To address the challenges of strong nonstationarity, severe noise interference, and insufficient detection accuracy of arc fault signals under complex operating conditions, this paper proposes a series arc fault detection method that integrates optimized feature enhancement with physics-informed deep modeling. First, an improved dream optimization algorithm (IDOA) is introduced to adaptively optimize the parameters in wavelet energy spectrum reconstruction, enabling the enhancement of arc-sensitive frequency bands and improving signal representation capability. Subsequently, recurrence plot (RP) is employed to perform time-frequency mapping on the reconstructed signals, transforming one-dimensional nonstationary signals into two-dimensional structural features to characterize their nonlinear dynamic evolution. In the feature modeling stage, a time-frequency collaborative receptive field attention (TF-CRFA) mechanism is constructed to jointly model transient characteristics in the time domain and energy distribution in the frequency domain, thereby achieving adaptive enhancement of critical arc features. Furthermore, a physics-informed Transformer (PI-Transformer) is designed by incorporating physical priors, such as high-frequency energy mutation and nonstationarity of arc signals, into the model architecture and loss function, enabling unified global temporal dependency modeling and physical consistency constraints. Experimental results demonstrate that the proposed method outperforms comparative models in terms of detection accuracy, noise robustness, and adaptability to complex conditions and potential for engineering applications in arc fault detection.