The threat posed by solar flares to human technology underscores the need for reliable forecasting models to mitigate risks effectively. Our study introduces a novel deep learning forecasting approach while emphasizing the need for performance evaluation methods tailored to highlight current models’ limitations better. We discuss shortcomings in existing evaluation metrics such as the True Skill Statistic (TSS) and the Heidke Skill Score (HSS) and propose the Matthews Correlation Coefficient (MCC) as a consistent alternative to both of them. We introduce Persistent-Relative-Skill Scores (PRSS), as well as metrics evaluation on the restriction of time windows presenting a change of activity (Activity-Changes (AC) performances). We found that models reaching state-of-the-art performances with traditional metrics can be less efficient and explanatory than a simple persistent model and struggle to forecast change in activity significantly better than random guesses. We introduce the Patch-Distributed-CNNs (P-CNN), which allows for performing full-disk forecasts while providing event probabilities in solar sub-regions and position predictions. This new framework offers similar information to Active-Region-based forecasting models while bypassing the problem of unrecorded and misattributed flares in Active-Region flare catalogues that are detrimental to machine learning training. As a result, the model also operates independently of prior feature extraction and AR detection, thus offering promising operational utility with minimal external dependencies. We propose a methodology for constructing balanced and independent Cross-Validation folds to train and evaluate full-disk forecast models accurately. Models combining SDO/AIA EUV images as inputs show improved performances compared to employing SDO/HMI photospheric magnetograms, with a TSS of 0.74 for the C+ model and 0.62 for the M+ model.