Improving Time-Series Forecasting Performance Using Imputation Techniques in Deep Learning

This study investigates the effectiveness of various missing data imputation techniques on the performance of deep learning models for time-series forecasting using the Beijing PM2.5 dataset. Imputation methods including Mean, Mode, Median, Multiple Imputation by Chained Equations (MICE), Last Observation Carried Forward (LOCF), and K-Nearest Neighbors (KNN) are compared across different deep learning architectures: Recurrent Neural Networks (RNN), Long Short-Term Memory Networks (LSTM), Bidirectional LSTM Networks (Bi-LSTM), and Gated Recurrent Units (GRU). Performance is evaluated using metrics such as Mean Absolute Percentage Error (MAPE), Root Mean Squared Error (RMSE), and Coefficient of Determination (R2). Results indicate that the KNN imputation method consistently outperforms other techniques across all models, achieving lower MAPE (0.186) and RMSE (18.495) values. Furthermore, KNN-imputed models exhibit higher R2 values, indicating better model fit than other imputation methods. These findings underscore the importance of preprocessing in time-series forecasting and highlight the potential of KNN imputation to enhance predictive accuracy in real-world applications. Robust imputation strategies like KNN promise to improve forecasting accuracy and reliability, particularly in domains where accurate predictions are crucial for decision-making. Future research may focus on refining and integrating imputation techniques with advanced deep-learning methodologies to enhance forecasting capabilities in diverse application scenarios.