Performance analysis of predictive models for determining standard behavior of pluviometric series

Abstract

The present study deals with the verification of different predictive analysis models to explore and understand the standard behaviour of rainfall series, using the Engenheiro Ãvidos reservoir as the study area. First, the rainfall data was collected, and the gaps were filled in, together with the analysis of the study series consistencies. Soon after, the Thiessen method was applied to determine the precipitation corresponding to the reservoir. A historical series equivalent to a period of 57 years (1963-2019) was obtained. The second stage of the study proceeded with the historical data statistical analysis, performed with the R Studio software, and finally, the predictive analysis using the Auto-Regressive Integrated Moving Average Model (ARIMA), Exponential Smoothing (SE), Networks Neural (RN) and Multiple Linear Regression (RLM). With the results obtained, it was observed that all models showed the need for statistical adjustments, but that did not make them unfeasible for use. The precipitation showed better adherence to the observed data when compared with the test series in all models, with the RLM being the most appropriate. Precipitation forecasts rarely exceeded the observed values, being a more present characteristic for the ARIMA and RN models. Thus, the precipitation forecast obtained is potentially useful for studies involving water engineering, as well as for simulation and optimization models.