A new rainfall prediction model based on ICEEMDAN-WSD-BiLSTM and ESN

Area of Science:

• Climate Science
• Hydrology
• Artificial Intelligence in Environmental Science

Background:

• Precipitation patterns are key indicators of regional climate evolution and distribution.
• Accurate precipitation prediction is vital for flood and drought control planning.
• Precipitation data often exhibits complex nonlinear and uncertain characteristics.

Purpose of the Study:

• To develop and evaluate an advanced hybrid model for predicting regional precipitation.
• To address the challenges posed by uncertainty and nonlinearity in precipitation time series.
• To enhance the accuracy of precipitation forecasts for practical applications in climate management.

Main Methods:

• Utilized Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) for signal decomposition.
• Applied Wavelet Signal Denoising (WSD) to preprocess the decomposed components.
• Employed a hybrid model combining Bi-directional Long Short-Term Memory (BiLSTM) for high-frequency components and Echo State Network (ESN) for low-frequency components.

Main Results:

• The proposed ICEEMDAN-WSD-BiLSTM and ESN hybrid model achieved high prediction accuracy.
• The model demonstrated an average relative error of 2.64%.
• Achieved a Nash-Sutcliffe Efficiency (NSE) coefficient of 0.91, outperforming other models.

Conclusions:

• The hybrid model effectively captures temporal patterns and evolutionary characteristics of precipitation.
• The findings offer significant theoretical and practical value for flood prevention and regional sustainable development.
• This approach provides a robust framework for improving climate system monitoring and prediction.