Architecture-Level Fusion of SCADA and ERA5 Reanalysis Data for Enhanced Winter Fault Detection in Offshore Wind Turbines

Abstract

Wind energy is a crucial part of the global shift towards low-carbon energy systems; however, the economic uncertainties of offshore wind energy remain driven by high offshore operation and maintenance (O&M) costs in harsh environments. Predictive maintenance is vital in offshore operations to minimize downtime and avoid catastrophic failures. But current techniques are mainly based on supervisory control and data acquisition (SCADA) data, which is susceptible to environmental effects and can become unreliable in winter. This thesis assesses the benefit of incorporating weather data for anomaly detection in offshore wind turbines during winter conditions. Using the CARE to Compare benchmark dataset, a three step approach is proposed. First, a SCADA-based one-dimensional neural network (1D-CNN) is reproduced to serve as a baseline. A data-level fusion model is then used to increase SCADA inputs using meteorological information produced by ERA5. Lastly, an architecture-level fusion model is developed, which relies on two parallel branches with a global gating mechanism to control the weight of environmental information. The performance of the models is evaluated on the full test set and the winter subset (December February). On the winter subset, the SCADA-only baseline has a high recall (0.9500) and low precision (0.6472), which produces a significant number of false alarms. Under the same winter evaluation, data-level fusion also enhances precision (0.8126) and reduces recall (0.8249), so fusion is more conservative in its detection behaviour. The proposed architecture level fusion model further achieves optimum performance with a precision of 0.7995, a recall rate of 0.9234 and the highest F1-score of 0.8570 and reduces false positives by over 55 percent as compared to the baseline. These results show how weather context can improve predictive maintenance in winter and can be integrated in effective ways to achieve better predictive maintenance performance. In terms of operational implications, lowering false alarms facilitates maintenance scheduling by reducing wastage of resources on false alarms. Finally, the research exports the model's detection of winter abnormalities as data inputs to an Operations Research optimization model, bridging the gap between predictive maintenance and maintenance operations in offshore wind farms.