%0 Journal Article
%A ZHANG Dai-feng
%A CUI Dong-wen
%T Prediction of Daily Inflow Runoff of Three Gorges Reservoir Using Regularized Extreme Learning Machine Optimized by Three New Swarm Intelligent Algorithms
%D 2024
%R 10.11988/ckyyb.20230272
%J Journal of Changjiang River Scientific Research Institute
%P 16-24
%V 41
%N 7
%X Accurate prediction of daily inflow runoff is crucial for optimizing reservoir operation. To enhance the precision of daily inflow runoff forecasting, a prediction model integrating Wavelet Packet Transform (WPT), Dung Beetle Optimizer (DBO),Giant Trevally Optimizer (GTO), and Mud Ring Algorithm (MRA) optimized Randomized Extreme Learning Machine (RELM) is proposed and applied to forecasting daily inflow runoff in the Three Gorges Reservoir. Initially, WPT is utilized to decompose the daily runoff time series into a periodic term component and a fluctuation term component. Subsequently, by employing DBO, GTO, and MRA to optimize the input layer weights and hidden layer bias of RELM, the WPT-DBO-RELM, WPT-GTO-RELM, and WPT-MRA-RELM models are established. These models are then employed to predict and reconstruct the periodic and fluctuation components of daily inflow runoff. Comparative models such as WPT-DBO-ELM, WPT-GTO-ELM, and WPT-MRA-ELM based on Extreme Learning Machine (ELM), as well as WPT-DBO-BP, WPT-GTO-BP, and WPT-MRA-BP based on BP neural network, along with unoptimized WPT-RELM, WPT-ELM, and WPT-BP models and undecomposed DBO-RELM, GTO-RELM, and MRA-RELM models are utilized for analysis. Results indicate that: 1) The mean absolute percentage error (MAPE) of WPT-DBO-RELM, WPT-GTO-RELM, and WPT-MRA-RELM models for predicting the daily inflow in Three Gorges Reservoir is 0.512%, 0.519%, and 0.762% respectively, with Mean Absolute Error (MAE) of 54.05 m<sup>3</sup>/s, 55.97 m<sup>3</sup>/s, and 86.76 m<sup>3</sup>/s, Root-Mean-Square Error (RMSE) of 84.99 m<sup>3</sup>/s, 84.81 m<sup>3</sup>/s, and 128.18 m<sup>3</sup>/s, a determination coefficient ≥0.999 4, Theil Inequality Coefficient ≤0.005 17, showing superior prediction accuracy and generalization ability when compared to the other 12 models. 2) DBO, GTO, and MRA effectively optimize the parameters of RELM networks and enhance prediction performance. 3) Incorporating a regularization term in RELM prevents overfitting, boosts model generalization ability, and outperforms ELM and BP networks. 4) The proposed models exhibit high prediction accuracy, low computational complexity, proving to be efficient for estimating daily inflow runoff time series.
%U http://ckyyb.crsri.cn/EN/10.11988/ckyyb.20230272