Parallelization Strategies for the Feature Space Partition Algorithm Applied to Fault Detection and Stability Analysis in Smart Grids

• Saulo Andrade Almeida UFRJ
• Silvana Rossetto UFRJ
• Carolina Gil Marcelino UFRJ

Resumo

This work presents the parallelization of the supervised learning algorithm Feature Space Partition (FSP), originally implemented sequentially in Python. To address computational challenges, multiple parallelization strategies were explored, including implementations based on CPU and GPU, while preserving the original algorithmic flow to ensure that classification performance remained unaffected. Among the approaches tested, CPU-based multiprocessing yielded the most effective results, achieving a reduction of 36.44% in execution time without compromising the predictive accuracy of the model. FSP has demonstrated high classification accuracy, ranging from 80% to 99% in identifying fault patterns in distribution networks and stability patterns in smart grid systems. This reinforces the viability of using existing Python-based parallel processing libraries to scale machine learning algorithms for critical electrical power grid applications.