Autism Spectrum Disorder (ASD) is a neurodevelopmental condition involving social communication difficulties and repetitive behaviors. EEG-based Computer-Aided Diagnosis (CAD) using machine learning offers a promising approach for early, objective ASD detection, though traditional models often lack transparency. This study introduces the Graphical Attention Spectral Temporal Convolutional Network (GASTCN), an explainable deep learning model that integrates temporal, spectral, and graphical EEG features for ASD classification. Spectrograms are generated via Short-Time Fourier Transform and processed through the proposed Diverse Selection Extraction (DSE) Block, which expands and compresses feature representations to diversify and refine time-frequency features. Inter-channel dependencies are modeled using attention guided by a learnable token, enabling the model to summarize and focus on the most relevant EEG channels. Under stratified group 5-fold cross-validation, GASTCN achieved an AUROC of 94.33% and recall of 96.10%, with standard deviations of 3.40% and 1.36%, respectively. Statistical analysis using two explainability metrics, label-adjusted SHAP and importance-weighted spectral power, identified the delta and alpha bands at the C3 electrode as key ASD biomarkers.