposes a novel data quality assessment (DQA) framework to guide ML modelling for building energy applications, addressing the gap between data characteristics and ML algorithm suitability. The framework evaluates data quality (DQ) across five dimensions: completeness, comprehensiveness, range, consistency, and information entropy. Using 45 office buildings from the building data genome project 2 (BDG2) dataset, the framework was comprehensively investigated and validated using five typical ML models in the building sector including similar day model, LightGBM, long short-term memory (LSTM), transformer, and ensemble model, covering models from simple statistical methods to deeper ML models. The results demonstrate that statistical models such as the similar day model are sufficient for high-DQ scenarios, whereas low-DQ scenarios require deep ML models such as LSTM and transformer. DQ-based model selection approach reduces CVRMSE errors by 38%–40% on average. The proposed framework guides building owners and energy managers to select suitable ML models, thereby enhancing smart control, improving energy efficiency, and reducing carbon emissions. It also paves the way for standardized DQA in future research, effectively bridging the gap between data science and practical ML application in building energy systems.