Interpretable machine learning for coastal wind prediction: Integrating SHAP analysis and seasonal trends

Abstract Accurate wind speed prediction plays an important role in developing effective coastal management strategies and risk assessments, especially in coastal region managements to reduce erosion damage. In offshore wind energy, precise forecasts optimize wind farm layout and operations, maximizing energy yield and minimizing downtime. Additionally, accurate wind speed forecasts significantly improve maritime transportation safety by predicting hazardous conditions. Understanding wind patterns is also important for coastal ecosystem management and safer navigation activities. However, accurate wind speed prediction in dynamic coastal environments remains challenging due to (1) limited applications of robust machine learning (ML) models tailored to coastal meteorological complexity, (2) insufficient integration of interpretable feature analysis with predictive modeling for actionable insights, and (3) gaps in understanding how seasonal and diurnal wind patterns influence model performance in understudied regions like tropical Queensland. This study focuses on Abbot Point, Queensland, Australia, using meteorological data collected hourly from January 1 to December 31, 2023 (Latitude: -19.9496; Longitude: 148.0482). It evaluates three machine ML models—Linear Regression (LR), Decision Tree Regressor (DT), and Random Forest (RF)—to identify the most reliable approach for wind speed forecasting. The dataset includes wind direction, air temperature, relative humidity, precipitation, and barometric pressure as feature variables, with wind speed as the target variable. Novel integration of SHapley Additive exPlanations (SHAP) analysis and seasonal decomposition addresses interpretability gaps, while rigorous validation across training (70%), testing (15%), and validation (15%) datasets ensures model robustness. The RF model consistently outperformed others across training, validation, and test datasets, achieving the lowest mean square error (MSE: Train 0.183, Validation 0.875, Test 0.803), highest R 2 (Train 0.966, Validation 0.831, Test 0.844), and superior Nash–Sutcliffe Efficiency (NSE: Train 0.96, Validation 0.83, Test 0.84). These results reflect the model's robust ability to capture complex relationships in the data. In contrast, LR and DT exhibited moderate accuracy, with higher MSE and lower NSE values, struggling particularly with consistency and extreme values. Complementary analyses, including wind rose plots and time series of wind speed, relative humidity, and barometric pressure, revealed high-risk periods characterized by strong winds (> 10 m/s), high humidity (> 90%), and low barometric pressure (< 1000 hPa). Seasonal analysis revealed spring/summer peaks in hazardous winds (> 10 m/s), with diurnal cycles (24-h periodicity) significantly influencing prediction accuracy—a pattern underemphasized in prior coastal ML studies. This study bridges critical gaps by demonstrating how interpretable ML enhances coastal wind prediction through: a) quantitative validation of RFR's superiority over traditional models in handling coastal meteorological variability, b) SHAP-driven identification of dominant predictors (wind direction, pressure) for targeted monitoring, c) Seasonal-temporal analysis framework for site-specific risk mitigation strategies. These findings confirm the interactions between meteorological variables that intensify storm risks and coastal hazards. Key insights include the dominant influence of southeast and south-southwest winds (100°–200°) and the critical role of barometric pressure in driving extreme wind events. Also, findings enable improved storm surge modeling and early warning systems by providing 6-h wind forecasts with 84% accuracy, directly informing coastal defense alignment with dominant wind-driven erosion patterns. This approach addresses the critical need for ML applications that combine predictive power with operational interpretability in coastal management contexts. The integration of ML models with detailed meteorological patterns supports the identification of high-risk periods, enabling targeted interventions such as strengthening coastal defenses and issuing early warnings. This study underscores the value of ML techniques, particularly RF, in enhancing predictive frameworks for coastal risk management and promoting sustainable, resilient coastal environments.