Design of an algorithm for cleaning big data of electric energy in Energy Internet based on heterogeneous computing and low-rank model

To better address the processing requirements of complex datasets and accurately eliminate redundant information and noise, this study proposes a cleaning algorithm for power and energy big data in the Energy Internet that integrates heterogeneous computing with low-rank modeling. The research develops a CPU–GPU heterogeneous computing framework within the multi-heterogeneous Energy Internet architecture to enhance the computational efficiency of power–energy big data processing. Within this framework, entity recognition is performed on power–energy big data through heterogeneous record analysis, followed by anomaly detection using low-rank modeling to distinguish between normal and abnormal entity data. The algorithm implements selective processing: insignificant anomalies with negligible impact on overall data analysis are automatically discarded, while known anomalies are systematically repaired using a correlation-based multidimensional time-series data correction algorithm, thereby completing the comprehensive cleaning process. Experimental results demonstrate that the proposed algorithm effectively identifies entities in Energy Internet power data while successfully detecting, repairing, and cleaning abnormal entity data. When processing 500 GB of power big data, the algorithm completed cleaning in just 8.2 s, maintaining a post-cleaning data quality (measured by the proportion of correctly recorded data) above 0.96. Even with high proportions of missing data, it maintained a near-perfect quality value close to 1. Comparative analysis confirms the algorithm’s exceptional data cleaning performance across power datasets with varying missing data ratios, establishing its robustness and reliability for practical Energy Internet applications.