This study addresses the challenge of 24-h ahead tropospheric ozone forecasting in the complex atmospheric basin of Seville, Spain. To capture the highly non-linear dynamics of photochemical pollutants, a hybrid CNN-LSTM-Attention framework is proposed, utilizing a 72-h sliding window in a multivariate approach. To explicitly prevent temporal data leakage and ensure robust evaluation, a purged walk-forward cross-validation scheme with fold-isolated preprocessing was implemented. Furthermore, novel optimization strategies, including Stochastic Weight Averaging (SWA) and Bayesian hyperparameter tuning, were applied during training to enhance structural robustness and convergence stability. Evaluated on an unseen 2023 testing set, deep sequence models demonstrated clear hierarchical superiority over traditional tree-based baselines. The CNN-LSTM-Attention SWA model achieved the highest performance ( \(R^2 = 0.699\) , MSE = 368.30), significantly mitigating the temporal lag and systematic amplitude damping observed in simpler architectures. Conversely, while XGBoost reduced computational overhead to a few minutes, it systematically underestimated ozone peak concentrations. Ultimately, the synergy between convolutional feature extraction, attention mechanisms, and SWA stabilization yields superior generalization. This establishes a robust foundation for early warning systems in Mediterranean environments, successfully capturing the amplitude and phase of diurnal ozone spikes where conventional models fail.