Naslov (eng)

Exploring Machine Learning Approaches for Predicting the Resilience of Water Resources System under Hazardous Events

Autor

Kazaković, Aleksandra
Simić, Višnja
Ćirović, Vukašin

Publisher

Opis (eng)

Climate change-induced extreme weather events affect water systems, causing operational challenges and functional failures. Earthquakes and consequential landslides are other common causes of disruption. Assessing system resilience is crucial to avoid the failure of water systems. This research aimed to explore various approaches in developing a machine learning model to predict the water system's robustness and the system's recovery time from external or internal hazards (rapidity). The dataset was obtained by simulating the system dynamics model and hazard model of the hydroelectric power plant. A thorough examination of the data preceded the model’s construction. Random Forest (RF) and Artificial Neural Network (ANN) models were fitted to the training dataset. The ANN model fine-tuned using the Keras-tune approach yielded a high R2 score. To overcome the imbalanced dataset problem, the synthetic minority oversampling technique (SMOGN) was utilized. Due to highly imbalanced data for the Robustness values over 0.2, even the implementation of the SMOGN technique could not yield an R2 score over 0.8. The dataset was also modeled as a classification problem, using K-means clustering to group Robustness and Rapidity values into classes. The best classification model obtained was compared with the existing Fuzzy rule-based model which enables comprehensible reasoning using natural language. Precision, Recall and F1 score values of the ANN Keras-tuned model were better than the same metrics for the fuzzy model, but the explanatory capability was lost.

Jezik

engleski

Datum

2024

Licenca

© All rights reserved

Predmet

exploratory data analysis, imbalanced data, machine learning

Deo kolekcije (1)

o:243

Institut za vodoprivredu "Jaroslav Černi"