Using Perceptron Network based on Autoregressive to Classify Evaporation Data

Abstract

The study of climatic conditions, atmospheric fluctuations and their effects is very important for diagnosing environmental and climatic features and their impact on various areas related to human life and other living organisms. In this study, the variable of the time series of evaporation quantities will be studied and classified using several atmospheric influences, such as the time series of air temperatures in the maximum and minimum cases, as well as the time series of relative humidity in the maximum and minimum cases, which are considered direct, cross-sectional influences on the time series of evaporation depending on The principle of regression between the input variables and the dependent variable. The time series is characterized by its reliance on the principle of autoregressive (AR), which generates a driving force in the form of longitudinal effects that help in making classifications for future time periods outside the study samples, but it conflicts with the nature of non-linear climate data, which may cause a state of uncertainty. Neural networks are one of the methods that deal well with non-linearity in data, and the Perceptron Neural Network (PNN) is one of the oldest neural networks that is mainly used to classify data in general and time series in particular, with a structure that is far from complexity and reduces effort and time. More than one type of learning function was used after choosing the appropriate transformation function for the nature of the study data. The types of functions used in the perceptron network, in addition to its structure, depend mainly on the principle of regression and its use in forecasting and classification. The use of climate time series data requires achieving homogeneity in the data to obtain the best classification results. Therefore, the data were aligned in time depending on the nature of the atmosphere and temperatures during the years of study to ensure greater homogeneity in the data and thus greater accuracy in the results. The study data were taken from the Iraqi agricultural meteorological station in Mosul station and were used as a real case in this study. The results of comparisons between seasons of data as well as different learning functions showed high accuracy at times and acceptable at other times, which makes it possible to use the perceptron network as a recommended matter to obtain accurate classification results that achieve the intended purpose of its use.