Local and Long-range Convolutional LSTM Network: A novel multi-step wind speed prediction approach for modeling local and long-range spatial correlations based on ConvLSTM

Accurate wind speed prediction is crucial for enhancing the stability and economic efficiency of power system operation, particularly in wind power grid integration. However, existing methods face challenges as they fail to explicitly model local and long-range spatial correlations simultaneously, thereby limiting the performance of wind speed prediction to a certain extent. To overcome these challenges, this study develops a novel method, namely, LLConvLSTM, from the perspective of modeling local and long-range spatial correlations in wind speed, which leverages Deformable Convolution V2 and Coordinate Attention for multi-step spatiotemporal wind speed prediction. A ConvLSTM encoder–decoder architecture is designed for end-to-end spatiotemporal wind speed prediction. The Residual Deformable Convolution Module (RDCM) increases additional offsets and modulation scales in the spatial sampling locations, enhancing the capability to capture local spatial correlations. Dense Coordinate Attention Module (DCAM) embeds spatial positional information into the channel attention. DCAM improves the representability of long-range spatial correlations. Experimental results based on wind speed data from 253 virtual wind turbines demonstrate that the proposed approach significantly outperforms existing methods throughout the entire year and months. Moreover, the proposed method achieves Mean Squared Error (MSE) of 0.1199, 0.3446 and 0.5798 for multi-step wind speed prediction, representing reductions of 22.47% to 40.91% compared with existing methods. These findings highlight the significance of modeling local and long-range spatial correlations in enhancing the accuracy and stability of wind speed prediction. Future research will design a universal method capable of handling turbine data from any location and emphasize long-term forecasting in wind speed prediction. [Display omitted] •This study integrates local and long-range spatial correlations in wind speed prediction.•Residual Deformable Convolution Module is utilized to capture local spatial correlations of wind speed flows outstandingly.•Dense Coordinate Attention Module is applied to construct long-range spatial correlations of wind speed flows adequately.•The MSE of proposed approach in multi-step prediction is generally reduced by more than 20% compared to other methods.