Smart grids rely on Automatic Generation Control (AGC) systems to maintain frequency stability and regulate power exchange between interconnected areas. However, their increasing interconnectivity creates critical cybersecurity vulnerabilities. This paper proposes a novel integrated approach for detecting False Data Injection Attacks (FDIAs) in two-area AGC systems, where attackers manipulate sensor measurements or control signals to deceive operators. Our three-component methodology combines: (1) an autoencoder for anomaly detection through reconstruction error analysis, (2) an Extreme Learning Machine (ELM) classifier for identifying seven attack categories (pulse, ramp, scale, step-up/down, and coordinated variants), and (3) t-distributed Stochastic Neighbor Embedding (t-SNE) for enhanced interpretability through dimensionality reduction and visualization. This creates an intuitive visual map where similar attack patterns cluster together, simplifying identification for system operators. Experimental results demonstrate exceptional performance with 97.29% classification accuracy, low undetected attack risk (0.152), and minimal false alarm rate (0.0017). These metrics significantly outperform traditional threshold-based methods (80–85% accuracy) and recent single-technique machine learning approaches (90–95% accuracy). Our false alarm rate is approximately 3–5 times lower than comparable methods—critical in operational environments where false alarms erode trust. Nyquist stability analysis reveals differential impacts across attack types, with scale and step-down attacks severely compromising system stability (gain margin 0.39). The framework's effectiveness stems from strategically integrating complementary techniques: the autoencoder's unsupervised learning, ELM's computational efficiency, and t-SNE's visualization power, creating a robust defense against increasingly sophisticated cyberattacks targeting critical power infrastructure.