Grid operation data evaluation model based on Keras-Spark with attention-optimized GRUs

To address the challenge of low efficiency in processing massive heterogeneous data for power grid operational state assessment, this study proposes an intelligent assessment framework integrating distributed computing and adaptive temporal modeling. It employs an in-memory computing-based parallel preprocessing pipeline to optimize high-dimensional data throughput. Combined with a gated recurrent unit's dynamic weighting mechanism, it enhances feature extraction for critical operational data (e.g., fault transients, load fluctuations), synergistically improving assessment efficiency and accuracy for massive data. Experimental results show the model reduces assessment response time from 32 seconds (traditional methods) to 4.7 seconds, achieves 95.3% transient disturbance identification accuracy, and increases steady-state deviation detection rate to 97.1%. Load prediction mean absolute error decreases by 2.8% and peak error by 5.2%. Model training requires only 8.3 hours, achieving a 12-fold speedup over single-machine environments and an 89% resource utilization increase. The proposed model effectively enhances the real-time capability and efficiency of massive power grid data assessment, providing a feasible solution for dynamic operational risk management.