Securing smart cities in the evolving Internet of Things (IoT) demands innovative security solutions that extend beyond conventional theft detection. This study introduces temporal convolutional networks and gated recurrent units (TCGR), a pioneering model tailored for the dynamic IoT-SM dataset, addressing eight distinct forms of theft. In contrast to conventional techniques, TCGR utilizes Jaya tuning (TCGRJ), ensuring improved accuracy and computational efficiency. The technique employs ResNeXt for feature extraction to extract important patterns from IoT device-generated data and Edited Nearest Neighbors for data balancing. Empirical evaluations validate TCGRJ's greater precision (96.7%) and accuracy (97.1%) in detecting theft. The model significantly aids in preventing theftrelated risks and is designed for real-time Internet of Things applications in smart cities, aligning with the broader goal of creating safer spaces by reducing hazards associated with unauthorized electrical connections. TCGRJ promotes sustainable energy practices that benefit every resident, particularly those with disabilities, by discouraging theft and encouraging economical power consumption. This research underscores the crucial role of advanced theft detection technologies in developing smart cities that prioritize inclusivity, accessibility, and an enhanced quality of life for all individuals, including those with disabilities.