Benchmarking physics-informed machine learning-based short term PV-power forecasting tools

Uncertainty is one of the core challenges posed by renewable energy integration in power systems, especially for solar photovoltaic (PV), given its dependence on meteorological phenomena. This has motivated the development of numerous forecasting tools, recently focused on physics informed machine learning (ML). Virtually, every paper claims to provide better accuracy than the previous, yet the replicability of these studies is very low, motivating unfair, or erroneous comparisons. This paper reviews and compares the most relevant ML-methods identified in the literature (Random Forest, Support Vector Regression, Convolutional Neural Networks, Long–Short Term Memory and a Hybrid of the last two) with two statistical methods: persistence and an Semi-Parametric Auto-Regressive model. Furthermore, we propose a methodology to integrate a PV-performance model in ML models to forecast power several hours ahead with 5-min resolution. A basic dataset including power production and meteorological measurements is expanded with physics-informed features that capture the relationship between weather and PV operational state, while keeping strong correlation towards the intrinsic feature. This allows the models to learn about the physical interdependence of different features, potentially yielding a higher accuracy than conventional methods. Then, we also propose a physics-informed feature selection to tighten the search-space of the best performer. A case study of a PV array in Denmark is used for validation using both the original and expanded datasets. Results show how the best ML models consistently used physics-informed features in all cases.