Gap Repair In Water Level Measurement DataUsing Neural Networks

Abstract

The paper presents a method for completing missing values in time series by applying neural network computation. Conventional methods to complete fragmentary time series like linear interpolation are chosen for small amounts of missing values; problems occur when there are larger intervals of missing values. So far such gaps are repaired by estimating techniques that use parameter estimation of mathematical models. However such an approach fails when there is not enough information for calibrating the model or when the model is too simplified for reliably completing the data series. Neural networks with their generalisation and memory properties are predestined for this category of problems. In the proposed method a two-layer feed-forward network is used to build spatial patterns out of time series intervals with some variations on one hand and on the other hand directly model processing is used for producing time series information as a function of additional and external information. A representative values approach combines characteristics of both the spatial and the direct approach. Parameters in the approach are the number of representativ e values in each input pattern and the sampling distance of representativ e values in the original time series. The discussed example refers to water level data that contains a basic tidal oscillation (which is computable by astronomical components) and additional set-up which is mainly a function of local meteorological conditions like wind speed and wind direction.