<p>Human Activity Recognition (HAR) plays an important role in healthcare, smart environments, and human-computer interaction. However, Current methods face challenges such as unbalanced data, irrelevant features, and complex movement patterns that reduce accuracy. To address these issues, we propose GCTnet, a deep learning architecture that combines Generative Adversarial Networks (GANs) for class balancing, Convolutional Neural Networks (CNNs) for spatial feature extraction, and Transformers with self-attention to model long-term dependencies in activity sequences. The architecture also incorporates Three-Way Clustering (3WC) to filter GAN-generated synthetic samples before model training, ensuring that only high-quality data is used for training, which enhances the robustness and diversity of the model. GCTnet is designed for deployment in scenarios where real-time HAR is needed, such as on edge devices like smartphones, wearables, and IoT sensors. The model has been tested and evaluated on multiple benchmark HAR datasets, including UCI HAR, WISDM, and PAMAP2, providing a comprehensive validation of its performance across diverse real-world scenarios. GCTnet is optimized for deployment on edge networks, enabling real-time, on-device processing, thereby reducing latency and bandwidth usage while improving data privacy. Experimental results demonstrate that GCTnet achieves 95.79% accuracy on the UCI HAR dataset, outperforming traditional models like CNN+LSTM (93.51%) and more advanced architectures like CNN+LSTM with Attention. Additionally, GCTnet outperforms state-of-the-art models on other datasets such as WISDM and PAMAP2, with accuracy improvements of up to 3-5%. The proposed model is tested on multiple datasets and validated for its generalizability across a wide range of activity recognition tasks, making it an ideal solution for efficient HAR applications on edge devices in wireless networks.</p>