Electrical power networks are heavily monitored systems, requiring operators to perform intricate information synthesis before understanding the underlying network state. Our study aims at helping this synthesis step by automatically creating features from sensor data. We propose a feature extraction approach using a grammar-guided evolution, which outputs interpretable and physically consistent features. Operations restrictions on physical units are introduced in the learning process through interactively-built context-free grammars. They ensure coherence with physical laws, dimensional-consistency and also introduce technical expertise in the created features. We compare our approach to other state-of-the-art feature extraction methods on a real dataset taken from the French electrical network sensors and evaluate interpretability on human and functional levels.