Assessing the Effectiveness of Statistical and Temporal Imputation Methods for Bi-LSTM-Based Forecasting on Environmental and Climate Time Series Data

Abstract

Time series data in climatology and environmental research are highly susceptible to missing values that can disrupt temporal structures and degrade forecasting performance. This study evaluates the effectiveness of several imputation methods in improving the predictive performance of a Bidirectional Long Short-Term Memory model across three missing-data mechanisms: Missing Completely at Random, Missing at Random, and Missing Not at Random. The compared methods include mean, median, mode, k-nearest neighbors, multiple imputation by chained equations, and last observation carried forward, with data deletion serving as the baseline. All datasets were normalized using the min–max technique, and model hyperparameters were optimized through Particle Swarm Optimization. Performance was assessed using mean absolute percentage error, root mean square error, and the coefficient of determination. The findings indicate that proper imputation significantly enhances forecasting accuracy compared to deleting incomplete observations. In Dataset 1, the last observation carried forward achieved the best performance with a coefficient of determination of 0.923 and a root mean square error of 3.373. Similarly, Dataset 2 showed optimal results with the same method, producing a coefficient of determination of 0.950 and a root mean square error of 14.458. The most substantial improvement was observed in Dataset 3, where mean imputation reduced the mean absolute percentage error from 3.219 to 0.329 while increasing the coefficient of determination to 0.986. These results highlight the critical role of selecting an imputation strategy in deep learning-based time series forecasting and provide practical guidance for handling incomplete environmental datasets.