With the accelerated construction of new-generation power systems, smart terminals centered around autonomous and controllable embedded operating systems play a crucial role in ensuring the safe, stable, and efficient operation of the power grid. Tasks in power system terminals exhibit typical mixed-criticality characteristics, comprising both high-criticality safety-protection tasks and low-criticality monitoring and communication tasks. During system overload or transient faults, traditional Mixed-Criticality Scheduling (MCS) algorithms, to guarantee the determinism of high-criticality tasks, often sacrifice or even completely discard all low-criticality tasks. This practice severely undermines the situational awareness and data analysis capabilities of the smart grid, running counter to the development goal of grid intelligence. To address this issue, this paper proposes an Adaptive Fault-Tolerant Mixed-Criticality Scheduling (AFT-MCS) algorithm for power system services. While ensuring the hard real-time constraints of high-criticality tasks, the algorithm introduces an adaptive scheduling strategy based on task value, aiming to achieve“graceful degradation” of system performance. First, AFT-MCS designs a multi-level graceful degradation mechanism that employs multi-dimensional fault-tolerance measures, such as reducing task activation frequency and task migration, to replace the traditional task-dropping strategy, thereby maximizing the quality of service for low-criticality tasks. Second, the algorithm introduces a value-based dynamic promotion mechanism that adjusts task priority factors based on historical execution performance to prevent the loss of critical low-criticality data streams due to long-term resource starvation. Finally, this paper constructs a lightweight Q-Learning decision model to dynamically select the optimal scheduling action at runtime based on multi-dimensional system states, achieving online adaptive optimization of the scheduling policy.