In recent years, Industry 4.0 has introduced a shift in power grid infrastructures into smart grids (SG) that embed advanced communication, sensing, and analytics capabilities. Cyberattacks may cause outages or blackouts, leading to safety hazards, human casualties, and costly restoration efforts. In this article, using only data from the electrical activity of a smart grid, we present a model based on convolutional neural networks (CNN) that effectively distinguishes a wide variety of cyberattacks from normal operational events and natural faults. Our experiments show that the CNN outperforms conventional machine learning classifiers in all performance metrics, achieving 93% F1-score, compared to 88% for the best baseline. This network-agnostic, ML-based energy auditing approach offers a potentially easy to implement and cost-effective solution for various industrial environments while providing a reactive threat detection capability that may remain resilient even when adversaries target the network infrastructure or replay fault patterns.