Explainable AI in solar PV forecasting: A systematic review of models, explainers, and physical drivers

Engineering notes

Engineering notes will be added by the Power for Solar editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为光伏效率、钙钛矿太阳能电池、储能技术、太阳能热利用、BIPV、并网技术等高价值论文补充中文说明。

Original abstract

Solar photovoltaic (PV) forecasting supports secure grid operation, yet advanced machine learning models are hard to trust when their decisions are opaque. This PRISMA 2020 guided systematic review examines explainable artificial intelligence (XAI) in solar forecasting. Of 951 retrieved records, 173 met strict eligibility, requiring quantitative forecasting evaluation plus an explicit explanation output. A sensitivity analysis of the decisive screening rule, which excluded 299 records using attention only as an architectural component, supported its precision, and dual independent re-screening of 75 excluded records found no false exclusions. Long short-term memory (LSTM) models dominated (48 studies), ahead of transformers (39) and convolutional neural networks (CNN, 32) used mainly for image-based nowcasting. SHAP prevailed among explainers (69 studies), ahead of feature importance and permutation methods (39), attention visualization (31), and LIME (14), and 47 studies combined several methods. Quality-stratified synthesis showed these rankings are robust, with the ordering unchanged among the 114 high-quality studies. Irradiance ranked first across targets, while temporal encodings and ambient temperature were stable secondary contributors. Humidity, wind, and cloud proxies shifted with season and climate zone, making global rankings conditional. Horizon mattered, with longer horizons using tabular weather and post hoc attribution and ultra-short horizons using sky images. Cross-study benchmarking remains fragile, and quantitative evaluation of explanation fidelity, stability, and runtime is rare. A critical appraisal of the dominant explainers documents computational cost, instability, sensitivity to correlated inputs, and unverified faithfulness, alongside a reproducibility gap, with only 1.7% of studies releasing both code and data.

Links and sources

• PDF from original source

Comments