Comparing machine learning algorithms for imputation of missing time series in meteorological data

This paper explores advanced feedforward neural networks specifically multi-layer perceptron (MLP), long short-term memory (LSTM), and convolutional neural networks (CNNs) as time-series imputation techniques to address the challenge of missing data in analytical contexts. The study evaluates their performance by introducing artificial data gaps of varying durations 3 days, 1 week, and 1 month. The results reveal that all three algorithms (MLP, LSTM, and CNN) exhibit the ability to estimate incomplete data, yet with differing accuracies. LSTM and CNN outperform in filling short-term gaps (3 days and 1 week) with R² values of 77% and 70% for LSTM, and 58.4% and 69.7% for CNN. MLP also demonstrates effectiveness, achieving accuracies of 74.9% for a 3-day gap and 67.7% for a 1-week gap. Notably, CNN proves the most accurate for monthly data gaps, attaining an R² value of 70.1%. The findings suggest that the selection of imputation techniques should consider the specific time gap, with CNN highlighted as particularly effective for monthly gaps. In conclusion, this study provides valuable insights for researchers and practitioners engaged in imputing missing data in time-series analysis.