Physics-Informed State Space Models for Reliable Solar Irradiance Forecasting in Off-Grid Systems
Physics-Informed State Space Models for Reliable Solar Irradiance Forecasting in Off-Grid Systems
Authors: Mohammed Ezzaldin Babiker Abdullah Date: 2026-04-13 Paper ID: openalex:2604.11807
Summary
This paper introduces the Thermodynamic Liquid Manifold Network, a physics-informed deep learning architecture designed for robust solar irradiance forecasting in off-grid photovoltaic systems. By projecting meteorological data onto a Koopman-linearized Riemannian manifold and enforcing celestial geometry constraints, the model eliminates common issues like phase lags and nocturnal power artifacts. Validated over a five-year period in a semi-arid climate, the framework demonstrates high predictive accuracy and extreme efficiency, making it suitable for edge-deployed microgrid control.
Key Contributions
- Introduces the Thermodynamic Liquid Manifold Network, achieving an RMSE of 18.31 Wh/m2 and a Pearson correlation of 0.988 for solar irradiance forecasting.
- Integrates a Spectral Calibration unit and a multiplicative Thermodynamic Alpha-Gate to ensure strict compliance with celestial geometry.
- Demonstrates complete elimination of nocturnal generation anomalies and sub-30-minute phase response during rapid cloud transients over a five-year testing horizon.
- Provides an ultra-lightweight, edge-deployable model with 63,458 parameters suitable for autonomous off-grid microgrid controllers.
Key Concepts
- thermodynamic-liquid-manifold-network: A neural network architecture that projects meteorological data into a Koopman-linearized Riemannian manifold to enforce physical constraints on solar irradiance.
Archivist Review
The review process prioritized the core methodological innovation—the Thermodynamic Liquid Manifold Network—as it provides a robust, reusable framework for embedding physical constraints into state-space models via Koopman linearization. Sub-components like the calibration unit and alpha-gate were rejected as they are domain-specific implementation details that do not yet warrant independent conceptual status. No open questions were approved, as the current evaluation presents the approach as a mature solution for the described task.
Approved Concepts
- Thermodynamic Liquid Manifold Network: It is the core architectural contribution, combining Koopman operator theory with geometric constraints for time-series forecasting.
Rejected Candidates
- [concept] Spectral Calibration unit (
spectral-calibration-unit) - subcomponent_of_broader_mechanism: This is a sub-module of the main architecture and lacks independent conceptual standing outside of this specific implementation. - [concept] Thermodynamic Alpha-Gate (
thermodynamic-alpha-gate) - subcomponent_of_broader_mechanism: This is a specific gate mechanism internal to the proposed architecture and not a widely applicable primitive.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2604.11807
- paper_id
- 2604.11807
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
-
- concept_slugs
- thermodynamic-liquid-manifold-network
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-04-16T05:46:39Z