The ability of smart homes, connected through Internet of Things sensors and devices, has changed daily life by offering convenience, efficiency, and personalization. These devices generate vast amounts of data from user actions and interactions, facilitating the transition from reactive, rule-based systems to proactive, intelligent ones capable of anticipating user needs. Action recommendation is one of these solutions, providing personalized suggestions or task automation for better smart home adaptability. This paper introduces enhanced hybrid sequential modeling for smart home environments that can make action suggestions based on changing conditions. The proposed system leverages temporally ordered data to explain the interrelations between user behaviors and contextual factors. To process data dynamically and better understand both short-term and long-term relationships, we suggest a model that combines temporal convolution networks with enhanced transformer blocks that are made better by relative positional encoding and global attention pooling. The evaluation on four benchmark datasets yields competitive performance with state-of-the-art methods, which indicates its potential to offer accurate personalized action recommendations.