Non-Intrusive Load Monitoring (NILM) aims to disaggregate a system’s total energy consumption into individual device power consumption. While Transformer models have gained popularity for tackling NILM, their accuracy may decline with increasing meter data volumes. In this study, we propose a Dilated Convolutional Gated Linear Unit Feed-Forward Network (DGPFFN) to address limitations in traditional Transformer models, such as inadequate local feature extraction and computational inefficiencies. The gated linear unit introduces nonlinear interactions among features, thereby enhancing complex feature extraction. Concurrently, dilated convolution effectively captures local features often overlooked by Transformer models, thereby enhancing the accuracy of power estimation for household appliances. Simulation results using the UK-DALE public dataset demonstrate that optimizing the Transformer model with DGPFFN yields superior prediction accuracy compared to the original Transformer model.