LGD-LSTM prediction network with VMD for nonlinear system

Since random input noise can affect the prediction accuracy and stability of the network, based on variational mode decomposition (VMD) technique, a network model combining local and global diffusion (LGD) and long short-term memory (LSTM) is proposed to predict nonlinear systems with input noise. By decompressing input signals into intrinsic mode functions (IMF) of different frequencies and eliminating them at high frequencies, the VMD technology can effectively achieve signal denoising. By using different convolution kernels in the branch, LGD network can extract the local and global feature information in the de-noising time series in parallel to improve the learning ability of the network. The LSTM network with gated structure is further used to extract the long-term dependencies in time series. Finally, the superiority of the proposed algorithm is verified by numerical simulation through comparison with existing prediction networks.