Optimal wind turbine positioning relies on understanding local wind patterns via wind direction sensors. However, such hardware adds complexity and costs. This research demonstrates a virtual sensing approach using Long Short-Term Memory (LSTM) neural networks to predict wind direction solely from historical wind speed data. The LSTM networks were trained on two years of 10-minute resolution data from three stations in Iran’s Markazi Province. The models accurately inferred wind directionality and speeds based on temporal analysis of speeds. Predictions closely matched measured wind direction per wind rose validation. The research indicates software-based artificial intelligence algorithms can effectively replace physical wind direction sensors, enabling simpler and cheaper wind farms. Operational reliability can be ensured via continual model updating. The approach also has promising implications for broader wind forecasting applications. Overall, the feasibility of transitioning from physical hardware to virtual software-based wind sensors is demonstrated.
