Multi-step forecasting of dissolved oxygen in River Ganga based on CEEMDAN-AdaBoost-BiLSTM-LSTM model

Area of Science:

• Environmental Science
• Water Resource Management
• Data Science

Background:

• Accurate Dissolved Oxygen (DO) prediction is crucial for effective water resource management.
• Non-linear and non-stationary characteristics of DO data pose challenges for traditional forecasting models.

Purpose of the Study:

• To develop and evaluate a novel hybrid model for multi-step DO forecasting.
• To improve the accuracy of DO predictions by leveraging data decomposition and ensemble learning techniques.

Main Methods:

• Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to decompose DO data into Intrinsic Mode Functions (IMFs).
• Hybrid prediction approach: AdaBoost with Bidirectional Long Short-Term Memory (BiLSTM) for high/medium-frequency IMFs, and standalone Long Short-Term Memory (LSTM) for low-frequency IMFs.
• Model validation using data from ten stations of the river Ganga, compared against various baseline and decomposition-based models.

Main Results:

• The proposed CEEMDAN-AdaBoost-BiLSTM-LSTM model demonstrated superior performance in DO forecasting.
• Significant improvements observed: 25.458% RMSE and 37.390% MAE reduction compared to CEEMDAN-BiLSTM.
• Further improvements noted: 20.779% RMSE and 28.921% MAE reduction compared to CEEMDAN-AdaBoost-BiLSTM.
• Statistical tests (Diebold-Mariano and t-test) confirmed the significant performance advantage of the proposed model.

Conclusions:

• Tailored prediction strategies for different frequency components (IMFs) enhance forecasting accuracy.
• The CEEMDAN-AdaBoost-BiLSTM-LSTM model offers a robust and accurate solution for multi-step DO prediction.
• This approach provides a valuable tool for optimizing water resource management and environmental monitoring.