A power load forecasting method in port based on VMD-ICSS-hybrid neural network

Aiming at the problem of load fluctuation at the power end of large ports, we propose a hybrid neural network joint model based on Mode Decomposition (MD) and Change Point Detection (CPD) to accomplish the load forecasting. In this study, a two-stage joint prediction model is constructed. First, the number of Intrinsic Mode Functions (IMFs) in the Variational Mode Decomposition (VMD) process was dynamically adjusted by introducing an improved Signal Energy (SE) evaluation metric. Subsequently, a Bidirectional Gated Recurrent Unit (Bi-GRU) network is employed to predict these IMFs, and the potential effect of the breakpoints on the prediction outcomes is investigated using the Iterative Cumulative Sum of Squares (ICSS) method. Finally, the eigenmode functions are summed and reconstructed, and then combined with the breakpoint data as inputs for the second stage prediction. To ensure the efficiency of the second stage prediction, the Mogrifier Long-and Short-Term Memory (Mogrifier-LSTM) network structure is improved. In the two-stage model, the adaptive tuning of hyperparameters is implemented by a Hunter-Prey Optimization (HPO) algorithm based on a redesigned chaotic mapping strategy. During the simulation, various neural network topologies were employed to confirm the effectiveness of the model in port power load forecasting.