Dual-task coupled forecasting of wind speed and wind power generation based on multimodal spatiotemporal feature fusion

Accurate prediction of wind speed and wind power is crucial for optimizing wind farm operations and improving renewable energy utilization. This paper proposes a dual-task coupled prediction model integrating multi-source data, including wind speed, wind direction, temperature, humidity, pressure, and wind power. The model incorporates dual decomposition, multi-dimensional feature extraction, adaptive multi-time-window fusion, and error correction to enhance prediction accuracy and stability. First, Singular Spectrum Analysis (SSA) and Improved Complementary Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) are applied for dual denoising, while 3D data tensors are constructed from multi-height wind speed and meteorological factors to capture spatiotemporal patterns. A multi-scale Convolutional Neural Network (CNN) feature extraction module is then used to explore complex interactions. In the prediction phase, a dual-task structure is designed with an improved loss function, while the Improved Snake Optimization (ISO) algorithm adaptively optimizes the weighted fusion of multi-time-window predictions. Finally, a Variational Mode Decomposition-Gate Recurrent Unit (VMD-GRU) error correction mechanism refines residuals to further improve accuracy. Experiments on two datasets show that the proposed model significantly outperforms benchmarks across multiple indexes (MSE, MAE, MAPE, RMSE, R²), demonstrating its effectiveness. Taking Dataset 1 as an example, the wind speed forecasting results of the proposed method demonstrated improvements over the baseline model, with the MSE, MAE, MAPE, RMSE, and R² enhanced by 93.7 %, 85.3 %, 79.5 %, 81.5 %, and 35.9 %, respectively. Ablation studies confirm that combining multi-feature extraction, dual-task coupling, adaptive fusion, and error correction significantly enhances wind speed and wind power prediction, compared to the baseline model, improvements of 95.3 %, 81.4 %, 78.8 %, 79.1 %, and 24.4 % were achieved across the five aforementioned indexes, respectively, providing valuable insights for wind power system optimization. • Multilevel wind data is stacked into 3D tensors with meteorological features for spatiotemporal modeling. • A dual-task model fusing 1D, 2D, and 3D CNNs improves wind speed and power prediction accuracy. • ISO-based weight fusion and hybrid loss improve the prediction accuracy and model stability.