2026· IEEE Open Journal of the Industrial Electronics Society· Vol 7, pp. 1037-1050· 0 citations· 47 references
TL;DR
Comprehensive evaluation including ensemble comparisons, ablation studies, and uncertainty metrics, namely prediction interval coverage probability and mean prediction interval width, confirms high coverage and sharpness, while feature-representation fusion ensures consistent, low-uncertainty predictions.
Abstract
This article presents a short-term (1-h ahead) physics-guided multirepresentation feature fusion framework for photovoltaic (PV) power prediction with uncertainty quantification based on quantile regression for interval forecasting. The proposed hybrid methodology integrates feature-representation fusion, deep learning architectures, including long short-term memory (LSTM) and gated recurrent unit (GRU), and machine learning models, such as support vector regression (SVR) and random forest (RF), along with dimensionality reduction techniques, namely principal component analysis (PCA) and autoencoder (AE), and a fully defined physics-based digital twin (DT) model with explicit irradiance-to-power equations and parameter calibration. The DT residual, computed causally using only training data, is incorporated as a physics-informed feature to capture unmodeled nonlinearities. PIs are learned using the pinball (quantile) loss function, replacing heuristic assumptions, and feature-representation fusion reduces uncertainty while enhancing robustness. Validation on real meteorological datasets from Izki and Manah stations (2017–2023) shows that conventional feature sets achieve root mean squared error (RMSE) values between 0.14–0.17 with coefficient of determination (<inline-formula><tex-math notation="LaTeX">$R^{2}$</tex-math></inline-formula>) ranging from 0.47–0.61. Incorporating DT residuals significantly improves performance, while maintaining realistic generalization performance, with LSTM achieving RMSE <inline-formula><tex-math notation="LaTeX">$\approx$</tex-math></inline-formula> 0.0105 and <inline-formula><tex-math notation="LaTeX">$R^{2}$</tex-math></inline-formula> up to 0.9979. SVR and RF models also benefit, achieving RMSE <inline-formula><tex-math notation="LaTeX">$\leq 0.0525$</tex-math></inline-formula> and <inline-formula><tex-math notation="LaTeX">$R^{2} \geq 0.9467$</tex-math></inline-formula>. Comprehensive evaluation including ensemble comparisons, ablation studies, and uncertainty metrics, namely prediction interval coverage probability and mean prediction interval width, confirms high coverage and sharpness, while feature-representation fusion ensures consistent, low-uncertainty predictions. Further experiments on a physically inspired synthetic dataset with explicitly defined generation process and temporal resolution of 1000 samples demonstrate similar trends. Using DT residual features, LSTM and GRU achieve <inline-formula><tex-math notation="LaTeX">$R^{2} > 0.998$</tex-math></inline-formula> with RMSE <inline-formula><tex-math notation="LaTeX">$\approx 0.042$</tex-math></inline-formula>, while RF reaches <inline-formula><tex-math notation="LaTeX">$R^{2} = 0.897$</tex-math></inline-formula> and a relatively higher RMSE <inline-formula><tex-math notation="LaTeX">$\approx 0.326$</tex-math></inline-formula>. PCA and AE representations improve computational efficiency but provide comparatively lower predictive accuracy. Direct comparison between individual models and the proposed ensemble highlights the robustness and stability of the fusion strategy, particularly in high-noise conditions. In summary, the framework offers a reproducible, physically grounded, and uncertainty-aware solution for next-generation solar energy forecasting, explicitly quantifying prediction uncertainty and leveraging physics-informed residuals to enhance predictive reliability and interpretability across both real and synthetic scenarios.
Reliable short-term photovoltaic (PV) power forecasts are pivotal to grid balancing, intraday market clearing, and asset optimization. This study provides an in-depth comparative analysis of four state-of-the-art neural architectures–a convolutional-recurrent hybrid (CNN-LSTM), an LSTM-Autoencoder, a standalone LSTM, and a time-series Transformer–trained on three years of irradiance and power data with one-minute sampling interval from two meteorologically contrasting PV sites (semi-arid Colorado and desert Nevada). Each model is evaluated in a probabilistic forecasting setting, where kernel density estimation (KDE)-based residual post-processing and quantile extraction are used to obtain calibrated prediction intervals. The benchmark adopts an irradiance-driven forecasting framework, using historical irradiance observations as inputs and future PV power as the prediction target, thereby enabling a controlled assessment of each architecture’s ability to model the underlying irradiance-to-power relationship. Robustness is probed through three stressors: truncated training histories (1–3 years), temporal coarsening (1-, 5-, and 15-minute records), and up to 30% randomly or block-removed observations. Across all tests, CNN-LSTM consistently delivers the narrowest and most reliable intervals, leveraging its convolutional front-end to detect rapid cloud-edge ramps while the LSTM tail preserves long-range context. Compared with a strong LSTM benchmark, the hybrid reduces mean interval width by 22%, lowers the Continuous Ranked Probability Score (CRPS; a scoring rule measuring the distance between the forecast distribution and the observed outcome - lower is better) by 11%, and boosts empirical coverage by 5% (achieving 98.2% reliability). These advantages remain intact under data scarcity, coarse sampling, and substantial missingness, highlighting the critical role of architectures that unite local feature extraction with sequential memory. The findings confirm that high-fidelity point forecasts and rigorously quantified uncertainty can be achieved simultaneously, providing a clear path toward more dependable PV dispatch, reserve allocation, and market participation.
Saloni Dhingra, G. Gruosso, G. Storti Gajani· Neural computing & applicati...· 0 citations
Photovoltaic (PV) power forecasting is essential for the reliable integration of solar energy into modern power systems. Although recent machine learning and ensemble-learning models achieve high forecasting accuracy, their performance often deteriorates under variable weather conditions, leading to large prediction errors and reduced reliability. This limitation highlights the need for forecasting frameworks that explicitly address error heterogeneity rather than focusing solely on global performance metrics.This study proposes a hybrid PV forecasting framework that combines stacking ensemble learning with a targeted residual correction strategy. The stacking architecture integrates three complementary ensemble models, namely Random Forest, Extra Trees, and XGBoost, whose outputs are combined through a Ridge regression meta-learner. To improve forecasting robustness, a residual-learning mechanism is introduced to identify difficult observations based on prediction errors and irradiance conditions. A dedicated correction model is subsequently applied to these hard-to-predict samples.The proposed framework is evaluated using real-world PV generation and meteorological data collected from a desert-climate photovoltaic installation. Experimental results show that ensemble-based models achieve excellent overall forecasting accuracy, with Extra Trees providing the best global performance. Furthermore, weather-regime and sensitivity analyses reveal that forecasting errors are strongly influenced by irradiance variability and atmospheric conditions. The findings demonstrate that analyzing error distribution and forecasting robustness provides valuable insights beyond conventional aggregate metrics, contributing to the development of more reliable photovoltaic forecasting systems.
Khawla Oufrit, A. Mouadili, M. Zazoui· EPJ Web of Conferences· 0 citations
In response to stochastic fluctuations in large-scale wind power integration and the resulting peak-shaving challenges, high-precision forecasting with physical consistency is essential for grid safety. To address the inefficiency of physical models and poor interpretability of data-driven methods, this paper proposes a hybrid FCM-WGM-BiLSTM-Transformer (FW-BTP) framework integrating Fuzzy C-Means (FCM) clustering, Weighted Grey Model (WGM) trend extraction, and a coupled BiLSTM-Transformer module. Physical prior knowledge—wind turbine power curves—is embedded into the loss function via a Physics-Guided Loss Regularization (PGL) mechanism. Validated on measured data from a Xinjiang wind farm, the model achieves an R2 of 0.9967, MAE of 6.11, and RMSE of 11.19. The proposed model reduces R2 by 37% compared to the newer model KAN, and compared to the better-performing recurrent baseline model (BiLSTM, MAE = 8.75 MW), the proposed FW-BTP model reduces the MAE by 30% (to 6.12 MW). Ablation studies confirm the WGM reduces LogCosh loss from 9.57 to 5.12, and SHAP analysis verifies sensitivity to trend and physical wind speed features. The framework balances accuracy, robustness, and interpretability, supporting refined scheduling in modern power systems.
Engineering use of AI forecasting models requires not only high nominal accuracy but also predictable behavior under uncertain inputs. In photovoltaic (PV) forecasting, this requirement is especially challenging because numerical weather prediction (NWP) errors are temporally correlated, state dependent, and physically coupled across variables. Existing evaluations, however, often rely on perfect forecast assumptions or simplistic perturbations that do not reflect these characteristics. This study presents a physically constrained robustness evaluation framework based on simulation, using virtual PV power as a controlled response variable to isolate the propagation of input uncertainty from confounders at the plant level. Six representative machine learning and deep sequence models, including PatchTST, GRU, N-HITS, and LightGBM, are evaluated under dynamic NWP perturbations with heteroscedasticity modulated by clear-sky conditions and Erbs reconstruction that preserves radiation consistency. The results show that sequence models provide stronger noise filtering and temporal resilience than a strong tabular baseline under medium to high disturbance regimes. SHapley Additive exPlanations (SHAP) and Integrated Gradients (IG) further support a feature reallocation tendency at the case level, in which predictive reliance shifts from corrupted future forecasts toward more stable historical observations and deterministic physical priors. A Pareto analysis of accuracy under clean conditions, robustness, and computational latency then translates these findings into engineering implications for robustness assessment and model selection under forecast uncertainty.
As the integration of photovoltaic (PV) systems into modern electrical grids continues to grow, there is an increasing demand for accurate short-term forecasting of both global horizontal irradiance (GHI) and PV power output. However, most existing deep learning models weakly couple GHI and power predictions, degrade under cloudy and partially cloudy conditions, and offer limited interpretability. To address these challenges, this study proposes the spatio-temporal fusion network (STFNet). This explainable hybrid model integrates convolutional neural networks for local feature extraction, unidirectional Long Short-Term Memory networks for temporal modelling, and XGBoost for nonlinear prediction, enhanced by a novel cloud-conditioned bidirectional GHI injection mechanism. The core innovation is a bidirectional gated fusion that couples a dedicated GHI branch and a PV branch in both directions: a GHI→PV blend gate transfers irradiance physics into the power pathway. In contrast, a PV→GHI residual gate allows recent power dynamics to refine the irradiance estimate, with both gates modulated by cloud-type conditioning so that the coupling adapts to prevailing sky conditions. Extensive evaluations on the SHEERM dataset across 15, 30, and 60 min horizons, using mean absolute error, root mean square error (RMSE), and R2 metrics stratified by season and cloud type, demonstrate that STFNet achieves a better overall performance among the eleven models compared, with an R2 of 0.9413 for GHI (RMSE 55.06 W m−2) while maintaining competitive PV performance (R2 0.9069, RMSE 58.49 W m−2). A controlled ablation isolates the bidirectional coupling as the sole driver of the GHI improvement, which is large (∼17% RMSE reduction over the unidirectional gate) and highly reproducible across random seeds; the improvement is consistent across all seasons, horizons, and cloud buckets. LIME and SHAP analyses further reveal that lagged PV power, clear-sky GHI, precipitable water content, and sinusoidal time encodings are the most influential predictors, enhancing both accuracy and physical interpretability for reliable solar forecasting and grid integration.
C. Otuka, Dongsheng Cai, C. Ukwuoma et al.· Engineering Research Express· 0 citations
Sound prediction of existence of future generation based on stochastic sources is a critical issue because of high nonlinearities, high transitions in the environment, and the nature of inherent uncertainty in observational information. The paper proposes an integrated architecture of deep learning, which is the Hierarchical Regime-Adaptive Probabilistic Network (HRAPN) that aims to overcome these constraints using an end-to-end learning framework. The solution selection boasts of hierarchical representation induction with a latent regime adaptation mechanism that modulates dynamically the internal model behavior in a non-stationary environment. Besides that, attention guided dependency synthesis module performs informative temporal context aggregation selectively to allow an efficient long horizon modeling without impaired performance. In contrast to the more traditional deterministic approaches, HRAPN uses probabilistic model of output in order to explicitly model predictive uncertainty, which enhances robustness and reliability of the estimated decisions. The system takes directly heterogeneous and multivariate data, without explicit features or domain pre-treatment. It is experimentally assessed that the presented method achieves higher results as compared to existing baselines in accuracy, stability, and uncertainty calibration in various forecast periods. The findings validate the performance of regime perceiving, hierarchical abstraction and probabilistic inference in the same learning process. The proposed HRAPN model offers a scaffoldable and adaptable evaluation of the dynamic generation modeling under both variable operating conditions with data. The suggested framework attains an overall accuracy of 96.6%, illustrating its robust prediction reliability and exceptional performance relative to current methodologies.
Bal Krishna Saraswat, Sonu Lal, Anshu Malhotra et al.· 2026 International Conferenc...· 0 citations