Murtaza, S., Raj, S., Yun, G. Y., Park, D. J., Kim, J. H., Park, G., & Moon, J. W. (2025). Adaptive neural temporal hybridization for missing data imputation in building energy use datasets: An integrated LNN-LSTM weighted model. Journal of Building Engineering, 102, 113774. https://doi.org/10.1016/j.jobe.2025.113774
Abstract
Accurate and complete building energy consumption data is essential for optimizing energy efficiency, forecasting demand, and supporting energy management systems. However, missing data from sensor malfunctions or communication failures can reduce the effectiveness of data-driven decision-making. This study introduces the integrated LNN-LSTM weighted model (ILLWM), a novel imputation approach that combines the adaptability of liquid neural networks (LNN) with the temporal modeling capabilities of long short-term memory (LSTM) models. Imputed values are generated using an RMSE-based weighted approach. ILLWM was tested on real-time energy consumption data from three building types, missing completely at random scenarios with missing rates of 20 %, 30 %, and 40 %. Results showed ILLWM significantly outperformed other imputation methods, including Soft-Impute, KNN, RF, SVM, MLP, Transformer networks, LSTM, and LNN. For commercial buildings with 40 % missing data, ILLWM achieved RMSE reductions of 76.9 % and 89.6 % over LNN and LSTM, respectively. For hospital buildings, improvements included RMSE reductions of 6.12 % over LNN and 31.93 % over LSTM. The ILLWM closely matched actual data, outperforming traditional and machine learning approaches. These results demonstrate the potential of the ILLWM to enhance data reliability, enabling more accurate energy demand forecasting and the development of sustainable energy management strategies in diverse building environments.
