Some magnetic features in active regions, related to strong solar flares, are considered as “anomaly” features in a machine learning algorithm. An unsupervised auto-encoder network has been trained to identify such anomalies and is used to predict occurrence of strong flares.