Accuracy verification of TimesNet that predicts water levels applied to long-term observed data with multiple periodic patterns

Abstract

This study applied TimesNet (one of the neural network models) to long-term, large-sample-size, and time-series data of water levels, observed in drainage management with pump operations in a low-lying watershed. Time series data, in general, contain complicated and multiple patterns. TimesNet is a prediction method that can automatically extract multiple periodic features, contained in the input data, using spectral analysis, then transforming to two-dimensional information that arranges continuous data according to these periodic features, and finally predicting future continuous data. Applied to long-term water level data, this method performed worse in prediction accuracy, when compared with conventional prediction methods. However, it was confirmed that the method better reproduced flood waveforms during the specific drainage period involving the largest flood event.