This study proposes an AI-based real-time fatigue detection and warning system for wind power O&M worker safety. Wind turbine maintenance operations involve prolonged, repetitive tasks in confined spaces at heights exceeding 80 m, which inevitably lead to worker fatigue accumulation and pose significant risks of severe industrial accidents. To overcome the limitations of subjective self-reporting methods and intrusive wearable sensor-based approaches, a camera-based non-invasive fatigue detection system was implemented on a low-cost embedded platform. The system incorporates an image preprocessing technique (InvertL) for low-light conditions, a multi-model ensemble-based fatigue classification algorithm combining rule-based and KNN approaches through soft-voting, a state transition logic integrating a finite state machine with hysteresis, and a multithreaded real-time processing architecture. Experimental validation demonstrated that the proposed system achieves an F1-score of 0.85, a false alarm rate of 0.71%, and an average processing speed of 25 FPS, confirming its technical feasibility as a worker fatigue monitoring system for wind power O&M environments.

This study proposes a hybrid wind power prediction model that combines a normalized wind speed-based power curve with residual learning to reduce prediction errors associated with air density variations and to compensate for residuals that cannot be explained by the power curve alone. First, measured SCADA wind speeds were converted into normalized wind speeds based on the air density correction concept specified in IEC 61400-12-1 using ERA5 temperature and atmospheric pressure data. A physical power curve, PC( ), was then created using the normalized wind speed. Subsequently, the difference between the actual power output and PC(Vn) was defined as the residual, and a Random Forest-based residual learning model was developed using normalized wind speed, temperature, and atmospheric pressure as input variables. The final power output was obtained by combining the physical power curve prediction with the predicted residual. Analysis of residual bias and variance across wind speed bins showed that the application of normalized wind speed reduced structural errors, particularly in the medium wind speed region, compared with the conventional power curve based on measured SCADA wind speed. Furthermore, the residual learning model effectively reduced the overall dispersion of prediction errors. Performance evaluation demonstrated that the proposed Physical PC( ) + Residual RF model achieved reductions of 24.4%, 30.2%, 24.4%, and 30.3% in RMSE, MAE, MAPE, and wMAPE, respectively, compared with the Physical PC( ) model without residual correction. In addition, temperature and atmospheric pressure were found to provide additional explanatory power during the residual learning stage even after being used in the normalized wind speed calculation. The results indicate that the proposed hybrid framework successfully combines the interpretability of a physics-based power curve with the predictive capability of machine learning-based residual correction. Therefore, the proposed approach can be effectively applied to wind farm power prediction and performance analysis.

This study aimed to establish an efficient and accurate power generation forecasting algorithm using machine learning to provide practical and actionable insights for power plant operators. The analysis used meteorological data and actual power generation data measured in a real-world wind farm located in South Korea. This study proposes a stacking ensemble model that combines density-based spatial clustering of applications with noise (DBSCAN)-based data cleaning and physical feature transformation to resolve the uncertainty of wind power data and maximize prediction precision. First, the DBSCAN clustering algorithm was applied to eliminate abnormal noise on the power curve. To reflect the physical laws of wind power generation, a Log V³ transformation was performed to incorporate the cubic relationship between wind speed and power output. Subsequently, a stacking ensemble architecture was designed using Random Forest, XGBoost and Linear Regression as base models and another Random Forest as the meta-model. Experimental results demonstrated that the proposed model achieved a significant reduction in error by decreasing the MAPE. This study demonstrates that the engineering design of data preprocessing plays a critical role in improving model reliability and is expected to serve as a practical forecasting framework for virtual power plants (VPP) and energy management systems (EMS). To verify the robustness of the proposed model, further investigation is required to analyze errors associated with the use of numerical weather prediction (NWP) data as input for day-ahead forecasting.

This study evaluates the performance of different long-term reference datasets for measure-correlate-predict (MCP) at an offshore wind site on the southwest coast of Korea. Short-term wind measurements from the HeMOSU-1 met mast were combined with ERA5 reanalysis, the high-resolution EMD-WRF mesoscale model, and near-surface winds from a KMA offshore light beacon, and long-term corrected wind speeds were derived over a representative one-year overlap within a 20-year reference period. The KMA light beacon showed the highest direct correlation with the met-mast data, but its lower height and coastal influence limited its ability to represent the offshore wind field, so MCP-corrected winds based solely on this reference can deviate substantially from actual hub-height conditions. By contrast, the reanalysis and mesoscale model dataset better captured the offshore climatic characteristics and yielded higher MCP accuracy. The results highlight that, beyond simple scalar correlations, selecting a reference whose sector-wise distributions of wind-speed frequency and total wind energy are consistent with those at the target site is important to avoid systematic bias in long-term MCP estimates.

The IEC 61400-1 Normal Turbulence Model (NTM), originally developed based on relatively flat European terrain, may not adequately represent the turbulence characteristics of complex mountainous terrain prevalent in South Korea, where approximately 70% of the land area consists of mountainous regions. This study investigates this discrepancy by analyzing turbulence characteristics measured at a lattice type mast installed in mountainous complex terrain in Gangwon Province, South Korea, over approximately 21 months (September 2023 to June 2025). The measured turbulence intensity (TI) was systematically compared with the IEC 61400-1 Ed.4 NTM design criteria. The results showed that, contrary to the NTM's assumption of decreasing TI with increasing wind speed, the measured mean TI increased from 0.144 at 5 m/s to 0.167 at 10 m/s, exhibiting a pattern fundamentally different from the NTM prediction. It was found that the measured 90th percentile TI exceeded the IEC C class above 5m/s, the B class above 7m/s, and the A class above 9m/s, and even the A+ class at 11-14m/s. These findings demonstrate a clear discrepancy between the IEC NTM and actual turbulence conditions in Korean mountainous terrain, and suggest that more conservative, site-specific turbulence design parameters are required for wind turbine installations in such regions.