Short-Horizon Data-Driven Joint Forecasting of Wind Speed and Direction for Future Aware Wind Farm Control

Abstract

Preview-based wind-farm control depends on accurate short-horizon estimates of wind speed and direction. Yet at five-minute resolution, practical models must be low-latency, and must outperform persistence modelling. This paper presents a two-step, control-oriented data-driven forecasting model for one-step-ahead prediction of the wind vector components. In step 1, the study compares a Long short-term memory network (LSTM), a 1-D convolutional neural network (CNN), and a causal temporal convolutional network (TCN) against a one-step persistence baseline, under a unified training and evaluation protocol. In step 2, the study retains the best-performing architecture and enhances the input representation using variational mode decomposition (VMD) and mutual-information (MI) lag analysis, yielding a VMD–MI–TCN model. The study uses three consecutive years of five-minute wind speed and direction from NREL’s Wind Resource. Forecasts are evaluated on a held-out test set using wind-speed errors, circular wind-direction errors, and error-based skill scores relative to persistence. Overall, the proposed workflow provides a basis for constructing and benchmarking lightweight wind speed and wind direction forecasting models suitable for real-time control integration.