VMD-CNN-BiGRU water level prediction model based on dual optimization

To reduce the non-linearity and non-stationarity of water level data and to solve the problem of large prediction errors in water level forecasting, a dual-optimized hybrid prediction model, VMD-CNN-BiGRU water level prediction model was proposed.First, the Crowned Porcupine Optimization (CPO) algorithm was used to optimize the number of modal decomposition and the quadratic penalty factor of the Variational Mode Decomposition (VMD) to ensure the optimal settings of the decomposition process and obtain intrinsic mode functions that characterize the water level data features.Then the Sparrow Search Algorithm (SSA) was introduced to optimize the parameters of CNN-BiGRU (Convolutional Neural Network-Bidirectional Gated Recurrent Unit) model to improve its prediction capability.Finally, all the predicted component sequences were integrated to obtain the final water level prediction results.To verify the effectiveness of the proposed hybrid prediction model, measured data from the Chen erhe Station of Yangtze River were used, and the CPO algorithm, SSA algorithm, and Particle Swarm Optimization (PSO) algorithm were respectively applied to optimize the parameters of the VMD and CNN-BiGRU models for validation.Meanwhile, ablation experiments were conducted to verify the necessity of the hybrid model.The results showed that in the optimization of VMD and CNN-BiGRU parameters, the CPO algorithm achieves the best optimization effect for VMD with lower fitness values and better convergence performance, while the SSA algorithm performs best in optimizing CNN-BiGRU.In the ablation experiments, after introducing VMD decomposition into CNN-BiGRU, the Mean Absolute Error (M_AE), Mean Absolute Percentage Error (M_APE), and Root Mean Square Error (R_MSE) decrease by 33%, 27%, and 24%, respectively, while the coefficient of determination (R²) increases by 0.021;after further introducing the SSA and CPO optimization algorithms, the M_AE, M_APE, and R_MSE decrease by 63%, 64%, and 60%, respectively, and R² increases by 0.050.This further indicates that each module of the hybrid model is indispensable in the water level prediction task.