A Comparative Evaluation of Various Imputation Methods for LSTM-Based Climatological Time Series Forecasting
TL;DR
It is demonstrated that imputation effectiveness depends on alignment with missingness structure rather than methodological complexity, providing evidence-based guidance for robust environmental time-series forecasting.
Abstract
Missing values substantially degrade the reliability of environmental time-series forecasting; however, prior studies largely evaluate imputation methods in isolation without systematically linking missingness mechanisms to deep learning forecasting performance. To address this gap, this study proposes a mechanism-aware comparative framework that evaluates deletion and six imputation methods (Mean, Median, Mode, LOCF, KNN, and MICE) across three environmental time-series datasets with naturally occurring missing values, using LSTM as the forecasting model. The novelty lies in jointly analyzing statistical error (MAPE, RMSE), goodness-of-fit (R²), and statistical significance to identify structurally aligned imputation strategies under different missingness patterns. Experimental results show that deletion as baseline consistently produces the worst performance (MAPE: 5.91429; 7.35000; 2.84881), whereas imputation reduces proportional error by more than 70% on average (p < 0.05). LOCF performs best under temporal dependency (MAPE 0.73959; R² 0.92757), KNN achieves the most balanced performance under MCAR-like behavior (R² 0.94086), and Mean imputation yields the lowest error in MAR-structured data (MAPE 0.41560; R² 0.97077). These findings demonstrate that imputation effectiveness depends on alignment with missingness structure rather than methodological complexity, providing evidence-based guidance for robust environmental.