Electricity theft in smart grids poses a significant threat to energy security, leading to billions in financial losses and grid instability worldwide. Traditional detection methods, including hardware-based solutions and machine learning (ML) models, are often costly, reliant on labeled data, and lack scalability. Deep learning (DL) approaches, while more advanced, face challenges such as overfitting to static datasets and inefficiency in adapting to evolving consumption patterns and new cyberattacks, requiring frequent and computationally expensive retraining. In this context, we propose a novel deep learning framework, Deep Adaptive Feature Learning for Theft Detection (DAFL-TD), tailored for smart grid environments. The architecture of DAFL-TD integrates a Temporal Feature Extraction Network (TFEN), which captures temporal dependencies in electricity usage, with an Adaptive Feature Learning Network (AFLN) that leverages both labeled and unlabeled data for adaptive feature extraction and classification. The novelty of DAFL-TD lies in its ability to handle fluctuating, imbalanced data and dynamically update its feature representation without the need for extensive retraining, making it highly scalable for real-world smart grid applications. Extensive evaluations on the State Grid Corporation of China (SGCC) dataset demonstrate that DAFL-TD achieves a 13.84% improvement in AUC compared to state-of-the-art models, alongside superior precision as measured by MAP metrics. These results underline the efficacy of DAFL-TD as a robust, scalable, and efficient solution for real-time electricity theft detection, significantly enhancing the resilience and security of smart grids.