Choosing the right model for PV power estimation

Published 13 October 2026Updated 13 October 2026SolarOZ technical summary

Based on research by Caixia Li and collaborators, published in Energy 323 (2025) 135744.

Short answerPhysical models are useful when operating history is limited and interpretability matters; machine-learning models can improve accuracy with representative data; larger, more variable plants may benefit from hybrid methods and stronger irradiance inputs.

Why this matters for plant teams

Converting solar-resource information into expected PV power supports monitoring, diagnostics, energy management and grid integration. The appropriate model depends on the plant scale, data quality, operating variability and decision—not only the lowest headline error.

Physical models: useful when historical data is limited

The study's physical model used irradiance, temperature and module electrical parameters to estimate PV output. For the 12 MW plant it achieved an nRMSE of 5.29% without historical power data. This supports physical modelling where a transparent estimate is needed or operating history is limited.

Machine learning: higher accuracy with suitable history

The study compared KNN, Random Forest and Support Vector Regression. In the mid-sized plant, SVR produced the best reported overall accuracy, with an nRMSE of 3.97%. The result depends on having sufficient and representative historical data.

Why large and variable plants are harder

At the 100 MW plant, both the physical and SVR models lost accuracy, particularly in the mid-power range under variable irradiance. Rapid cloud movement, nonlinear irradiance behaviour, seasonal variation and soiling increased the challenge. Using measured irradiance directly in the physical model improved the reported accuracy by more than 40%, making irradiance input quality a central finding.

A physical-AI hybrid approach

The researchers combined a Temporal-aware Multi-level Mixer with a physical IV model. The hybrid retained a physical model structure while using AI to represent nonlinear irradiance patterns. In the large-plant scenario it reduced RMSE by about 9% compared with the reported SVR result and achieved an nRMSE of 5.64%.

Practical interpretation

Use physical models when historical data is sparse and transparency matters.
Use machine learning when a clean and representative operating history is available.
For large or highly variable sites, investigate irradiance input quality before attributing error only to the model class.
Match the model to the operational question: benchmarking, diagnosis, monitoring or forecasting.

Assumptions and limitations

This page summarises one published study; its reported errors are specific to the studied datasets, plants and model configurations. They should not be treated as guaranteed performance for another asset. Site-specific validation remains necessary.

Original source

Li, C. et al. “Assessing solar-to-PV power conversion models: Physical, ML, and hybrid approaches across diverse scales.” Energy 323 (2025) 135744. Open the paper via DOI.