Emprego de redes neurais artificiais na previsão de precipitação pluviométrica

Abstract

The meteorological services are relevant to various human activities that are significantly affected by climatic and meteorological conditions, for example, agriculture and tourism. Due to the high cost, meteorological stations are usually implemented to meet the demand of a wide region, making it difficult to accurately understand the meteorological information of small towns. Technological advances in the machine learning area make it possible to develop techniques to facilitate rain forecasting, especially on a local scale, aiming to make it more practical and accessible. This fact motivates several researches to use algorithms, such as Artificial Neural Networks (ANN), for the prediction of pluviometric precipitation. The objective of this work is the use of ANN for the forecast of pluviometric precipitation (millimeters of rain) in the region of the Cerro Largo Campus of the Federal University of the Fronteira Sul (UFFS). The variables air temperature, relative humidity, dew point temperature, atmospheric pressure, wind speed, wind direction, solar irradiation and rainfall for the algorithm training were selected, recorded at the surface level by the Meteorological Station located on the Campus Cerro Largo of UFFS, for a period of three years (2016-2019).The ANN used in this work are classified as Multilayer Perceptron and were trained by the backpropagation method. Two models were developed, one for classification (to predict the occurrence of precipitation) and the other for regression (to estimate the volume of rainfall). The classification model was evaluated from the confusion matrix, accuracy, precision, sensitivity, specificity and F-score. Considering these metrics, the results varied between 94.35% (sensitivity) and 77.60% (specificity) of correct answers in the validation step. To assess the performance of the regression model, the determination coefficient (R²), the Pearson correlation coefficient (R), the mean absolute error (MAE) and the mean square error (MSE) were used. The results were R² = 0.71 and R = 0.84, MAE = 0.34 mm/day and MSE = 0.22 mm²/day, considering the validation data. With this study, it is concluded that the ANN is a promising technique for analysis of meteorological data, with the potential to be implemented in applications that demand the forecast of pluviometric precipitation, as in an alert system of heavy rains, in an accessible way and with reduced costs.