A Model Tree Generator (MTG) Framework for Simulating Hydrologic Systems: Application to Reservoir Routing

Abstract

Data-driven algorithms have been widely used as effective tools to mimic hydrologic systems. Unlike black-box models, decision tree algorithms offer transparent representations of systems and reveal useful information about the underlying process. A popular class of decision tree models is model tree (MT), which is designed for predicting continuous variables. Most MT algorithms employ an exhaustive search mechanism and a pre-defined splitting criterion to generate a piecewise linear model. However, this approach is computationally intensive, and the selection of the splitting criterion can significantly affect the performance of the generated model. These drawbacks can limit the application of MTs to large datasets. To overcome these shortcomings, a new flexible Model Tree Generator (MTG) framework is introduced here. MTG is equipped with several modules to provide a flexible, efficient, and effective tool for generating MTs. The application of the algorithm is demonstrated through simulation of controlled discharge from several reservoirs across the Contiguous United States (CONUS).

Status and prospects for drought forecasting: opportunities in artificial intelligence and hybrid physical–statistical forecasting

“…AI models commonly used for drought forecasting or similar applications with the potential to be used for drought prediction include fuzzy logic based models ( [126–129]), genetic algorithm ([130,131]), genetic programming ([132]), clustering methods such as K-means and nearest neighbour [133–135], as well as a variety of ML models such as artificial neural networks (ANNs; [26,118,136,142–]), support vector regression ([72]), support vector machine (SVM; [143,144]), decision tree ([145,146]) and random forest ([147–151]). More recently, boosting algorithms, such as XGBoost [152] and deep generative models [153–155] including variational autoencoders (VAEs; [156]), and generative adversarial networks (GANs; [157]) have shown great promise for drought forecasting performance improvement.…”

Investigation of hydrometeorological influences on reservoir releases using explainable machine learning methods

“…Since the time series observations such as inflow and storage contain both natural factors and human operating management experience, data-driven ML models have been increasingly being applied in assist of the reservoir management (Chang et al, 2016;Liu et al, 2017Liu et al, , 2019Bozorg-Haddad et al, 2018;Uysal et al, 2018;Niu et al, 2019;Zolfaghari and Golabi, 2021). Many studies have explored the applicability of using the linear regression models (e.g., ridge and lasso regressions) and decision treesbased models (e.g., random forest, extreme gradient boosting tree) to forecast the controlled reservoir releases (Zhang et al, 2018;Rahnamay Naeini et al, 2020;. In these ML models, the interpretation of the prediction is straightforward because these models have a transparent process and internal logic to allow end-users to understand the mathematical mapping from inputs to outputs.…”