This study presented a novel method of combining two NILM feature groups to utilize the complementary nature of the two feature groups and reduce feature size. We proposed a method of extracting 6 standard physical features (like RMS current, RMS voltage, etc.) along with 4 geometric features derived from the V-I trajectory shape and then integrating them into a 10-element feature vector. The method was evaluated on the public PLAID and WHITED datasets by deploying a Random Forest model, achieving classification accuracies of 97.82% and 97.69%, which are similar to the results of image-based methods. By using the proposed feature set and simple ML model, the computational cost was significantly reduced while maintaining comparable efficiency and prediction accuracy. This approach has lower requirements for memory storage and computational resources than image-based methods, which makes it more practical in many edge-based NILM applications.