Understanding transition states is pivotal in the design of efficient chemical processes and catalysts. However, identifying transition states is challenging due to the resource-intensive and iterative nature of current computational methods. This study integrates neural network potentials with physical models to enhance the transition state prediction. Different neural network potentials and transition states locating algorithms are benchmarked. By combining NequIP with the energy-weighted Climbing Image-Nudged Elastic Band (EW-CI-NEB) method, we achieved highly accurate transition state predictions, significantly surpassing semiempirical methods in accuracy and greatly outpacing density functional theory in efficiency. Additionally, the transferability of the model was evaluated using a NequIP model trained on a refined subset of the dataset, and the model's performance was further improved through active learning. This method can directly search for transition states in given reactions or serve as an efficient tool for generating initial guesses of transition state structures, significantly reducing manual effort.