The reliable operation of photovoltaic (PV) systems requires effective methods for distinguishing anomalies from normal environmental variations. This study presents an Artificial Intelligence of Things (AIoT) framework for real-time PV monitoring and anomaly diagnosis. The proposed system integrates custom data acquisition, MQTT communication, edge computing, an LSTM Autoencoder, and Random Forest classification. Experimental validation was conducted under real operating conditions at three scales: a 500 W microinverter-based system, a 3.9 kWp grid-connected installation, and a 62.5 kWp PV subsystem. The evaluated conditions included normal operation, environmental variations, inverter clipping, shading, dust accumulation, and localized hotspot conditions. The Random Forest classifier achieved 99.02% test accuracy for the classification of seven operating conditions, while the hybrid strategy subsequently identified localized and progressive anomalies through inter-string comparison and persistence-based rules. The results demonstrate the feasibility of low-latency PV anomaly diagnosis using edge intelligence with reduced dependence on continuous cloud processing.