Wind power forecasting supports power-system planning by reducing uncertainty in operational decision making. Because meteorological conditions such as wind speed are inherently stochastic, accurate day-ahead and longer-horizon wind-power predictions remain difficult. A natural way to address this difficulty is to combine weather information from multiple points on a geographical grid with temporal information from historical wind-farm power outputs. Our proposed CNN-RNN architecture combines convolutional neural networks (CNNs) and recurrent neural networks (RNNs) to extract spatial and temporal information from multi-dimensional input data for day-ahead predictions. The method adopts an ultra-wide learning view by combining data from multiple numerical weather prediction models, wind farms, geographical locations, and atmospheric levels. We also evaluate global forecasting, in which one model is trained across multiple wind farms, and a hybrid approach that passes convolutional features to a tree ensemble, represented by Light Gradient Boosting Machine (LGBM), instead of to fully connected layers. The results show that our proposed CNN-RNN architecture outperforms various baseline models when trained globally, but does not consistently do so when trained independently on each farm. We also observe that passing CNN-extracted spatial features to LGBM improves LGBM performance, providing further evidence of CNN’s spatial feature extraction capabilities.