Toward trustworthy AI: An explainable framework for deterministic PV power forecasting using uncertainty estimation to separate model and data anomalies

Abstract

The increasing deployment of deep learning models in energy forecasting has led to improved performance, but their “black-box” nature hinders transparency, particularly in photovoltaic (PV) power prediction where trust and interpretability are crucial. This work a novel framework that enhances the explainability of Long Short-Term Memory model for PV power prediction. First, our methodology diagnoses prediction failures by analyzing the interplay between model uncertainty and Normalized Mean Absolute error (NMAE), systematically categorizing them into data anomalies (e.g., outliers, array degradation) and model anomalies (e.g., model overlap). Second, we validate these diagnoses using a suite of explanatory tools, such as Multivariate Outlier Detection (MOD) analysis and SHAP-weighted clustering, to confirm the root cause. Finally, we mitigate the identified issues using transfer learning strategy. Applied to a large-scale, real-world dataset, our framework demonstrates significant, validated improvements. For instance, after removing transient outliers from a representative array, its NMAE was reduced from 478.7% to 2.7%. The final performance for this array achieved a Mean Absolute Percentage Error (MAPE) of 6.9% and an R-squared (R 2 ) of 0.998. • Novel explainability framework for PV power prediction is proposed. • Diagnoses prediction errors into model and data-driven anomalies. • Uncertainty and error metrics diagnose root causes of failures. • Transfer learning is used to validate the framework's diagnoses.