Imputation of Missing Data by Characteristic Analysis of Household Water Metering Data and Deep Learning-Based Prediction Study

초록

Smart water grid technologies have been widely adopted as a key component of digital transformation in water resource management, where real-time household water consumption data collected from smart water meters serve as fundamental inputs. However, these datasets often contain numerous outliers and missing values due to communication errors, which degrade data reliability and hinder accurate analysis. This study proposes an improved framework for outlier detection and missing data imputation tailored to the characteristics of cumulative household water consumption data. The proposed imputation methods were evaluated against conventional approaches using error metrics, and the results demonstrated significant improvements in accuracy, with RMSE values substantially lower than those of the reference method. In addition, prediction models with varying levels of complexity were explored to examine how improved data quality influences forecasting performance. The results indicate that, although data preprocessing enhances data reliability, prediction performance remains limited due to the inherent variability and stochastic nature of household water consumption data. Prediction models with varying levels of complexity were constructed and evaluated using the corrected datasets. The performance of the models varied depending on dataset characteristics, and no single model consistently outperformed others. Overall, this study highlights the critical role of data quality improvement in smart water management systems and provides practical insights into missing data imputation, while suggesting that further advancements in prediction require additional explanatory variables and more sophisticated modeling approaches.

키워드