Data-Driven Modelling of PV Production to Increase Forecast Accuracy of Multi-Inverter Plants

The accuracy of photovoltaic (PV) production forecasts is critical when implemented in codes that manage complex energy systems with renewable energy sources (RES). Generally, simple analytical models estimate the non-dispatchable energy production for applications such as optimized scheduling of multi-energy systems and energy management systems (EMS), using solar radiation and environmental temperature as inputs. This work aims to assess the impact of overcoming this simple approach, training data-driven models of the PV production output with environmental forecasts and real PV production measurements. Three machine learning (ML) algorithms are considered: one multiple linear regression (MLR), one multi-layer perceptron (MLP), and one extreme gradient boosting (XGBoost). Several models are trained and their performances are compared, using a complete 3-year real dataset of 6 different PV plants, located in different regions of Italy. Results show that the data-driven approach always outperforms the conventional one, and the best ML model obtained is an MLP, reaching 8.49% for the mean absolute error (MAE%), 29.74 kW for the root mean square error (RMSE), and 0.62 for the coefficient of determination (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$R^{2}$</tex>). Moreover, this paper shows the optimal handling of a dataset used to estimate the production of an entire solar plant, consisting of multiple inverters. The modeling approach for individual inverters has proven to be generalizable to multiple plants and robust enough to model accurately, overcoming plant-specific issues and improving performances up to a maximum of 8.51% when compared with an overall plant modeling approach.