Deep Learning para estimativa na escala mensal das temperaturas do ar para a região Centro-Oeste do Brasil

Abstract

Among the climatic elements that influence agricultural production, there is the air temperature, which in turn depends on other factors and elements. In this sense, the study of the climate for the estimates of the minimum, average and maximum values of air temperature is important, especially in regions that do not have an acceptable density of meteorological stations, missing data or inconsistent values, as is the case of the Midwest region of Brazil. This work was carried out with the objective of studying the use of Deep Learning Artificial Neural Networks, capable of estimating the maximum (Tmax), minimum (Tmin) and average (Tmed) air temperatures, on a monthly scale, for the Midwest of Brazil. Data were obtained from 1990 to 2020, from 81 meteorological stations available in the states that comprise the Brazilian Midwest region. The data were subdivided into training, validation and test plots. The independent variables of altitude, latitude, longitude, relative humidity and wind speed were combined in nine scenarios (S1 to S9) to study the best combination and number of neurons for structuring the ANN layers. Three configurations of hidden layers and neurons in each layer were selected and analyzed. Regarding the output layer, it is composed of three neurons that relate to air temperatures (maximum, average and minimum) on the monthly scale. In this way, Deep Learning ANNs were generated for each of the nine scenarios, configuration of intermediate layers, for each month of the year (January to December) and air temperatures. The adjustments of the ANNs were evaluated for the portion of the test data and through statistical indices of Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Pearson's correlation coefficient (r), coefficient of determination (R²), agreement index (d) and performance index (PI). The best scenarios in the combination of independent variables for most months were S1, S2, S3, S4 and S6, which presented R² ≥ 70%. The ANNs developed in the configurations of hidden layers M3, M2 and M1, in that order, obtained the best statistical indices of r ≥ 70%; RMSE and MAE ≤ 1.0 °C; index (d) ≥ 0.8 and PI classified as excellent. Using a large volume of data linked to the application of Deep Learning, it was possible to estimate the monthly averages of air temperatures Tmax, Tmed and Tmin, with an acceptable degree of accuracy, for the Midwest region of Brazil.