New AI Model Improves Utility Data Imputation Accuracy

This paper introduces MBDiff, a Multi-view Behavior-aware Diffusion Model designed to address the critical problem of missing values in utility data, such as electricity, water, and gas consumption. Missing data can severely impact billing accuracy, forecasting, and supply management. Existing imputation methods often overlook rich user behavior information, which could provide valuable insights. MBDiff tackles this by incorporating two main components. First, a multi-view User Behavior Extraction module learns comprehensive user behavior from global, local, and instance-level perspectives, overcoming the challenge of learning from long-term, diverse, and incomplete data. Second, a behavior-aware conditional diffusion model, featuring a reference selection module and a conditional attentional denoising network, efficiently imputes the missing data. The model was implemented and evaluated in collaboration with a large municipal utility provider in Florida. Experimental results demonstrate that MBDiff significantly outperforms state-of-the-art baselines, showing improvements of 7.04% for electricity and 29.1% for water usage datasets in block missingness imputation. This highlights the effectiveness of leveraging user behavior for more accurate and reliable utility data management.