Photovoltaic power forecasting under atypical operating conditions validated by explainable artificial intelligence

• Robust PV forecasting framework for atypical conditions. • Optimized BiLSTM with Bayesian tuning improves accuracy across datasets. • Robustness assessed under data loss, noise contamination, and sensor drift. • RFE-based XAI supports robustness-oriented global interpretability. • Deterministic and interval forecasts remain stable under degraded inputs. This study proposes an integrated photovoltaic (PV) power forecasting framework designed for reliable operation under both stable and atypical operating conditions. While conventional forecasting approaches are typically validated under standard scenarios, real-world PV systems may experience atypical situations that significantly degrade prediction accuracy and system reliability, such as temporary sensor failures, measurement contamination with white noise, and sensor drift. The proposed framework is structured as a unified pipeline that combines an efficient data preprocessing procedure, including data cleaning, quality enhancement, nighttime filtering, and temporal resolution adjustment, with a forecasting module based on a bidirectional long short term memory (BiLSTM) network optimized through Bayesian optimization (BO). It also incorporates a dedicated validation stage under systematically simulated atypical operating conditions to quantify robustness, as well as explainable artificial intelligence (XAI) techniques to assess input variable relevance and interpret model behaviour across degradation scenarios. The framework is evaluated using open access univariate and multivariate datasets from PV installations in Belgium and Australia. Its performance is benchmarked against multilayer perceptron networks, support vector machines, and random forests, consistently achieving superior results with nRMSE improvements exceeding 60% across all datasets. The results confirm the robustness, adaptability, and interpretability of the proposed framework, even under significant data loss and measurement perturbation scenarios. Additionally, the model provides both deterministic and interval forecasts, reinforcing its suitability for real-time PV power prediction and energy management applications.