Explainable deep learning techniques for wind speed forecasting in coastal areas: Integrating model configuration, regularization, early stopping, and SHAP analysis
Ahmet Durap
Abstract
Abstract Accurate wind prediction is critical across engineering disciplines. For coastal infrastructure, it determines wave loads and storm surge resilience, directly impacting millions in vulnerable low-lying regions. The energy sector relies on precise forecasts to optimize wind farm output and stabilize power grids, while agriculture uses wind data to time pesticide applications and protect crops. Aviation and shipping industries leverage predictions for fuel-efficient routing and hazard avoidance, and urban engineers integrate wind models for skyscraper design and air pollution management. As climate change amplifies wind extremes, advancing predictive capabilities has become an urgent cross-sector priority for adaptive planning and risk mitigation. In coastal applications, empirical wave models (e.g., SWAN and WAVEWATCH III) heavily depend on accurate wind inputs, where errors can lead to underestimation of extreme events and compromise structural safety. This study introduces a novel deep learning framework, integrating advanced data preprocessing, structured neural networks, and explainable AI techniques, to enhance short-term (hourly) wind speed forecasting for coastal engineering applications, addressing the gap in region-specific deep learning frameworks for operational forecasting. The proposed method in this study addresses critical gaps in traditional methods by combining physical constraints with data-driven learning. It presents an innovative framework for wind speed data processing and prediction, integrating deep learning architectures with comprehensive meteorological analysis. Our research implements a sophisticated neural network model that processes high-frequency wind data from Bowen, incorporating multiple environmental parameters through a systematic data pipeline. The methodology encompasses three key components: (1) advanced data preprocessing, including time series standardization and cyclical feature encoding; (2) a deep learning architecture featuring three hidden layers (128-64-32 nodes) with ReLU activation and dropout regularization; and (3) comprehensive performance evaluation using five-fold cross-validation. The model achieved remarkable accuracy metrics: R 2 = 0.957, RMSE = 0.449 m/s, demonstrating robust performance across varying weather conditions. Analysis revealed distinct performance patterns across wind speed ranges (low-speed MAE: 0.295 m/s; high-speed MAE: 0.433 m/s). The SHAP (SHapley Additive exPlanations) analysis provided deeper insights into feature importance and model interpretability, revealing Wind Direction (0.713 SHAP value) as the most influential predictor, followed by Relative Humidity (0.609) and Barometric Pressure (0.563). Temporal features (month, hour, and day) exhibited lower but consistent influence (SHAP values < 0.239). This research advances the field of environmental data science by providing: (1) a reproducible framework for wind speed prediction, (2) insights into feature significance and model behavior, and (3) practical applications for renewable energy planning and meteorological forecasting. The demonstrated methodology offers a foundation for future research in environmental modeling and time series prediction.