Characterizing Coastal Stratocumulus Cloud Impacts On Solar Irradiance For Short-Term Forecasting
Abstract
Short-term solar forecasting is essential for optimizing photovoltaic systems. In coastal regions, the variability caused by morning stratocumulus clouds challenges precise forecasts. We use meteorological data from Antofagasta, Chile, to detect important variables for solar irradiance: correlations highlight a strong influence of wind speed, net radiation, and relative humidity on solar irradiance, and Convergent Cross Mapping identifies a causal influence of dry bulb and dew point temperatures and atmospheric pressure. These variables and their importance are incorporated into a Long Short-Term Memory model with attention to predict time series of solar irradiance, demonstrating an improved performance over the model without attention, achieving lower MSE, MAE, RMSE, and MAPE values, as well as reduced variability across cross-validation folds, indicating both higher accuracy and more consistent results. SHapley Additive exPlanations analysis confirms a greater variable importance for the attention model.
Keywords
Solar forecasting, Coastal stratocumulus, Solar variability, Machine Learning