Local Input–Output Traceability for Multimodal Solar Power Predictions by Integrating Transitional Neural-Backed Decision Tree

Decreasing the randomness of renewable energy sources is the priority for the stability of novel power systems. Renewable energy prediction models have been studied extensively with higher precision. However, these models have become much more complicated and opaquer, meanwhile accuracy improvements almost reach the convergence. Major prediction deviations are still inevitable and how the deviations occurred is inexplicable in those black-box models. This prediction interpretability problem arises puzzling power system operators. Specifically, advanced solar power forecasting technologies have proposed multimodal prediction models that involve various input forms, such as remote-sensing cloud images, exacerbating the forecast opacity. Hence, this study focuses on the interpretability issue of deep-learning-based multimodal solar power predictions, and proposes a post-hoc local traceability method. Based on neural-backed decision trees, the method can decouple solar power forecast outputs into an inference hierarchy and weather transition probabilities. Effects of multimodal inputs can be also quantified with Shapley values in the method. By providing qualitative results of input effects and the prediction inference process, the proposed method increases local interpretability while maintaining forecast accuracy.