Abstract:To address the contradiction between high-concurrency processing and low-latency response of electric power measurement data, this paper proposes a distributed computing architecture oriented toward real-time measurement. By analyzing the characteristics of Flink and Spark Streaming, a scheduling model based on task priority and adaptive resource load is constructed to achieve dynamic resource management for streaming data. The model attains an event detection accuracy of 96.92%, a peak throughput of 132.5 GB/min, protocol processing capability of 3492 packets/s, and a 31.7% reduction in computing resource overhead. Actual tests show that the fault recovery time is shortened to 153.1 ms (minimum 116 ms), and data processing integrity reaches 98.77% (peak 99.9%). During 72 hours of continuous operation, the model maintained a throughput of 124.7 GB/min and a detection accuracy of 94.8%, with a comprehensive resource optimization rate of 31.7% (20.9% reduction in computational complexity and 48.1% reduction in memory overhead), verifying its robustness and elastic optimization capability under complex multi-voltage level conditions, and providing effective reference for building real-time analysis platforms in power systems.