<p>Multicolor soliton dynamics is essential for understanding the soliton interactions in fiber lasers and their triggered nonlinear effects, such as spectral modulation and the switching of soliton states. When studying complex and unsteady dynamics of multicolor solitons in the passive mode-locked fiber laser, not only traditional numerical methods based on the nonlinear Schrõdinger equation require high computational costs and are inefficient, but also single neural network models struggle to capture key dynamic features, such as phase modulation and energy fluctuations during soliton collisions. By effectively extracting the spatial characteristics of soliton interactions and capturing their pulse and spectral evolution, this paper proposes a dual-channel convolutional-recurrent neural network model. This dual-channel model processes the real and imaginary components of the optical complex field data, and accurately predicts the transient evolutionary characteristics of two-color and three-color solitons in the unsteady state, steady state, and their transition process, while capturing key nonlinear dynamical phenomena such as soliton collisions and energy redistribution. Compared with recurrent neural network, this combined model reduces the normalized root mean square error in predicting the positions of soliton collisions by approximately 39%, demonstrates excellent performance in multidimensional dynamic analysis, and offers new tools and insights for optimizing the design of fiber laser.</p>