Applicering av Long Short-Term Memory för prediktion av markdeformation på svensk järnväg : Utveckling av ett artificiellt neuralt nätverk för prediktion av kommande marksättningar på järnvägssträckan mellan Mölndal och Torrekulla

Sammanfattning: The purpose of this study is to evaluate whether it is possible to predict future land subsidence on the railway line between Mölndal and Torrekulla. The prediction was made using Long Short-Term Memory; an artificial neural network with RNN architecture. The study used the research strategy "design and creation" to develop a neural network in the form of an artefact, it used a document study as a data collection method and a quantitative data analysis. To enable a successful prediction, the study used secondary data consisting of already collected geological data from a three-year project that began in January 2016 and ended in September 2018. Through a technology called Interferometric synthetic-aperture radar (InSAR), an active satellite system that transmits radar signals to the earth for measurements of the ground level. The already collected InSAR data was supplemented with weather data from a nearby weather station, the artefact could achieve a predictive accuracy of 99.1% of all values within an approved prediction interval from the Swedish Transport Administration of + -1 millimetres. The result of the created artefact also showed that the model also managed to achieve a prediction accuracy of 86% of all values within the range + -0.4 millimetres. With the model's prediction accuracy score and an RMSE average of 0.22 for all points, the assessment was made that a prediction of future land subsidence is possible on the railway line between Mölndal and Torrekulla.