EWT多重分解与若干新型元启发式算法优化的多层感知器月径流预测

To improve the accuracy of monthly runoff time series prediction, enhance the performance of multi-layer perceptron (MLP), and compare and verify the optimization effects of four new metaheuristic algorithms on benchmarking functions and instance objective functions, including catch fish optimization algorithm (CFOA), flood algorithm (FLA), arctic puffin optimization (APO), educational competition optimization (ECO), and particle swarm optimization (PSO), an EWT^III-CFOA/FLA/APO/ECO/PSO-MLP prediction model was proposed. The model was validated through a monthly runoff time series instance at Mengda Hydrological Station. Firstly, EWT^Ⅰ was used to decompose the monthly runoff time series into fluctuation and trend terms. FuzzyEn was used to determine the complexity. EWT^Ⅱ and EWT^Ⅲ were used to decompose the more complex fluctuation terms. Secondly, based on the training sets of each component, MLP weight and bias (hyperparameter) were constructed to optimize the instance objective function. Six benchmarking functions were selected as comparative verification functions. The CFOA/FLA/APO/ECO/PSO algorithm was used to perform extreme value optimization and comparative analysis on the benchmarking function and instance objective function, respectively. Finally, the EWT^Ⅰ/EWT^Ⅱ/EWT^Ⅲ- CFOA/FLA/APO/ECO/PSO-MLP model was established to train, predict, and reconstruct each decomposed component. The results show that the EWT^III-FLOA/FLA/APO/ECO-MLP model is superior to other comparative models in fitting and prediction accuracy, with better prediction accuracy. The CFOA/FLA/APO/ECO/PSO algorithm has similar overall rankings for benchmarking function optimization, instance objective function optimization, and EWT^Ⅰ/EWT^Ⅱ/EWT^Ⅲ- FLOA/FLA/APO/ECO/PSO-MLP model prediction accuracy. As the optimization performance of the algorithm gets stronger, the MLP hyperparameter obtained through optimization gets better. The prediction accuracy of the EWT^Ⅰ/EWT^Ⅱ/EWT^Ⅲ-FLOA/FLA/APO/ECO/PSO-MLP model improves with the increase of EWT decomposition weight. As a concise and efficient time series decomposition method, EWT^III can decompose the original monthly runoff series into more easily modeled and predictable components with a larger scale.