Re-evaluating neural networks for solar power estimation: features, baselines and held-out-day validation
This study re-evaluates an undergraduate neural network research project on solar-power estimation to examine how electrical inputs and validation procedures affect model performance. The dataset contains 180 observations across 20 measurement days. Two experimental tracks compare electrical and environmental inputs (Track A) with environmental measurements and observation hour (Track B). Feedforward networks implemented in scikit-learn, TensorFlow/Keras and PyTorch are evaluated against linear regression and a no-fitting electrical-product baseline for Track A using seven outer grouped folds, nested model selection where applicable and three neural network seeds. In Track A, linear regression achieves a mean MAE of 0.978 W, compared with 1.115 W for the electrical product baseline and 1.069–1.624 W across the neural-networkvariants. In Track B, linear regression achieves the lowest mean MAE of 11.835 W, compared with 11.951 W for TensorFlow, 12.032 W for PyTorch and 12.179 W for the tuned scikit learn network. These descriptive results show no consistent advantage from neural-network complexity; no significance tests were performed. The small dataset and absence of external validation limit generalisation. Both tracks estimate power fromcontemporaneous measurements; neither establishes future forecasting capability.
Authors
- Oluwafisayo Ibrahim (ORCID: https://orcid.org/0009-0006-2625-0675)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-17
- DOI
- https://doi.org/10.5281/zenodo.24320495
- Primary Topic
- Solar Radiation and Photovoltaics
- Type
- preprint