Enhancing transformer-based wind speed forecasting on new meteorological and turbine-level SCADA datasets

Accurate wind speed forecasting is crucial for mitigating grid instability and supply–demand imbalances caused by the intermittent nature of wind power. To support reproducible research under heterogeneous sensing conditions, we construct and publicly release three complementary datasets: KZMET, a 15-year reanalysis-based meteorological dataset covering 25 cities in Kazakhstan; VNMET, a multi-year mast-based wind measurement dataset from 10 locations in Vietnam; and WTSL, a high-resolution (1-minute) turbine-level SCADA dataset containing over 17 million records and labeled operational alarms. Together, these datasets cover three practical forecasting scenarios: large-scale meteorological benchmarking, field measurement-based validation, and turbine-level operational forecasting. Addressing these heterogeneous forecasting scenarios requires models that can capture both long-range temporal dependencies and fine-grained local nonlinear patterns. Although Transformers have shown strong performance in time-series forecasting, many existing Transformer-based approaches still rely on linear embedding layers, which may limit their ability to represent local temporal structure and nonlinear feature interactions in wind data. Motivated by the heterogeneous characteristics of the released datasets, we propose a nonlinear embedding layer, termed CME, which combines a 1D convolutional neural network to extract local temporal patterns and a two-layer multilayer perceptron to model nonlinear feature interactions. Integrated into a patch-based Transformer, this design forms the CME-Patchformer architecture. We evaluate CME-Patchformer on the three datasets across multiple forecasting horizons against 12 baseline models. Results show that the proposed model achieves competitive performance, improves MAE and RMSE in many settings, and provides a strong reference architecture for the released benchmarks with modest inference cost.