Hybrid models for time series forecasting of the dam monitoring data

Resumo

Time series forecasting is a result that contributes to analysis and decision-making and can be applied
in several areas of knowledge. In the area of dam structural safety, this practice is little explored, although the
equipment used in the monitoring feeds an extensive database. This work aims to apply a hybrid methodology for
forecasting time series, integrated into the processing of data collected by monitoring instruments of a concrete
dam. Wavelet Decomposition will perform the time series processing, then to separate the series components, an
Autoregressive Integrated Moving Average model will be fitted. Residuals resulting from this
mathematical/statistical model will be modeled through Artificial Neural Networks of Radial Basis Functions. The
linear combination of these models will generate the time series forecast itself. The combination weights will be
defined by solving a nonlinear programming problem. The investigated approaches will be compared and selected
according to the smallest mean absolute percentage error measure. The partial series models do not need to have
high performance for the prediction proposed to be satisfactory. The proposed approaches for predicting the test
set have MAPE of less than 0.57%, while ARIMA and ANN-RBF models used separately reached values of up to
4.33%. The results indicate gains in forecasting assertiveness, aiding decision-making, which aims to create
preventive measures to ensure dam safety.