In the context of escalating electricity demand and the critical challenge of electricity theft in smart grids, this paper presents a novel dual-attention fusion transformer model, termed DaFT, for improved electricity theft detection. By effectively integrating both horizontal and vertical attention mechanisms, our approach captures the multidimensional periodicity inherent in electricity consumption data, with a particular focus on weekly and day-specific patterns. Utilizing a comprehensive dataset from the State Grid Corporation of China, we rigorously analyze consumption behaviors to discriminate between legitimate users and electricity thieves. Experimental results reveal that DaFT outperforms state-of-the-art methods in terms of Mean Average Precision (MAP) and Area Under the Curve (AUC) metrics, achieving a MAP@100 of 0.988 and an AUC of 0.827 under optimal training conditions. Our results highlight the eligibility of the scheme to address the pressing issue of electricity theft, thereby contributing to the security and sustainability of smart grid systems. This work not only advances the field of electricity theft detection, but also opens avenues for applying similar methodologies to other time-series data analysis.