Solar Irradiance Probabilistic Forecasting Using Machine Learning, Metaheuristic Models and Numerical Weather Predictions

International audienceSolar-power-generation forecasting tools are essential formicrogrid stability, operation, and planning. The prediction ofsolar irradiance (SI) usually relies on the time series of SI andother meteorological data. In this study, the considered microgridwas a combined cold- and power-generation system, located inTahiti. Point forecasts were obtained using a particle swarmoptimization (PSO) algorithm combined with three stand-alonemodels: XGboost (PSO-XGboost), the long short-term memory neuralnetwork (PSO-LSTM), and the gradient boosting regression algorithm(PSO-GBRT). The implemented daily SI forecasts relied on an hourlytime-step. The input data were composed of outputs from thenumerical forecasting model AROME (Météo France) combined withhistorical meteorological data. Our three hybrid models werecompared with other stand-alone models, namely, artificial neuralnetwork (ANN), convolutional neural network (CNN), random forest(RF), LSTM, GBRT, and XGboost. The probabilistic forecasts wereobtained by mapping the quantiles of the hourly residuals, whichenabled the computation of 38%, 68%, 95%, and 99% predictionintervals (PIs). The experimental results showed that PSO-LSTM hadthe best accuracy for day-ahead solar irradiance forecastingcompared with the other benchmark models, through overalldeterministic and probabilistic metrics