Wind Speed Prediction Method Based on EMD-LSTM with Integrated Spatial Features

In urban wind speed prediction, when forecasting wind speed at a single station, it is crucial to take into account the impact of wind speeds at surrounding monitoring stations. To address this issue, a wind speed forecasting approach utilizing Empirical Mode Decomposition-Long Short-Term Memory (EMD-LSTM) with integrated spatial features is proposed. This method employs Empirical Mode Decomposition (EMD) to decompose wind speed data in the time model. Subsequently, multiple LSTM models are applied to predict each decomposed component, and their predictions are aggregated to formulate the wind speed forecast, utilizing temporal features. In the spatial model, vector decomposition and interpolation of wind speeds at nearby stations are performed. The spatial independence weights are determined by calculating the distance between the interpolated results and actual values. Finally, in the model fusion stage, Bayesian optimization of weights is conducted based on spatial independence, resulting in the prediction of wind speed values that integrate both temporal and spatial features. Experimental findings suggest that the proposed model, considering both time and space features, achieves an average absolute percentage error (MAPE) of 20.661% for wind speed prediction and 22.038% for wind direction prediction, demonstrating relatively high accuracy. Additionally, the model exhibits a mean absolute error (MAE) of 7% and a root mean square error (RMSE) of 16%, effectively enhancing the precision of wind speed forecasting in comparison to LSTM.