Temporal Networks‐Based Approach for Nonstationary Hydroclimatic Modeling and its Demonstration With Streamflow Prediction

Abstract

Lack of stationarity in most of the hydroclimatic variables is no longer a topic of debate rather a reality. It may be hypothesized that alternative methodologies are needed to deal with such nonstationarity and to improve the skill of hydroclimatic modeling/prediction. We propose the concept of temporal networks in hydroclimatic modeling as a potential solution to this problem. As a typical case, complex association among different hydroclimatic variables and streamflow is considered as an illustrative problem. Evolution of temporal networks over time, obtained through Graphical Modeling (GM), depicts the changes in the model inputs as well as model parameters over time. The proposed concept indicates that the time interval after which the model needs to be updated/recalibrated, referred to as Optimum Recurrence Interval (ORI), is problem specific and is optimized to achieve the best model performance. The proposed concept not only depicts a notable change in the potential predictors for the high and low flow months, but also establishes the different extent of temporal variability for different months, and hence the ORI of model recalibration. As compared to its time‐invariant counterpart, the temporal networks‐based approach shows higher efficacy in capturing the extreme flow events due to its inherent time‐varying characteristics. We recommend the concept of temporal networks to be promising in the context of climate change to capture the time‐varying association. In general, the concept can be applied to other hydroclimatic variables where a time‐varying association is expected due to various reasons including the impacts of climate change.

Feedback From Vegetation to Interannual Variations of Indian Summer Monsoon Rainfall

“…By doing this multiple times and computing mean and standard deviation of MI s and TE s , we can perform a t ‐test on these quantities to evaluate statistical significance thresholds for MI and TE at a 95% confidence level. It should be noted that concepts of network have been used extensively in hydrometeorology (e.g., Boers et al., 2013; Dai et al., 2019; Dutta & Maity, 2020; Keshtkar et al., 2013; Konapala & Mishra, 2017; Morrison & Stone, 2014); however, applications of causality based networks are limited (Chauhan & Ghosh, 2020; Goodwell & Kumar, 2017; Jiang & Kumar, 2019).…”

Dynamic Bayesian-Network-Based Approach to Enhance the Performance of Monthly Streamflow Prediction Considering Nonstationarity

“…Training Period Testing Period NGM-BNs Here, we also developed the NGM-BN models separately for different months of the year considering cyclostationarity, which is called a month-wise strategy of prediction [31]. The corresponding results are shown in Figures 5 and 6 and Table 2.…”

Information Content of Hydrologic Data across Space: Streamflow Predictions using Machine Learning

“…Recently, machine learning (ML) has gained popularity among hydrologists (Karpatne, 2018;Kratzert et al, 2018a;Kratzert et al, 2018b;Chandawala et al, 2019;Kratzert et al, 2019a;Kratzert et al, 2019b;Bennet & Nijssen, 2020;Dutta and Maity, 2020;Konpala et al, 2020;Fang et al, 2021;Gauch et al, 2021a;Gauch et al, 2021b;Herath et al, 2021;Lee et al, 2021;Razavi 2021;Sadler et al, 2022). In some studies, ML has been used as a tool for searching some optimal conceptual/process-based representation of a watershed hydrologic system (e.g., Chandawala et al, 2019), but in most of the recent studies, Long-Short Memory Network (LSTM; a variant of recurrent neural networks which is especially suitable for time series prediction) has been used to predict streamflow.…”