Wind energy represents one of the leading renewable energy sectors and is considered instrumental in the ongoing decarbonization process. Accurate forecasts are essential for a reliable large-scale wind power integration, allowing efficient operation and maintenance, planning of unit commitment, and scheduling by system operators. However, due to non-stationarity, randomness, and intermittency, forecasting wind power is challenging. This work investigates a multi-modal approach for wind power forecasting by considering turbine-level time series collected from SCADA systems and high-resolution Numerical Weather Prediction maps. A neural architecture based on stacked Recurrent Neural Networks is proposed to process and combine different data sources containing spatio-temporal patterns. This architecture allows combining the local information from the turbine’s internal operating conditions with future meteorological data from the surrounding area. Specifically, this work focuses on multi-horizon turbine-level hourly forecasts for an entire wind farm with a lead time of 90 h. This work explores the impact of meteorological variables on different spatial scales, from full grids to cardinal point features, on wind power forecasts. Results show that a subset of features associated with all wind directions, even when spatially distant, can produce more accurate forecasts with respect to full grids and reduce computation times. The proposed model outperforms the linear regression baseline and the XGBoost regressor achieving an average skill score of 25%. Finally, the integration of SCADA data in the training process improved the predictions allowing the multi-modal neural network to model not only the meteorological patterns but also the turbine’s internal behavior.