Alapfogalmak
Developed a data-parsimonious machine learning model for near-term solar irradiance forecasting under data transmission constraints.
Kivonat
The study introduces a machine learning model for short-term solar irradiance forecasting, focusing on data parsimony. The model incorporates sky camera images and meteorological station data to predict irradiance deviations from the persistence of cloudiness model. Key insights include the importance of feature selection, the impact of input sequence length on model performance, and the effectiveness of a noise signal in improving forecast accuracy. The study achieves a mean absolute error of 74.34 W/m2 compared to a baseline of 134.35 W/m2.
Abstract:
Developed a machine learning model for short-term solar irradiance forecasting.
Utilized sky camera images and meteorological station data for predictions.
Achieved a mean absolute error of 74.34 W/m2 compared to baseline models.
Introduction:
United States' target for net-zero greenhouse gas emissions by 2035 necessitates renewable energy integration.
Accurate solar PV forecasts crucial for grid stability amid increasing renewable sources.
Model Architecture:
CNN-LSTM structure used with dropout layer to prevent overfitting.
LSTM identifies sequential patterns, dense layers transform intermediate representations.
Results:
Model achieved MAE of 75.20 W/m2, outperforming baseline POC model at 134.35 W/m2.
Feature importance analysis identified key features impacting prediction accuracy.
Statisztikák
Five years of data from NREL Solar Radiation Research Laboratory were used to create three rolling train-validate sets.
For the chosen test data, the model achieves a mean absolute error of 74.34 W/m2 compared to a baseline 134.35 W/m2 using the persistence of cloudiness model.