Short-term multi-step wind speed forecasting with multi-feature inputs using Variational Mode Decomposition, a novel artificial intelligence network, and the Polar Lights Optimizer

Wind energy is vital to the energy industry, but wind speed fluctuations pose a challenge to stable grid integration. To improve forecasting accuracy, this paper integrates artificial intelligence and decomposition method to propose a novel forecasting framework; the Polar Lights Optimizer is applied for parameter tuning. It conducts 3–6 step multi-step forecasting on wind speed time series with a sampling frequency of 1 h. Variational Mode Decomposition is employed to decompose four feature types, thereby reducing the complexity of the time series and facilitating the construction of multi-feature inputs. In the neural network component of the artificial intelligence model, a Graph Convolutional Network is used to update and aggregate multi-feature inputs by treating time steps as nodes. Sequence Squeeze-and-Excitation is applied to add channel attention, capturing deep correlations among multiple features. Cascaded LSTM is employed for time series modeling, and Linear layers are used for output forecasting. Evaluated on the public BSG dataset, the framework proposed in this paper outperforms multiple advanced models or frameworks, achieving R2 scores of 0.9857, 0.9792, 0.9706, and 0.9614 in 3/4/5/6-step forecasting. Ablation and significance tests confirm the contribution of each component. The coefficient of variation is used to assess model stability across multiple runs.