Research on short-term precipitation forecasting method based on CEEMDAN-GRU algorithm.

Area of Science:

• Environmental science
• Data science
• Meteorology

Background:

• Accurate precipitation forecasting is crucial for disaster management, urban planning, and agricultural productivity.
• Existing time series models face challenges in capturing complex, nonlinear precipitation patterns.

Purpose of the Study:

• To develop and evaluate an improved model for short-term precipitation forecasting.
• To enhance the accuracy and stability of precipitation predictions using advanced decomposition and recurrent neural network techniques.

Main Methods:

• The study combines Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) for signal decomposition and Gated Recurrent Unit (GRU) for time series pattern recognition.
• Precipitation data from January 1, 2019, to December 31, 2022, was utilized for model development and validation.
• The proposed CEEMDAN-GRU model was benchmarked against 12 other models, including CEEMDAN-LSTM, EMD-GRU, and standalone GRU and LSTM.

Main Results:

• The CEEMDAN-GRU model demonstrated superior performance compared to benchmark models in short-term precipitation forecasting.
• Key performance metrics included an R² of 0.7915, Mean Absolute Error (MAE) of 0.05382, and Mean Squared Error (MSE) of 0.09081.
• The integration of the Adam optimizer with adaptive learning rate reduction facilitated enhanced model convergence and prediction reliability.

Conclusions:

• The CEEMDAN-GRU model represents a significant advancement in short-term precipitation forecasting accuracy and stability.
• This hybrid approach effectively leverages signal decomposition and deep learning for robust meteorological predictions.
• The findings support the application of CEEMDAN-GRU for critical weather-dependent operational planning.