Peak Solar Intelligence

The Methodology

Data Clean-Up & Pipeline

Raw satellite data is noisy. We apply a rigorous Physics-Guided cleaning pipeline before the data ever touches the neural network.

Crucially, Night Masking forces all solar values to true zero when the sun is below the horizon, eliminating "night noise" and easing the model's burden.

This includes a strict Chronological Split (2010-2015 Train / 2020-2024 Test) to act as a temporal stress test against climate shifts.

The Approach

Physics-Guided vs. Informed

We faced a critical choice: Physics-Informed (PINN) vs. Physics-Guided (PGNN).

The Engine

Physics-Guided Core

Instead of relying on AI to implicitly learn the laws of physics, we inject explicit solar geometry directly into the learning process.

A local clear-sky calculation engine computes the theoretical irradiance baseline, anchoring the model with 15 engineered features.

By decoupling the deterministic geometry from stochastic clouds, the AI only needs to learn the residuals — solving the Phase Lag problem entirely.

Training Protocol ℹ️

A rigorous two-stage training strategy was adopted. During Bayesian search, candidates trained for 50 epochs. The final model retrained for 100 epochs.

Remarkably, the physics-guided architecture exhibited inherent numerical stability — no gradient clipping was needed, a testament to the well-conditioned input space provided by engineered features.

The Optimizer

Bayesian Intelligence

Manual hyperparameter tuning is computationally irresponsible. We employed Bayesian Optimization with the Expected Improvement Plus (EI+) acquisition function.

The algorithm iteratively updates a probabilistic surrogate model, balancing exploration (uncertain regions) and exploitation (promising regions) across 30 iterations.

The Results

The Complexity Paradox

Our research proved that in high-noise meteorological tasks, Explicit Physical Constraints outperform Self-Attention mechanisms.

The "Simpler" Physics-Hybrid model achieved a 36% Error Reduction compared to heavy Transformer-based baselines.

Live Demonstration

Interactive Phase Lag Simulator ℹ️

See the difference for yourself. Drag the Cloud Intensity slider to simulate a sudden dust storm or cloud event.

5-Year Robustness ℹ️

To ensure operational validity in a changing climate, a long-term Stress Test was conducted on an independent dataset spanning 2020–2024.

The R² remained consistently above 0.995 throughout the entire five-year horizon — proving the model learned true physical laws, not temporary weather patterns.