Multilevel prediction of missing time series dam displacements data based on artificial neural networks voting evaluation

The dams are very important objects (production of electric energy, flood management, drought control, etc.) but they are also a great danger for areas downstream because there is always risk of dam failure. To prevent dam failure it is important to perform regular dam monitoring. Precise geodetic surveying of a number of discrete points, which accurately depict the characteristics of the dam, is a common monitoring method. It is not uncommon that some discrete points can't be monitored because of different obstacles or equipment limitations but it is possible to interpolate those dam displacements using prediction. In this research a non conventional approach for interpolation of missing dam displacements data using artificial neural networks (ANNs) is presented. This approach combines spatial and temporal aspects of data to its benefits to give good results using ANNs. Three ANN types (Feed Forward Back Propagation, Cascade Forward Back Propagation and Layer Recurrent Back Propagation) were used for dam displacements prediction. Additionally, a voting system with four functions (“minimum”, “maximum”, “mean of the closest two” and “mean of three”) is introduced to improve prediction results. The research showed that ANNs in combination with the voting system can provide precise prediction for missing dam displacements.