In this paper, solar irradiance short-term forecasts were performed considering time horizons ranging from 5 min to 30 min, under a 5 min timestep. Global Horizontal Irradiance (GHI) and Direct Normal Irradiance (DNI) were computed using Deep Neural Networks with One-Dimensional Convolution (CNN-1D), Long-Short Term Memory (LSTM) and CNN-LSTM layers on the benchmarking dataset FOLSOM, which is formed by predictors obtained by recursive functions on the clear sky index time series and statistical attributes extracted from images collected by a camera pointed to the zenith, characterizing endogenous and exogenous variables, respectively. To analyze the endogenous predictors influence on the accuracy of the networks, the performance was evaluated for the cases with and without them. This analysis is motivated, to our best knowledge, by the lack of works that cite the FOLSOM dataset using deep learning models, and it is necessary to verify the impact of the endogenous and exogenous predictors in the forecasts results for this specific approach. The accuracy of the networks was evaluated by the metrics Mean Absolute Error (MAE), Mean Bias Error (MBE), Root Mean Squared Error (RMSE), Relative root-mean-squared error (rRMSE), Determination Coefficient (R²) and Forecast Skill (s). The network architectures using isolated CNN-1D and LSTM layers generally performed better. The best accuracy was obtained by the CNN-1D network for a horizon of 10 min ahead reaching an RMSE of 36.24 W/m², improving 11.15% on this error metric compared to the persistence model.

The presence of pollutants in our atmosphere has become one of humanity’s greatest challenges. These pollutants, produced primarily by burning fossil fuels, are detrimental to human health, our climate and agriculture. This work proposes the use of a spatiotemporal graph neural network, designed to forecast ozone concentration based on the GraphSAGE paradigm, to aid in our understanding of the dynamic nature of these pollutants’ production and proliferation in urban areas. This model was trained and tested using data from Houston, Texas, the United States, with varying numbers of time-lags, forecast horizons (1, 3, 6 h ahead), input data and nearby stations. The results show that the proposed GNN-SAGE model successfully recognized spatiotemporal patterns underlying these data, bolstering its forecasting performance when compared with a benchmarking persistence model by 33.7%, 48.7% and 57.1% for 1, 3 and 6 h forecast horizons, respectively. The proposed model produces error levels lower than we could find in the existing literature. The conclusions drawn from variable importance SHAP analysis also revealed that when predicting ozone, solar radiation becomes relevant as the forecast time horizon is raised. According to EPA regulation, the model also determined nonattainment conditions for the reference station.