Rainfall Forecasting Using Data Decomposed Technique and Deep Learning

Abstract

Rainfall data is essential for forecasting, planning, and water resource management. Various rainfall forecasting models are currently used, particularly numerical weather prediction models, which require large amounts of data and computational resources. This study aims to forecast rainfall in the upstream catchment of Lam Ta Khong Dam to support reservoir inflow prediction. The study area is characterized by natural river flow (free flow) without hydraulic structures controlling the flow, so the inflow into the reservoir directly reflects natural runoff, making it highly suitable for studying rainfall and runoff behavior. This study hypothesizes that applying Variational Mode Decomposition (VMD) before forecasting helps the model learn key data characteristics by reducing signal complexity and distinguishing patterns within highly variable data, thereby improving forecasting accuracy. VMD is a signal decomposition technique that separates data into modes across different frequency ranges to reduce data complexity, while Stack Long Short-Term Memory (Stack LSTM) is a deep sequential neural network designed to enhance the model’s ability to capture complex temporal features. Therefore, the study is divided into two cases for performance comparison: Case 1 applies Stack LSTM alone, and Case 2 applies VMD combined with Stack LSTM. The results show that using VMD combined with Stack LSTM significantly improves forecasting accuracy. The one-day-ahead rainfall forecasting achieved an R value of 0.928, NSE of 0.861, RMSE of 2.858 millimeters, and MAE of 1.409 millimeters, which are better than using daily average rainfall directly. Therefore, applying data decomposition before forecasting effectively reduces errors and improves the model’s accuracy.