Abstract:

[Objective] This study took Hanjiang River Basin as the study area. To better monitor the runoff conditions in Hanjiang River Basin, the daily runoff data collected from Ankang and Baihe hydroelectric power stations were selected for prediction analysis. The original data included daily runoff from January 2005 to December 2012. [Methods] This study first employed Multivariate Variational Mode Decomposition(MVMD) to decompose the original daily runoff data from the two stations, reducing data complexity. Subsequently, the decomposed modes and the historical runoff data from the previous 7 days were reconstructed using the Pearson correlation coefficient method(used to measure inter-variable correlation). The modes with high correlation coefficients were superimposed and defined as fluctuation terms, while those with low correlation coefficients were superimposed and defined as random terms. For the prediction of fluctuation terms, the historical runoff from the previous 7 days was used as input, resulting in seven operating conditions. Then, the Microbial Enhanced Algorithm-Back Propagation(MEA-BP) model was used for multiple predictions, and the average values were taken, and evaluation indicators were employed to assess the seven operating conditions. For the prediction of random terms, the Grey Wolf Optimizer-Extreme Learning Machine(GWO-ELM) was used for multiple predictions, and the average values were taken, and evaluation indicators were also used for assessment. Finally, the predicted results were fused, and evaluation coefficients were derived using evaluation indicators, demonstrating the accuracy and stability of the model. [Results] For Ankang station, IMF1 and IMF5 showed correlation coefficients greater than 0.5 with R₁-R₇, indicating high correlation. Therefore, IMF1 and IMF5 were reconstructed as fluctuation terms. IMF2, IMF3, IMF4, and IMF6, with correlation coefficients all below 0.5 with R₁-R₇, were reconstructed as random terms. Similarly, for Baihe station, IMF1 and IMF5 had correlation coefficients exceeding 0.5 with R₁-R₇ and were reconstructed as fluctuation terms, while IMF2, IMF3, IMF4, and IMF6, with correlation coefficients all below 0.5 with R₁-R₇, were reconstructed as random terms. For the prediction of fluctuation terms, the seven operating conditions were specifically defined as: R₁,R₁-R₂,R₁-R₃, R₁-R₄,R₁-R₅, R₁-R₆,and R₁-R₇. The coefficients of determination(R²) for these seven conditions of fluctuation term prediction at Ankang station were 0.54, 0.73, 0.74, 0.72, 0.81, 0.73, and 0.60, respectively, while those at Baihe station were 0.65, 0.68, 0.72, 0.77, 0.82, 0.74, and 0.77, respectively. The optimal operating condition for both stations was condition 5(R₁-R₅). For the prediction of random terms, the R² for random term prediction at Ankang and Baihe stations was 0.80 and 0.74, respectively. Finally, the integrated prediction combining fluctuation and random terms under condition 5 yielded R² of 0.87 and 0.93 for the overall prediction at Ankang and Baihe stations, respectively, demonstrating excellent model performance. [Conclusions](1) The MVMD decomposition method can control the number of decomposition layers, ensuring complete signal feature extraction without overfitting while improving processing speed.(2) Pearson correlation coefficient method enhances prediction accuracy through decomposed data classification.(3) The MEA-BP can improve signal-to-noise ratio, adapt to complex environments, enhance learning efficiency and generalization ability, and reduce computational complexity.(4) The GWO-ELM algorithm integrates grey wolf optimizer with extreme learning machine, providing a fast and adaptive solution for time-series prediction with reduced overfitting and improved efficiency.(5) The overall combined model can efficiently and stably process large amount of data while ensuring high accuracy.

Key words: multivariate variational mode decomposition, correlation reconstruction, mind evolutionary algorithm, BP neural network, grey wolf optimizer, extreme learning machine algorithm