Variability and forecasting of air temperature in Elqui Valley (Chile)

A method to forecast air temperature T A in Elqui Valley (south of Chile’s Atacama Desert) using an artificial neural network (ANN) and meteorological time series data relevant to this zone, is proposed. This zone has one of the most sensitive climates in South America due to the influence of phenomena such as El Niño/La Niña , the Southeast Pacific Subtropical Anticyclone, Humboldt Current, and Madden–Julian Oscillation, in addition to the complex topography of Chilean east-west transverse valleys, and the Andes. We used a method that combines ANN and genetic algorithm (GA) for forecasting T A 1, 3, and 6 hours ahead. GA is introduced to optimize the weights update process in the proposed method. Our database contains 457,969 data, from 2004–2017, taken from eight stations throughout the valley and divided into three datasets: training set with 50% of overall data of each station, validation set with the subsequent 25% of data of each station, and prediction set with the last 25% of data. Several architectures were evaluated using the root-mean-square error (RMSE), mean absolute error (MAE) and correlation coefficient (R). The results show that the ANN+GA method represents a powerful technique for forecasting T A with RMSE from 0.99 to 4.51 [ ∘ C], MAE from 0.52 to 3.07 [ ∘ C], and R from 0.86 to 0.96. Also, we use the values obtained by ANN+GA method to investigate the spatial T A variability in elevation from the coast ∼ 30 [masl] to the Andean zone of this valley ∼ 5,000 [masl].