A combined monthly precipitation prediction method based on CEEMD and improved LSTM

Area of Science:

• Hydrology
• Climate Science
• Data Science

Background:

• Declining water resources necessitate accurate precipitation prediction for effective water management.
• Monthly precipitation data exhibit complex non-linear and non-stationary characteristics, posing challenges for traditional forecasting models.

Purpose of the Study:

• To develop a novel combined prediction method for monthly precipitation.
• To address the challenges of non-linearity and non-stationarity in precipitation data.
• To improve the accuracy and reliability of precipitation forecasting.

Main Methods:

• Decomposition of monthly precipitation series into stationary sub-sequences using Complementary Ensemble Empirical Mode Decomposition (CEEMD).
• Prediction of individual sub-sequences using a modified Long Short-Term Memory (LSTM) neural network.
• Optimization of LSTM hyperparameters via Particle Swarm Optimization (PSO) algorithm.
• Superposition of predicted sub-sequences to obtain the final precipitation forecast.

Main Results:

• The CEEMD-LSTM method successfully decomposed precipitation data, revealing local characteristics and nonlinear dynamics.
• Particle Swarm Optimization effectively optimized LSTM hyperparameters, reducing prediction randomness.
• The combined CEEMD-LSTM model demonstrated superior performance compared to traditional methods in forecasting monthly precipitation.
• The model accurately captured precipitation trends and exhibited higher prediction accuracy in a case study.

Conclusions:

• The proposed CEEMD-LSTM method offers a robust approach for monthly precipitation prediction, effectively handling complex data characteristics.
• This advanced forecasting technique provides valuable insights for rational water resource allocation and management.
• The study highlights the potential of hybrid decomposition-machine learning models in environmental forecasting.