From Explainable AI to Knowledge Extraction for Trustworthy Energy Forecasting Systems: A Systematic Review

Irina F. Iumanova, Pavel V. Matrenin, Alexandra I. Khalyasmaa

Modern artificial intelligence methods are increasingly used in power systems for renewable energy generation and electricity load forecasting. However, the limited interpretability of complex machine learning and deep learning models constrains their adoption in critical energy applications where transparency and trust are essential. Explainable Artificial Intelligence (XAI) provides tools for interpreting model behavior, yet its application to multivariate time series remains associated with significant methodological challenges. This paper presents a systematic review of XAI applications in solar power, wind power, and electricity load forecasting based on 154 peer-reviewed journal articles published between 2019 and 2026, identified through searches in Scopus, IEEE Xplore, ScienceDirect, and MDPI, following the PRISMA 2020 methodology. The review covers widely used forecasting architectures, including LSTMs, Transformers, and tree-based ensembles, as well as XAI methods. The analysis identifies a fundamental limitation of conventional XAI approaches for multivariate time series, referred to as the curse of dimensionality in XAI-based interpretation of time series, in which each time step is treated as an independent feature, resulting in explanations that are difficult to interpret in practice. To address this challenge, eight categories of XAI adaptations for time series forecasting are systematized. A classification of knowledge extraction mechanisms is proposed, including feature-level, temporal, regime-based, causal, diagnostic, model-level, and decision-support knowledge. The results demonstrate a gradual transition from explainability toward knowledge extraction, where XAI serves not only to explain individual forecasts but also to generate actionable knowledge about data, models, and energy processes. The review is limited to peer-reviewed English-language journal articles published between 2019 and 2026. The findings suggest that Knowledge Extraction represents a key mechanism for building trust in intelligent energy forecasting systems.