Accurate interpretation of wind turbine operational measurements is essential for reliable wind power forecasting and efficient wind farm operation. Modern wind turbines are equipped with supervisory control and data acquisition (SCADA) systems that continuously record key operational parameters, providing rich measurement data for data-driven modeling. However, although deep learning methods have demonstrated strong capability in capturing nonlinear relationships, most existing approaches lack physical interpretability, show limited robustness under extreme conditions, and often neglect fundamental aerodynamic constraints. To address these limitations, this paper proposes a physics-constrained graph attention network (PCGAT) for multi-horizon mid-to-long term (from 6 h to 3 d) wind turbine power forecasting using SCADA measurement data. The proposed model constructs a temporal graph in which each node represents a measurement time step and connects to its preceding neighbors to capture dynamic dependencies, while a multi-head graph attention mechanism extracts informative representations from the graph-structured time series. A hybrid loss function incorporating aerodynamic constraints is introduced to ensure physical consistency during model training. Experiments on real-world SCADA datasets demonstrate that the proposed approach achieves higher forecasting accuracy and stronger robustness than several state-of-the-art models across multiple prediction horizons.
Menglong Wu, Xiaotian Zhang, Wenfei Liu et al.· Measurement science and tech...· 0 citations
Rolling bearings are critical components in rotating machinery, and their failures may cause unexpected downtime and safety risks. However, conventional deep diagnostic models are often difficult to deploy on resource-constrained edge devices because of their high computational cost and memory consumption. This paper proposes a lightweight time–frequency attention network (LTFANet) for multi-fault diagnosis of rolling bearings on an edge platform. The proposed model directly processes one-dimensional vibration signals and employs multi-scale depthwise separable convolutions to capture impact and periodic fault features with low computational complexity. A lightweight frequency branch is introduced to enhance fault-frequency representation, while an efficient channel attention module adaptively emphasizes fault-sensitive features. Moreover, a severity-aware multi-task extension is introduced to jointly identify the fault location and degradation level. To further improve edge inference efficiency, knowledge distillation, structured pruning, and TensorRT-based acceleration are integrated into the deployment pipeline. Experiments on CWRU-10 and Paderborn achieve 97.20% and 90.25% accuracy, respectively, while LTFANet contains only 0.020 M parameters and requires 0.610 M FLOPs. Knowledge distillation increases the CWRU-10 accuracy to 98.50%, and the severity-aware extension achieves 95.18% severity accuracy. On the NVIDIA Jetson Nano, the pruned TensorRT FP16 implementation achieves an average inference latency of 0.520 ms and a throughput of 1923.08 samples/s. The framework provides an effective solution for real-time and low-cost bearing condition monitoring at the edge.
