Temporal Attention-Driven Ensemble Learning for Solar Power Generation Forecasting: The SOTAE Framework and Multi-Metric Comparative Evaluation
Abstract
Solar power generation forecasting is an essential element of effective integration of photovoltaic (PV) systems into a smart grid infrastructure, which makes it possible to plan dispatch effectively, schedule storage and manage reliability. Traditional methods of machine learning are stunted by domain blind architecture that does not reflect the complicated time dependence, non-linear interaction of features and multi-scale variation in solar power time series. The given paper presents a five-layer hybrid machine learning architecture called Solar-Optimized Temporal Attention Ensemble (SOTAE), which has been specifically designed to forecast solar power generation. In SOTAE, Dual-Signal Feature Enhancement (DSFE), Non-linear Feature Interaction Synthesis (NFIS), Heterogeneous Seven-Learner Out-of-Fold Stacking, Correlation-Gated Adaptive Residual Boosting (CG-ARB), and Polynomial-Augmented Meta-Learning have all been combined into a single architecture. SOTAE is also found to perform well on a large-scale solar power dataset (185,464 observations) with a high set of four evaluation metrics: $\mathrm{R} 2=0.8359, \text{RMSE}=210,652 \text{Wh}, \text{MAE}=23,665$ Wh and MAPE $=507.85 {\%}$. SOTAE has a higher performance compared to all the eight base algorithms such as Linear Regression, Decision Tree, K-Nearest Neighbour, AdaBoost, Support Vector Regression, Neural Network, Random Forest, and Gradient Boosting. The superiority of SOTAE in multi dimensions is proven through extensive visualisation using ten graphs (distribution analysis, correlation heatmaps, attention weight profiles, scatter comparisons, residual diagnostics, learning curves, and multi-metric radar charts) and conclusively proves the superiority of the approach. When a Wilcoxon signed-rank test is used, the statistical significance $(\mathrm{p}<0.001)$ is statistically significant. These findings define the existence of SOTAE as a powerful and operationally feasible model of solar energy prediction in renewable energy management models.