<p>Accurate and computationally efficient wave height prediction is critical for marine operations, coastal engineering, and offshore safety. However, existing models struggle to capture both short- and long-term dependencies in complex sea states. This study presents a hybrid forecasting framework for single-step wave height prediction that combines Long Short-Term Memory (LSTM) networks and Transformer encoders. To further refine the global dependencies captured by the Transformer, a sparse attention mechanism is integrated not to replace the internal self-attention calculation, but to serve as a feature-selective filter. This mechanism identifies and preserves only the most salient temporal correlations, effectively mitigating the noise introduced by redundant global information while maintaining high prediction accuracy. To further enhance reliability under extreme conditions, a knowledge-based expert system is incorporated, consisting of 21 domain-specific rules addressing physical constraints, safety thresholds, and control logic. These rules refine the model’s outputs through a confidence-weighted fusion approach. The model is evaluated using five years of half-hourly wave height data from the Muluraba buoy station.Experimental results are divided into two parts: a comparative study with six benchmark models and a detailed ablation analysis. Experimental results demonstrate that the proposed hybrid system outperforms classical statistical models (ARIMA), machine learning baselines (SVR), and state-of-the-art deep learning architectures (Informer, CNN-LSTM), achieving a correlation coefficient of 0.97, RMSE of 0.027, and MAE of 0.019. These results demonstrate the effectiveness of integrating neural networks with expert systems for robust wave forecasting.</p>