Real-time rainfall and runoff prediction by integrating BC-MODWT and automatically-tuned DNNs: Comparing different deep learning models

Journal of Hydrology, Год журнала: 2024, Номер 631, С. 130804 - 130804

Опубликована: Янв. 26, 2024

Язык: Английский

---

Enhancing hydrological modeling with transformers: a case study for 24-h streamflow prediction

Water Science & Technology, Год журнала: 2024, Номер 89(9), С. 2326 - 2341

Опубликована: Апрель 4, 2024

ABSTRACT In this paper, we address the critical task of 24-h streamflow forecasting using advanced deep-learning models, with a primary focus on transformer architecture which has seen limited application in specific task. We compare performance five different including persistence, long short-term memory (LSTM), Seq2Seq, GRU, and transformer, across four distinct regions. The evaluation is based three metrics: Nash–Sutcliffe Efficiency (NSE), Pearson's r, normalized root mean square error (NRMSE). Additionally, investigate impact two data extension methods: zero-padding model's predictive capabilities. Our findings highlight transformer's superiority capturing complex temporal dependencies patterns data, outperforming all other models terms both accuracy reliability. Specifically, model demonstrated substantial improvement NSE scores by up to 20% compared models. study's insights emphasize significance leveraging deep learning techniques, such as hydrological modeling for effective water resource management flood prediction.

Язык: Английский

---

Spatio-Temporal Graph Neural Networks for Streamflow Prediction in the Upper Colorado Basin

Hydrology, Год журнала: 2025, Номер 12(3), С. 60 - 60

Опубликована: Март 17, 2025

Streamflow prediction is vital for effective water resource management, enabling a better understanding of hydrological variability and its response to environmental factors. This study presents spatio-temporal graph neural network (STGNN) model streamflow in the Upper Colorado River Basin (UCRB), integrating convolutional networks (GCNs) spatial connectivity long short-term memory (LSTM) capture temporal dynamics. Using 30 years monthly data from 20 monitoring stations, STGNN predicted over 36-month horizon was evaluated against traditional models, including random forest regression (RFR), LSTM, gated recurrent units (GRU), seasonal auto-regressive integrated moving average (SARIMA). The outperformed these models across multiple metrics, achieving an R2 0.78, RMSE 0.81 mm/month, KGE 0.79 at critical locations like Lees Ferry. A sequential analysis input–output configurations identified (36, 36) setup as optimal balancing historical context forecasting accuracy. Additionally, showed strong generalizability when applied other within UCRB. These results underscore importance dependencies dynamics forecasting, offering scalable adaptable framework improve predictive accuracy support adaptive management river basins.

Язык: Английский

---

Runoff Prediction Based on Dynamic Spatiotemporal Graph Neural Network

Water, Год журнала: 2023, Номер 15(13), С. 2463 - 2463

Опубликована: Июль 5, 2023

Runoff prediction plays an important role in the construction of intelligent hydraulic engineering. Most existing deep learning runoff models use recurrent neural networks for single-step a single time series, which mainly model temporal features and ignore river convergence process within watershed. In order to improve accuracy prediction, dynamic spatiotemporal graph network (DSTGNN) is proposed considering interaction hydrological stations. The sequences are first input block extract features. captured by long short-term memory (LSTM) with self-attention mechanism. Then, upstream downstream distance matrices constructed based on topology basin, matrix sequence, spatial dependence combining above two through diffusion process. After that, residual next layer decoupling block, and, finally, results output after multi-layer stacking. Experiments conducted historical dataset Upper Delaware River Basin, MAE, MSE, MAPE, NSE were best compared baseline forecasting periods 3 h, 6 9 h. experimental show that DSTGNN can better capture characteristics has higher accuracy.

Язык: Английский

---

A Systematic Review of Deep Learning Applications in Streamflow Data Augmentation and Forecasting

EarthArXiv (California Digital Library), Год журнала: 2022, Номер unknown

Опубликована: Сен. 26, 2022

The volume and variety of Earth data have increased as a result growing attention to climate change and, subsequently, the availability large-scale sensor networks remote sensing instruments. This has been an important resource for data-driven studies generate practical knowledge services, support environmental modeling forecasting needs, transform earth science research thanks computational resources popularity novel techniques like deep learning. Timely accurate simulation extreme events are critical planning mitigation in hydrology water resources. There is strong need short-term long-term forecasts streamflow, benefiting from recent developments learning methods. In this study, we review literature that employ tackling tasks either improve quality streamflow or forecast streamflow. study aims serve starting point by covering latest approaches those topics well highlighting problems, limitations, open questions with insights future directions.

Язык: Английский

---

Real-Time Streamflow Forecasting Framework, Implementation and Post-Analysis Using Deep Learning

EarthArXiv (California Digital Library), Год журнала: 2022, Номер unknown

Опубликована: Март 13, 2022

Rainfall-runoff modeling and streamflow prediction using deep learning algorithms have been studied significantly in the last few years. The majority of these studies focus on simulation testing historical datasets. Deployment operation a real-time forecast model will face additional data computational challenges such as inaccurate rainfall assimilation with limited guiding difficulties. We proposed framework that includes pre-event training learning, acquisition, post-event analysis. implemented for 124 USGS gauged watersheds across Iowa to 120-hour rates since April 2021. This is first time models used predict operational settings at large scale, we anticipate seeing more implementations future.

Язык: Английский

---

TempNet – Temporal Super Resolution of Radar Rainfall Products with Residual CNNs

EarthArXiv (California Digital Library), Год журнала: 2022, Номер unknown

Опубликована: Сен. 22, 2022

The temporal and spatial resolution of rainfall data is crucial for environmental modeling studies in which its variability space time considered as a primary factor. Rainfall products from different remote sensing instruments (e.g., radar, satellite) have space-time resolutions because the differences their capabilities post-processing methods. In this study, we developed deep learning approach that augments with increased to complement relatively lower products. We propose neural network architecture based on Convolutional Neural Networks (CNNs) improve radar-based compare proposed model an optical flow-based interpolation method CNN-baseline model. methodology presented study could be used enhancing maps better imputation missing frames sequences 2D support hydrological flood forecasting studies.

Язык: Английский

---

Deep Temporal Neural Networks for Water Level Predictions of Watershed Systems

Опубликована: Апрель 27, 2023

Rainfall-runoff systems are complex hydrological environments that play a critical role in flood prevention. Currently, physics-based, process-driven computational models often used to forecast future flooding events. However, these physics-based computationally expensive and require intensive physical measurements of beyond remote data collection. There is growing body literature applies deep neural networks time-series for efficient, real-time predictions without the need complete virtual modeling system. deep-learning networks' robustness at forecasting far into remains an open question. In this study, we examine capabilities Long Short-Term Memory (LSTM) Temporal Convolutional Networks (TCN), state-of-the-art temporal networks, rainfall-runoff system depths. Specifically, study leverages primary, multi-modal, collected by sensors watershed Conner Creek, tributary Clinch River eastern Tennessee. These were 5-minute intervals over course 5 months. Notably, Creek consists four interconnected reservoir basins. We water level each basin independently times ranging from five minutes two hours future. Our results show both LSTM TCN can effectively model levels. when averaged across basins, has mean absolute error (MAE), with 95% confidence interval, 0.158 ± 0.049 ft 0.490 0.260 120 future, respectively. comparison, MAE 0.258 0.160 0.375 0.245 outperforms near lead time forecasting; however, retains greater relative accuracy larger periods (two hours). Nevertheless, be considered effective capturing trends systems, demonstrating them powerful tools use risk management systems.

Язык: Английский

---

Probabilistic multi-step ahead streamflow forecast based on deep learning

at - Automatisierungstechnik, Год журнала: 2024, Номер 72(6), С. 518 - 527

Опубликована: Июнь 1, 2024

Abstract The use of deep learning methods for fluvial flood forecasting is rapidly gaining traction, offering a promising solution to the challenges associated with accurate yet time-consuming numerical models. This paper presents two physics-inspired approaches specifically designed forecasting, each embracing different principles: centralized and federated learning. model utilizes an Encoder-Decoder technique handle input data varying types scales, while based on node-link graph seq2seq internal model. Both models are enhanced probabilistic head account inherent uncertainty in streamflow forecasts. objective these address limitations traditional leveraging potential improve speed, accuracy, scalability forecasting. To validate their effectiveness, were tested across cases. findings from approach emphasize importance catchment clustering before utilization demonstrate models’ ability generalize effectively catchments similar properties. On other hand, results method highlight model’s reliance test set falling within range training (Average NSE KGE sixth hour ahead 0.88 0.78, respectively). this limitation, suggests development future, such as or using Generative Adversarial Networks, generate highly extreme events, particularly context changing climate. implemented flexible operational framework open standards, ensuring adaptability usability various settings.

Язык: Английский

---

Spatio-temporal Causal Learning for Streamflow Forecasting

2021 IEEE International Conference on Big Data (Big Data), Год журнала: 2024, Номер unknown, С. 6161 - 6170

Опубликована: Дек. 15, 2024

Язык: Английский

---