• RWGF integrates wake-aware spatial priors with multi-step interval forecasting. • The dynamic wake graph is designed as a forecasting-oriented spatial prior. • Soft directional attenuation and wake superposition reduce rigid graph assumptions. • Horizon-wise regularisation improves interval-boundary stability across lead times. • Step-wise conformal recalibration improves empirical coverage consistency. As power systems integrate increasing shares of renewable energy, ultra-short-term wind farm power forecasting is required to provide not only accurate point estimates but also calibrated uncertainty information for short-term forecasting applications. However, wind farm forecasting remains challenging because inter-turbine dependence varies with wind direction, wake interaction and local operating conditions, while multi-step prediction intervals may become miscalibrated under non-stationary operating regimes. This paper proposes RWGF, a reliable wind farm power forecasting framework with dynamic wake-aware spatio-temporal coupling, for multi-step point and interval forecasting. In RWGF, a directed dynamic wake graph is constructed from turbine layout, real-time wind conditions and a simplified wake velocity-deficit model. The graph is used as a computationally efficient spatial prior rather than as a high-fidelity flow-field reconstruction, and is coupled with a multi-scale TCN–GRU module to learn lagged temporal dependence and non-stationary power evolution. To generate ordered and stable prediction intervals, the shared spatio-temporal representation is further combined with non-crossing quantile learning, horizon-wise width and boundary-smoothness regularisation, and step-wise conformal recalibration. Experiments on SCADA data from two real-world wind farms show that RWGF improves deterministic forecasting accuracy and provides a more balanced coverage–sharpness trade-off than representative temporal, graph-based and Transformer-based baselines. Additional analyses under intense-turbulence and rapid-ramping conditions suggest that the proposed framework maintains more stable empirical performance under disturbed operating regimes. The results suggest that combining wake-aware dynamic spatial priors with horizon-wise interval calibration can improve empirical calibration quality under the tested SCADA forecasting settings.
Wang et al. (Fri,) studied this question.