Interval-based solar photovoltaic energy predictions: A single-parameter approach with direct radiation focus
Carlos Roldán-Blay, Manuel Francisco Abad-Rodríguez, Víctor Abad-Giner, Xavier Serrano‐Guerrero
Abstract
This study introduces a novel solar photovoltaic (PV) power generation forecasting method for residential installations, focusing on the direct radiation parameter. It has been rigorously compared against four established methods: Linear Regression (Alt1), Gradient Boosting (Alt2), Gradient Boosting with Lags (Alt3), and Long Short-Term Memory (LSTM) Network (Alt4). Alt1 utilizes a simple linear equation, while Alt2 employs ensemble learning with decision trees. Alt3 leverages past 24-h PV generation data, and Alt4 utilizes specialized recurrent neural networks to address challenges of long-term dependencies in time series forecasting. The proposed method showed better performance in 2022, with a Mean Absolute Error (MAE) of 0.1490 kW and a Coverage Probability (CP) of 91.55 %, demonstrating its reliability and consistency in forecasting. Additionally, it obtained an Average Width of Intervals (AWI) of 0.3365 kW. Furthermore, the method significantly boosted solar energy utilization in a residential case, increasing average solar panel generation by 61.33 kWh/year and reducing the average price by 0.0188 €/kWh. These results highlight its effectiveness in enhancing self-consumption and cutting energy costs, presenting a precise and user-friendly forecasting tool for the solar energy sector. Particularly advantageous for residential use, it facilitates optimized solar energy utilization, contributing to the transition towards sustainable energy practices.