Aiming at the optimal scheduling of load-side resources in smart grids, this study proposes a multilevel collaborative scheduling framework based on a hierarchical market mechanism, which combines an improved hybrid genetic algorithm with a distributed model predictive control method to achieve the global optimization and dynamic response balance of load resources. First, a spot market load resource dispatch model based on dynamic cross-probability adjustment is constructed to reduce the total dispatch cost by optimizing the objective function, which covers the operating cost, network loss and power supply/demand balance. For the aggregation control demand of temperature-controlled load TCL, a distributed model predictive control LDMPC method based on Lyapunov function is designed to reduce the system delay by 9.19ms through localized communication, which is 25.6% lower than the traditional method, and at the same time, safeguard the comfort of users. In addition, an improved thermoelectric load resource allocation algorithm is proposed to coordinate the weight allocation of electric and thermal loads, which reduces the operating cost of the plant to 9860 RMB, saving 24.1% compared with the traditional algorithm. Simulation experiments show that the proposed method significantly outperforms the traditional scheduling strategy in terms of delay performance, integrated system output and cost optimization.