Ana Dodig
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Ana Dodig has a total of 14 co-authors. Here are the top co-authors by number of shared publications:
- Milan Stojković
Institute for Artificial Intelligence R&D of Serbia, Fruškogorska 1, 21000 Novi Sad, Serbia
4 shared publications - Luka Vinokić
Institute for Artificial Intelligence R&D of Serbia, Fruškogorska 1, 21000 Novi Sad, Serbia
3 shared publications - Milan Dotlić
Jaroslav Černi Institute Jaroslava Černog 80 11226 Pinosava Belgrade Serbia
3 shared publications - Veljko Prodanovic
Institute for Artificial Intelligence Research and Development of Serbia, Fruskogorska 1, Novi Sad 21000, Serbia
3 shared publications - Elisa Ricci
University of Trento, Trento, Italy and Fondazione Bruno Kessler, Trento, Italy
1 shared publication - Goran Kvascev
University of Belgrade
1 shared publication
Latest Publications
- Open Access
- 9 September 2026
- Published by Elsevier in SSRN Electronic Journal
Weather-Informed Streamflow Forecasting with Temporal Convolution Enhanced Graph Transformer Networks
In recent years, machine learning has emerged as a powerful tool for high-accuracy streamflow forecasting. However, most existing approaches do not fully exploit the inherent directional and hierarchical graph structure of hydrological systems. This work introduces Temporal Convolution Graph Transformer Network (TC-GTN), a novel hybrid model for streamflow forecasting that specifically integrates temporal convolution for temporal pattern extraction with a graph transformer for advanced relational reasoning. It utilizes structured graph representation of river network with accompanying meteorological stations where the transformer's attention mechanism is critical for better understanding of interactions between different nodes/stations and for capturing self-dependencies within each station. To benchmark performance, TC-GTN is compared against two strong baselines: a Graph Transformer without a temporal encoder and a Graph Transformer with a long short-term memory encoder. Experiments on the Drina - Lim River Basin dataset demonstrate that the TC-GTN achieves high performance, with an average coefficient of determination of 0.88 across all hydrological stations, outperforming baselines in 75% of evaluated cases. High-flow event analysis (streamflows above the 75th percentile) further shows that TC-GTN maintains strong accuracy during extreme conditions, achieving average coefficient of determination above 0.75 at major hydropower and flood monitoring stations. This robustness is further strengthened by the quantile-loss formulation, which provides balanced and stable performance for both overall and high-flow conditions, essential for effective flood risk mitigation and sustainable hydropower management. Beyond its hydrological application, TC-GTN offers a general design pattern for spatio-temporal prediction on directed heterogeneous graphs, contributing to the broader literature on spatio-temporal representation learning.
- 9 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessWeather-Informed Streamflow Forecasting with Temporal Convolution Enhanced Graph Transformer Networks
In recent years, machine learning has emerged as a powerful tool for high-accuracy streamflow forecasting. However, most existing approaches do not fully exploit the inherent directional and hierarchical graph structure of hydrological systems. This work introduces Temporal Convolution Graph Transformer Network (TC-GTN), a novel hybrid model for streamflow forecasting that specifically integrates temporal convolution for temporal pattern extraction with a graph transformer for advanced relational reasoning. It utilizes structured graph representation of river network with accompanying meteorological stations where the transformer's attention mechanism is critical for better understanding of interactions between different nodes/stations and for capturing self-dependencies within each station. To benchmark performance, TC-GTN is compared against two strong baselines: a Graph Transformer without a temporal encoder and a Graph Transformer with a long short-term memory encoder. Experiments on the Drina - Lim River Basin dataset demonstrate that the TC-GTN achieves high performance, with an average coefficient of determination of 0.88 across all hydrological stations, outperforming baselines in 75% of evaluated cases. High-flow event analysis (streamflows above the 75th percentile) further shows that TC-GTN maintains strong accuracy during extreme conditions, achieving average coefficient of determination above 0.75 at major hydropower and flood monitoring stations. This robustness is further strengthened by the quantile-loss formulation, which provides balanced and stable performance for both overall and high-flow conditions, essential for effective flood risk mitigation and sustainable hydropower management. Beyond its hydrological application, TC-GTN offers a general design pattern for spatio-temporal prediction on directed heterogeneous graphs, contributing to the broader literature on spatio-temporal representation learning.
- Open Access
- 9 May 2026
- Published by Elsevier in SSRN Electronic Journal
Exploring uncertainties across the modelling chain in ML-based streamflow forecasting
Operational streamflow forecasts underpin flood preparedness and reservoir operations, yet their utility is often constrained by poorly characterized and attributed predictive uncertainty. In machine-learning-based forecasting, uncertainty is frequently omitted or reported as a single aggregate output, leaving it unclear which parts of the end-to-end forecasting chain drive overconfidence and forecast degradation, particularly with increasing lead time. In this work, we develop an end-to-end uncertainty decomposition framework for operational streamflow forecasting that attributes predictive uncertainty across meteorological forcing choice, feature design, model architecture, hyperparameter optimization, and training variability, evaluated across multi-day horizons. The decomposition reveals a systematic, horizon-dependent shift in dominant uncertainty sources, with forcing-related contributions increasing with lead time while model-structure and feature choices remain influential at shorter horizons. During high-flow events, predictive intervals remain essential because pipeline heterogeneity can bias the central estimate even when ensemble dispersion widens appropriately. Tuning contributes little to the uncertainty budget but strongly affects compute–skill trade-offs, with Bayesian optimization delivering the most favorable cost–benefit performance under the tested constraints. Together, these results provide actionable guidance for operational freshwater management, showing where investment yields the largest reliability gains: model design at short lead times and forcing quality at longer lead times. This guidance can reduce the risk of costly or unsafe decisions in flood preparedness, reservoir operation, and other critical decision-making contexts in water management.
- 9 May 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessExploring uncertainties across the modelling chain in ML-based streamflow forecasting
Operational streamflow forecasts underpin flood preparedness and reservoir operations, yet their utility is often constrained by poorly characterized and attributed predictive uncertainty. In machine-learning-based forecasting, uncertainty is frequently omitted or reported as a single aggregate output, leaving it unclear which parts of the end-to-end forecasting chain drive overconfidence and forecast degradation, particularly with increasing lead time. In this work, we develop an end-to-end uncertainty decomposition framework for operational streamflow forecasting that attributes predictive uncertainty across meteorological forcing choice, feature design, model architecture, hyperparameter optimization, and training variability, evaluated across multi-day horizons. The decomposition reveals a systematic, horizon-dependent shift in dominant uncertainty sources, with forcing-related contributions increasing with lead time while model-structure and feature choices remain influential at shorter horizons. During high-flow events, predictive intervals remain essential because pipeline heterogeneity can bias the central estimate even when ensemble dispersion widens appropriately. Tuning contributes little to the uncertainty budget but strongly affects compute–skill trade-offs, with Bayesian optimization delivering the most favorable cost–benefit performance under the tested constraints. Together, these results provide actionable guidance for operational freshwater management, showing where investment yields the largest reliability gains: model design at short lead times and forcing quality at longer lead times. This guidance can reduce the risk of costly or unsafe decisions in flood preparedness, reservoir operation, and other critical decision-making contexts in water management.
- Open Access
- 15 April 2026
- Published by Elsevier in SSRN Electronic Journal
Improved Spatio-Temporal Streamflow Forecasting with Temporal Encoder Enhanced Graph Transformer Network
In recent years, machine learning has emerged as a powerful tool for achieving high-accuracy hydrological streamflow forecasting. However, most existing approaches do not fully exploit the inherent directional and hierarchical graph structure of hydrological systems. This work introduces TC-GTN (Temporal Convolution Graph Transformer Network), a novel hybrid model for streamflow forecasting that integrates temporal convolution for temporal pattern extraction with a graph transformer for advanced relational reasoning. It utilizes structured graph representation of river network with accompanying meteorological stations where the transformer's attention mechanism are critical for better understanding of interactions between different nodes/stations and for capturing self-dependencies within each station. To benchmark performance, TC-GTN is compared against two strong baselines: a Graph Transformer without a temporal encoder (GTN) and a Graph Transformer with a long short-term memory encoder (LS-GTN). Comprehensive experiments on the Drina-Lim River Basin dataset demonstrate that the TC-GTN achieves high performance, with an average R2 of 0.88 across all hydrological stations, outperforming two strong baselines in 75% of all evaluated cases. High-flow event analysis (flows above the 75th percentile) further shows that TC-GTN maintains strong accuracy during extreme conditions, achieving average R2 values above 0.75 at major hydropower and flood monitoring stations. This robustness is further strenghtened by the quantile-loss formulation, which provides the most balanced and stable performance for both overall and high-flow conditions, an essential feature for effective flood risk mitigation and sustainable hydropower management.
- 15 April 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessImproved Spatio-Temporal Streamflow Forecasting with Temporal Encoder Enhanced Graph Transformer Network
In recent years, machine learning has emerged as a powerful tool for achieving high-accuracy hydrological streamflow forecasting. However, most existing approaches do not fully exploit the inherent directional and hierarchical graph structure of hydrological systems. This work introduces TC-GTN (Temporal Convolution Graph Transformer Network), a novel hybrid model for streamflow forecasting that integrates temporal convolution for temporal pattern extraction with a graph transformer for advanced relational reasoning. It utilizes structured graph representation of river network with accompanying meteorological stations where the transformer's attention mechanism are critical for better understanding of interactions between different nodes/stations and for capturing self-dependencies within each station. To benchmark performance, TC-GTN is compared against two strong baselines: a Graph Transformer without a temporal encoder (GTN) and a Graph Transformer with a long short-term memory encoder (LS-GTN). Comprehensive experiments on the Drina-Lim River Basin dataset demonstrate that the TC-GTN achieves high performance, with an average R2 of 0.88 across all hydrological stations, outperforming two strong baselines in 75% of all evaluated cases. High-flow event analysis (flows above the 75th percentile) further shows that TC-GTN maintains strong accuracy during extreme conditions, achieving average R2 values above 0.75 at major hydropower and flood monitoring stations. This robustness is further strenghtened by the quantile-loss formulation, which provides the most balanced and stable performance for both overall and high-flow conditions, an essential feature for effective flood risk mitigation and sustainable hydropower management.
- 18 July 2025
- Published by Springer Nature
- p. 42-61
Evaluating Bayesian Approaches for Water Quality Classification: A Comparative Study
- 18 July 2025
- Published by Springer Nature
- p. 42-61
Evaluating Bayesian Approaches for Water Quality Classification: A Comparative Study
- Open Access
- 12 April 2024
- Published by IWA Publishing in Journal of Hydroinformatics
- Vol. 26 (5) , 1059-1079
A novel machine learning-based framework for the water quality parameters prediction using hybrid long short-term memory and locally weighted scatterplot smoothing methods
Water quality prediction is crucial for effective river stream management. Dissolved oxygen, conductivity and chemical oxygen demand are vital chemical parameters for water quality. Development of machine learning (ML) and deep learning (DL) methods made them widely used in this domain. Sophisticated DL techniques, especially long short-term memory (LSTM) networks, are required for accurate, real-time multistep prediction. LSTM networks are effective in predicting water quality due to their ability to handle long-term dependencies in sequential data. We propose a novel hybrid approach for water quality parameters prediction combining DL with data smoothing method. The Sava river at the Jamena hydrological station serves as a case study. Our workflow uses LSTM networks alongside LOcally WEighted Scatterplot Smoothing (LOWESS) technique for data filtering. For comparison, Support Vector Regressor (SVR) is used as the baseline method. Performance is evaluated using Root Mean Squared Error (RMSE) and Coefficient of Determination R2 metrics. Results demonstrate that LSTM outperforms the baseline method, with an R2 up to 0.9998 and RMSE of 0.0230 on the test set for dissolved oxygen. Over a 5-day prediction period, our approach achieves R2 of 0.9912 and RMSE of 0.1610 confirming it as a reliable method for water quality multistep parameters prediction.
- 12 April 2024
- Published by IWA Publishing in Journal of Hydroinformatics
- Vol. 26 (5) , 1059-1079
Open AccessA novel machine learning-based framework for the water quality parameters prediction using hybrid long short-term memory and locally weighted scatterplot smoothing methods
Water quality prediction is crucial for effective river stream management. Dissolved oxygen, conductivity and chemical oxygen demand are vital chemical parameters for water quality. Development of machine learning (ML) and deep learning (DL) methods made them widely used in this domain. Sophisticated DL techniques, especially long short-term memory (LSTM) networks, are required for accurate, real-time multistep prediction. LSTM networks are effective in predicting water quality due to their ability to handle long-term dependencies in sequential data. We propose a novel hybrid approach for water quality parameters prediction combining DL with data smoothing method. The Sava river at the Jamena hydrological station serves as a case study. Our workflow uses LSTM networks alongside LOcally WEighted Scatterplot Smoothing (LOWESS) technique for data filtering. For comparison, Support Vector Regressor (SVR) is used as the baseline method. Performance is evaluated using Root Mean Squared Error (RMSE) and Coefficient of Determination R2 metrics. Results demonstrate that LSTM outperforms the baseline method, with an R2 up to 0.9998 and RMSE of 0.0230 on the test set for dissolved oxygen. Over a 5-day prediction period, our approach achieves R2 of 0.9912 and RMSE of 0.1610 confirming it as a reliable method for water quality multistep parameters prediction.
- 1 February 2024
- Published by Springer Nature
- p. 362-374
Digital Solution to Estimate Solar Power Potential of Rooftops in City of Belgrade
- 1 February 2024
- Published by Springer Nature
- p. 362-374
Digital Solution to Estimate Solar Power Potential of Rooftops in City of Belgrade