Data preprocessing is an issue that is often recommended to create more uniform data to facilitate ANN learning, meet transfer function requirements, and avoid computation problems. In ANN typical transfer functions, such as the sigmoid logistic function, or the hyperbolic tangent function, cannot distinguish between two very large values, because both yield identical threshold output values of 1.0. It is then necessary to normalize (preprocess) the inputs and outputs of a network. Usually normalization is carried out using the minimum and maximum values obtained in the in-sample (calibration) data. Such a network will result in absurd output, if the out-of-sample (test) data contain values that are beyond the in-sample data range. This ultimately limits the application of ANN in forecasting continuously increasing or decreasing time series data. This study will present a novel and successful application of ANN, which is trained by the error backpropagation algorithm, in the context of forecasting beyond in-sample data range. The emphasis here is on continuously decreasing hydraulic pressure data forecasting that are observed at Mizunami underground research laboratory construction site, Japan. The ANN utilizes the sigmoid logistic function in its hidden and output layers.
|Title of host publication||Focus on Artificial Neural Networks|
|Editors||John A. Flores|
|Number of pages||18|
|Publication status||Published - 12 Feb 2021|
|Name||Focus on Artificial Neural Networks|