Title:	Predicting PV Self-Consumption in Villas with Machine Learning
Author(s):	F. Galli, N. Sommerfeldt
Keywords:	Economic Analysis, Grid-Connected, PV System, Modelling / Modeling, Prosumer
Topic:	PV Systems – Modelling, Design, Operation and Performance
Subtopic:	Design and Installation of PV Systems
Event:	38th European Photovoltaic Solar Energy Conference and Exhibition
Session:	5CO.12.1
Pages:	993 - 997
ISBN:	3-936338-78-7
Paper DOI:	10.4229/EUPVSEC20212021-5CO.12.1
Price:	0,00 EUR
Document(s):	paper

Abstract/Summary:

---

The price for solar photovoltaic (PV) generation fed into the network from a prosumer system is often lower than the price earned with self-consumption, thereby making it a critical input to the economic analysis. This study aims to create a self-consumption model that can be applied during the planning phase of a residential PV system in Sweden, when detailed knowledge of the system and/or building load are often limited. Eight regression models are trained using 1.08 million unique data points built from measured villa loads and simulated PV generation to calculate self-consumption with an hourly time step. Five of the trained models are found to deliver acceptable performance with an R2 greater than 0.9. Random Forest is the best performer with R2=0.985 and a mean absolute error of 1.5 percentage points while only requiring annual building load, PV generation, and latitude as inputs. This is noteworthy given that SC can vary by +/- 20 percentage points for a given solar fraction and is a marked improvement over previous estimation techniques. The resulting models are limited to Swedish villas and future studies using this technique in other regions and building types with local building loads are recommended.