Hybrid LSTM-RF Ensemble for Extreme Block Missingness Imputation in Multivariate Time Series

Abstract

Multivariate time-series data plays an important role in monitoring smart industrial heating systems. However, in real industrial environments, sensor failures and communication disruptions frequently lead to missing data problems, particularly in the form of continuous block missingness. Although deep learning approaches are capable of learning temporal patterns effectively, recurrent models such as Long Short-Term Memory (LSTM) networks often face performance degradation when handling long missing intervals. During prolonged gaps, the autoregressive prediction process causes estimation errors to accumulate over time, leading to prediction drift and unstable outputs. To address this limitation, this paper proposes a Hybrid LSTM-RF ensemble method for imputing extreme block missingness in multivariate time-series data. The proposed model combines the temporal learning capability of LSTM with the stable cross-variable mapping strength of Random Forest (RF) through a parallel weighted ensemble architecture. Experimental evaluation was conducted using the real-world GECCO 2015 industrial heating system dataset with simulated continuous missing gaps ranging from 1 to 48 hours under overall missingness rates of 10% to 50%. The proposed hybrid model was benchmarked against standalone LSTM, Recurrent Neural Network (RNN), RF, mean, median, and mode imputation methods. The results show that the Hybrid LSTM-RF model consistently achieved lower error rates and higher correlation scores across all testing scenarios, particularly under long-duration missing gaps. By reducing the effect of temporal drift during autoregressive inference, the proposed approach provides a more robust and reliable solution for missing data recovery in industrial IoT applications.