A monthly temperature prediction based on the CEEMDAN-BO-BiLSTM coupled model

Area of Science:

• Environmental Science
• Data Science
• Climate Modeling

Background:

• Accurate temperature prediction is vital for climate change monitoring, impacting agriculture, energy, and disaster warning systems.
• Nonlinear and nonsmooth signals in climate data present challenges for traditional prediction models.

Purpose of the Study:

• To develop a robust and accurate monthly average temperature prediction model.
• To leverage the strengths of CEEMDAN, BO algorithm, and BiLSTM for enhanced forecasting capabilities.

Main Methods:

• The study integrates the Complementary Ensemble Empirical Mode Decomposition (CEEMDAN) for signal decomposition.
• The Bayesian Optimization (BO) algorithm is employed for objective function optimization.
• A Bidirectional Long Short-Term Memory (BiLSTM) network is utilized for time-series prediction.

Main Results:

• The CEEMDAN-BO-BiLSTM model achieved an average absolute error of 1.17 and a root mean square error of 1.43.
• The model demonstrated superior prediction accuracy and adaptability compared to CEEMDAN-BiLSTM, EMD-BiLSTM, and BiLSTM alone.
• Friedman's test confirmed the model's efficiency and lack of over-modeling.

Conclusions:

• The CEEMDAN-BO-BiLSTM model is a feasible and effective approach for monthly average temperature prediction.
• This integrated model offers improved forecasting performance for climate-related applications.
• The findings provide valuable insights for enhancing climate forecasting strategies.