Shared bikes represent a crucial component of urban transportation. The randomness of user demand for shared bikes with fixed piles leads to unbalanced demand in time and space, and even the difficulty in renting a bike, which cannot meet the user demand during peak hours. Therefore, high-frequency users frequently travel to nearby stations to rent a bike for serving, which means that there are implicit demands. As for the hidden demand, firstly, the state changes of the site were described by the rental number and the return number, and the critical state of the reference site was determined by mining the user travel conditions of nearby sites. The hidden demand of the site was determined based on the site state change diagram and the demand judgment model. Then, according to the real needs of the site, the long short-term memory(LSTM) network prediction model was established, and the regional scheduling model of shared bicycles based on the real needs was established. The model takes the cost minimization as the goal, and obtains the path with minimum scheduling cost through genetic algorithm, which provides a reference for balanced scheduling based on real demand. The results demonstrate that, when transportation costs are similar, the scheduling method under real demand can alleviate the problem of users’ difficulty in renting a bike, thereby reducing the loss of high-frequency users.