As many dams were built around 1950, the expected life span of these dams are about tobe reached. With this, the need for monitoring and increased understanding of the damsstructural integrity increases. In order to prevent failures, two warning signals are defined;alert and alarm. The main difference being that the first indicates an unexpected changein behavior that needs to be addressed and evaluated in the near future, while the otherrequires that immediate action must be taken to ensure the safety of the dam.This report aims to evaluate the applicability of different models for designing alert values.In order to achieve this goal, two case studies have been performed. The first being onSchlegeis, an arch dam in Austria, and the second Storfinnforsen, a concrete buttress damin Sweden. The methods used are finite element modelling as well as data-based models.Data-based models work on the presumption that the dam behaviour is governed by variationsin environmental conditions such as temperature and water level. The report hasevaluated two commonly used data-based models, hydrostatic thermal time (HTT) and hydrostaticseasonal time (HST), as well as two machine learning based models artificial neuralnetworks (ANN) and boosted regression trees (BRT).The programs used in this report are BRIGADE plus for finite element method and MATLABfor the multi-linear regression analyses HTT and HST, as well as boosted regressiontrees. The neural networks were constructed in Python using TensorFlow and Keras API.The result from the case studies is that the commonly used data-based models HST andHTT perform well enough for creating predictions and alert levels when given a sufficientamount of historical data, approximately 3-5 years. Machine learning such as artificial neuralnetworks while comparable in prediction quality does not further increase the understandingof the dam behaviour and can due to the complexity of designing an appropriate networkstructure be less suited for this type of analysis. Finite element models can also capturethe behavior of the dam rather well. It is however not as accurate as data-based modelswhen sufficient data is available. An FE-model should be used for definition of alert valueswhen insufficient data exists after the dam conditions have been significantly altered, orwhen newly constructed. The main advantage that machine learning provides is that theyperform better for non-linear behavior than multi-linear regression.
| Identifer | oai:union.ndltd.org:UPSALLA1/oai:DiVA.org:kth-289385 |
| Date | January 2020 |
| Creators | Nilsson, Isak, Sandström, Leonard |
| Publisher | KTH, Betongbyggnad |
| Source Sets | DiVA Archive at Upsalla University |
| Language | English |
| Detected Language | English |
| Type | Student thesis, info:eu-repo/semantics/bachelorThesis, text |
| Format | application/pdf |
| Rights | info:eu-repo/semantics/openAccess |
| Relation | TRITA-ABE-MBT ; 20490 |
Page generated in 0.0023 seconds