This study addresses a fundamental question in music psychology: which specific, dynamic acoustic features predict human listeners' emotional responses along the dimensions of valence and arousal. Our primary objective was to develop and validate an interpretable computational model that can serve as a tool for testing and advancing theories of music cognition. Using the publicly available DEAM dataset, containing 1,802 music excerpts with continuous valence-arousal ratings, we developed a novel, theory-guided neural network. This proposed model integrates a convolutional pathway for local spectral analysis with a Transformer pathway for capturing long-range temporal dependencies. Critically, its learning process is constrained by established principles from music psychology to enhance its plausibility. A core finding from an analysis of the model's attention mechanisms was that distinct acoustic patterns drive the two emotional dimensions: rhythmic regularity and spectral flux emerged as strong predictors of arousal, whereas harmonic complexity and musical mode were key predictors of valence. To validate our analytical tool, we confirmed that the model significantly outperformed standard baselines in predictive accuracy, achieving a Concordance Correlation Coefficient (CCC) of 0.67 for valence and 0.73 for arousal. Furthermore, an ablation study demonstrated that the theory-guided constraints were essential for this superior performance. Together, these findings provide robust computational evidence for the distinct roles of temporal and spectral features in shaping emotional perception. This work demonstrates the utility of interpretable machine learning as a powerful methodology for testing and refining psychological theories of music and emotion.