AI-driven digital twin framework for multi-horizon wind power forecasting using Bi-LSTM attention networks

Wind energy is one of the fastest-growing renewable sources of electricity in the world; however, its variability and intermittent nature present issues for effective integration into a reliable grid. Accordingly, precise wind power forecasting is essential for economic dispatch, reserve scheduling, and power market participation. We present a new AI Digital Twin (AIDT) framework for multi-horizon wind power forecasting that integrates a physics-based wind farm model with an end-to-end deep-learning engine based on a bidirectional long short-term memory network enhanced by a multi-head attention mechanism, as a data assimilation module grounded on a Kalman filter, and as a statistical uncertainty quantification method. The framework is designed around a continuously updating cyber-physical representation of a wind farm, in which SCADA telemetry, numerical weather prediction outputs, and LiDAR-derived wind profiles are assimilated into the digital-twin state; in the present study, this synchronization loop is evaluated retrospectively against archived data rather than in a live online deployment. The proposed model is validated on two complementary, publicly available wind farm datasets: the NREL Western Wind dataset [Draxl et al., Appl. Energy 151, 355 (2015)], a WRF-based simulated/reconstructed wind-resource record, and the EirGrid Ireland dataset [EirGrid Group, Wind Generation Data—All-Island Grid, EirGrid System Operator, Dublin, Ireland, 2020], comprising measured operational generation records, together spanning a combined multi-year record of more than 350 000 data points across multiple meteorological regimes. Results show that AIDT outperforms six baseline models, achieving RMSE, MAE, and MAPE of 1.38 MW, 1.07 MW, and 5.1%, respectively, on the NREL test set, corresponding to improvements of 35.8%, 37.8%, and 35.4% over the best-performing baseline (a transformer). The 90% nominal prediction interval achieved a coverage probability of 91.4% with a normalized average width of 1.87 MW, indicating close-to-nominal calibration with comparatively narrow intervals. Horizon-wise analysis shows that the Kalman-filter-based state assimilation yields consistently lower RMSE than an open-loop variant of the same network from 1- to 48-h horizons, with the largest relative benefit at longer horizons, where sequential correction against SCADA measurements offsets the growing uncertainty of the exogenous NWP inputs. These findings indicate that combining physics-informed modeling with deep learning in a digital-twin-style architecture improves forecasting accuracy over the individual architectural and physics-only components, particularly for high-variability wind-ramp and storm regimes; confirming this framework under live, online operation with hardware-in-the-loop validation remains an important direction for future work.