Solar radiation intra-hour forecasting at Canary islands: Improving ramp events detection

The integration of high shares of photovoltaic (PV) energy in isolated insular power grids poses significant challenges due to the volatility of generation and the frequent occurrence of critical ramp events. This paper introduces a deep learning framework incorporating a cross-attention mechanism for intra-hour solar forecasting in such environments, trained under a composite objective function that combines Soft Dynamic Time Warping with a standard MAE term to better capture abrupt irradiance transitions. The methodology is validated using real operational data from a PV plant located in Tenerife power system, benchmarking the proposed architecture against Smart Persistence and a convolutional neural network (ECNN), and contrasting the composite loss against a standard RMSE objective under an identical architecture. The proposed model achieves a Forecast Skill of up to 14.94% relative to Smart Persistence, outperforming ECNN (12.71%). More importantly, despite comparable pointwise error metrics, the composite-loss model substantially outperforms its RMSE-trained counterpart in ramp event detection, particularly at longer lead times, where the RMSE-trained model fails to detect any ramp event. These results confirm that the double penalty problem affecting standard Euclidean losses is primarily driven by the choice of training objective rather than network capacity, highlighting the value of shape and timing aware losses for improving grid stability and energy storage management in insular energy systems.