Real-time streamflow forecasting in a reservoir-regulated river basin using explainable machine learning and conceptual reservoir module

An efficient parallel runoff forecasting model for capturing global and local feature information

Artificial intelligence has significantly accelerated the development of hydrological forecasting. However, research on how to efficiently identify the physical characteristics of runoff sequences and develop forecasting models that simultaneously address both global and local features of the sequences is still lacking. To address these issues, this study proposes a new PCPFN (PolyCyclic Parallel Fusion Network) prediction model that leverages the multi-periodic characteristics of runoff sequences and shares global features through a dual-architecture parallel computation approach. Unlike existing models, the PCPFN model can extract both the periodic and trend-based evolution features of runoff sequences. It constructs a multi-feature set in a "sequence-to-sequence" manner and employs a parallel structure of an Encoder and BiGRU (Bidirectional Gated Recurrent Unit) to simultaneously capture changes in both local, adjacent features and global characteristics, ensuring comprehensive attention to the sequence features. When predicting runoff data for three different hydrological conditions, the PCPFN model achieved R2 values of 0.97, 0.98, and 0.97, respectively, with other evaluation indicators significantly outperforming the benchmark models. Additionally, due to the opacity in feature distribution processes of AI models, SHAP (Shapley Additive exPlanations) analysis was used to evaluate the contribution of each feature variable to long-term runoff trends. The proposed PCPFN model, during parallel computation, not only utilizes the intrinsic features of sequences and efficiently handles the distribution of local and global features but also shares predictive information in the output module, achieving accurate runoff forecasting and providing crucial references for timely warning and forecasting.

HydroSignal: open-source internet of things information communication platform for hydrological education and outreach

This paper introduces HydroSignal, a low-cost, open-source platform designed to transform environmental monitoring and hydrological data visualization, making it accessible to educators, students, and the general public. By integrating internet of things (IoT) technology with a user-friendly web interface, HydroSignal enables real-time data acquisition and visualization of key hydrological parameters such as flood levels, turbidity, rainfall, soil moisture, and temperature. The platform's true contribution lies in its ability to empower users with hands-on, practical experience, enabling them to go beyond passive observation and actively engage with real-world hydrological data. Through four diverse use cases, HydroSignal demonstrates its flexibility in fostering data-driven decision-making, hands-on learning, and skill development across academic, educational, and community settings. By democratizing access to environmental data, HydroSignal not only simplifies the visualization of complex hydrological processes but also nurtures a deeper understanding of environmental monitoring, promoting STEM education, public engagement, and environmental awareness.

Research on a hybrid model for flood probability prediction based on time convolutional network and particle swarm optimization algorithm

Accurate flood forecasting in advance is crucial for planning and implementing watershed flood prevention measures. This study developed the PSO-TCN-Bootstrap flood forecasting model for the Tailan River Basin in Xinjiang by integrating the particle swarm optimisation (PSO) algorithm, temporal convolutional network (TCN), and Bootstrap probability sampling method. Evaluated on 50 historical flood events from 1960 to 2014 using observed rainfall-runoff data, the model showed, under the same lead time conditions, a higher Nash efficiency coefficient, along with lower root mean square and relative peak errors in flood forecasting. These results highlight the PSO-TCN-Bootstrap model's superior applicability and robustness for the Tailan River Basin. However, when the lead time exceeds 5 h, the model's relative peak error remains above 20%. Future work will focus on integrating flood generation mechanisms and enhancing machine learning models' generalisability in flood forecasting. These findings provide a scientific foundation for flood management strategies in the Tailan River Basin.

Attribution of hydrological droughts in large river-connected lakes: Insights from an explainable machine learning model

Large lakes play an important role in water resource supply, regional climate regulation, and ecosystem support, but they face threats from frequent extreme drought events, necessitating an understanding of the mechanisms behind these events. In this study, we developed an explainable machine learning (ML) model that combines the Bayesian optimized (BO) long short-term memory (LSTM) model and the integrated gradients (IG) interpretation method to simulate and explain lake water level variations. In addition, the hydrological drought trends and extreme drought events in Poyang Lake from 1960 to 2022 were identified using the standardized water level index (SWI) and run theory. The analysis revealed that the frequency of hydrological droughts in Poyang Lake increased from 1960 to 2022, especially in the autumn after 2003. By selecting the flows of the catchment and the Yangtze River as the input features, the BO-LSTM model accurately predicted the water level of Poyang Lake. The IG method was then used to interpret the prediction results from three aspects: the importance ranking of the input features, their roles in the seasonal drought trends, and their roles in extreme drought events. The results indicate that (1) the most influential factor affecting the water level of Poyang Lake was the inflow of the Ganjiang River in the catchment. (2) The increase in the lake outflow caused by the Yangtze River's draining effect was the reason for the intensification of the autumn drought in Poyang Lake. (3) The extreme hydrological drought events were primarily caused by low catchment inflows. Overall, this research provides a new approach that balances prediction accuracy with interpretability for predicting and understanding the hydrological processes in large river-connected lakes. Moreover, this method was also applied to the attribution analysis of hydrological drought in Poyang Lake, providing theoretical support for regional water resource management.

An explainable multiscale LSTM model with wavelet transform and layer-wise relevance propagation for daily streamflow forecasting

Developing an accurate and reliable daily streamflow forecasting model is important for facilitating the efficient resource planning and management of hydrological systems. In this study, an explainable multiscale long short-term memory (XM-LSTM) model is proposed for effective daily streamflow forecasting by integrating the à trous wavelet transform (ATWT) for decomposing data, the Boruta algorithm for identifying model inputs, and the layer-wise relevance propagation (LRP) for explaining the prediction results. The proposed XM-LSTM is tested by performing multi-step-ahead forecasting of daily streamflow at four stations in the middle and lower reaches of the Yangtze River basin and compared with the X-LSTM. The X-LSTM is formed by coupling the long short-term memory (LSTM) with the LRP. For comparison, the inputs of these two models are identified by the Boruta selection algorithm. The results show that all models exhibit good ability to forecast daily streamflow, however, the prediction performance decreases as the lead time increases. The XM-LSTM provides a better forecasting performance than the X-LSTM, suggesting the ability of the ATWT to improve the LSTM for daily streamflow forecasting. Moreover, the correlation scores analysis by the LRP shows that the ATWT can extract useful information that influences the daily streamflow from the raw predictors, and the water level has the most significant contribution to streamflow prediction. Accordingly, the XM-LSTM model can be viewed as a potentially useful approach for increasing the accuracy and explainability of streamflow forecasting.

Research on runoff process vectorization and integration of deep learning algorithms for flood forecasting

Accurate multi-step ahead flood forecasting is crucial for flood prevention and mitigation efforts as well as optimizing water resource management. In this study, we propose a Runoff Process Vectorization (RPV) method and integrate it with three Deep Learning (DL) models, namely Long Short-Term Memory (LSTM), Temporal Convolutional Network (TCN), and Transformer, to develop a series of RPV-DL flood forecasting models, namely RPV-LSTM, RPV-TCN, and RPV-Transformer models. The models are evaluated using observed flood runoff data from nine typical basins in the middle Yellow River region. The key findings are as follows: Under the same lead time conditions, the RPV-DL models outperform the DL models in terms of Nash-Sutcliffe efficiency coefficient, root mean square error, and relative error for peak flows in the nine typical basins of the middle Yellow River region. Based on the comprehensive evaluation results of the train and test periods, the RPV-DL model outperforms the DL model by an average of 2.82%-22.21% in terms of NSE across nine basins, with RMSE and RE reductions of 10.86-28.81% and 36.14%-51.35%, respectively. The vectorization method significantly improves the accuracy of DL flood forecasting, and the RPV-DL models exhibit better predictive performance, particularly when the lead time is 4h-6h. When the lead time is 4-6h, the percentage improvement in NSE is 9.77%, 15.07%, and 17.94%. The RPV-TCN model shows superior performance in overcoming forecast errors among the nine basins. The research findings provide scientific evidence for flood prevention and mitigation efforts in river basins.

Real-time reservoir operation using inflow and irrigation demand forecasts in a reservoir-regulated river basin

Real-time reservoir operation using inflow and irrigation demand forecasts can help reservoir system managers make effective water management decisions. Forecasting of inflow and irrigation demands is challenging, owing to the variability of the weather variables that affect inflows and irrigation demands. In this context, bias-corrected Global Forecasting System (GFS) forecasts are used here in a hybrid approach (reservoir module with Long Short Term Memory (LSTM)) to forecast the reservoir inflows. Concurrently, the bias-corrected GFS forecasts are used in irrigation demand module to forecast the irrigation demands. The 'Scaled Distribution Mapping' method is used to bias-correct the GFS data of 1-5 days lead. The study area is the Damodar river basin, India, consisting of five major reservoirs: Tenughat and Konar located upstream of Panchet, and Tilaya situated upstream of Maithon. With the upstream reservoir outflow forecasts, the inflows are forecasted in Panchet and Maithon reservoirs with NSE values of 0.88-0.96 and 0.78-0.88, respectively, up to a 5-day lead. The irrigation demand module with bias-corrected GFS forecasts forecasted the irrigation demands close to the irrigation demands with the observed weather data. The percentage errors in irrigation demand forecasts of the Kharif (June-October) season at 1-5 days lead are 9.45 %, -15.45 %, -20.52 %, -26.36 %, -27.31 %, respectively. On the contrary, percentage errors in irrigation demand forecasts of Rabi (November-February) and Boro (January-May) are in the range of 8.17-8.79 % and 3.48-8.06 %, respectively. With the inflows and irrigation demand forecasts, the Panchet and Maithon reservoirs satisfied the downstream demands and reduced the floods. The inflow and irrigation demand forecasts, based on the GFS forecasts, have substantial potential for real-time reservoir operation, leading to efficient water management downstream.

Forecasting hydrologic controls on juvenile anadromous fish out-migration with process-based modeling and machine learning

River herring (Alosa sp.) are ecologically and economically foundational species in freshwater streams, estuaries, and oceanic ecosystems. The migration between fresh and saltwater is a key life stage of river herring, where the timing and magnitude of out-migration by juveniles can be limited when streams dry and hydrologic connectivity is lost. Operational decisions by water managers (e.g., restricting community water use) can impact out-migration success; however, these decisions are largely made without reliable predictions of outmigration potential across the migration season. This research presents a model to generate short-term forecasts of the probability of herring out-migration loss. We monitored streamflow and herring out-migration for 2 years at three critical runs along Long Island Sound (CT, USA) to develop empirical understandings of the hydrologic controls on out-migration. We used calibrated Soil and Water Assessment Tool hydrologic models of each site to generate 10,000 years of daily synthetic meteorological and streamflow records. These synthetic meteorological and streamflow data were used to train random forest models to provide rapid within-season forecasts of out-migration loss from two simple predictors: current spawning reservoir depth and the previous 30-day precipitation total. The resulting models were approximately 60%-80% accurate with a 1.5-month lead time and 70-90% accurate within 2 weeks. We anticipate that this tool will support regional decisions on spawning reservoir operations and community water withdrawals. The architecture of this tool provides a framework to facilitate broader predictions of the ecological consequences of streamflow connectivity loss in human-impacted watersheds.

Streamflow classification by employing various machine learning models for peninsular Malaysia

Due to excessive streamflow (SF), Peninsular Malaysia has historically experienced floods and droughts. Forecasting streamflow to mitigate municipal and environmental damage is therefore crucial. Streamflow prediction has been extensively demonstrated in the literature to estimate the continuous values of streamflow level. Prediction of continuous values of streamflow is not necessary in several applications and at the same time it is very challenging task because of uncertainty. A streamflow category prediction is more advantageous for addressing the uncertainty in numerical point forecasting, considering that its predictions are linked to a propensity to belong to the pre-defined classes. Here, we formulate streamflow prediction as a time series classification with discrete ranges of values, each representing a class to classify streamflow into five or ten, respectively, using machine learning approaches in various rivers in Malaysia. The findings reveal that several models, specifically LSTM, outperform others in predicting the following n-time steps of streamflow because LSTM is able to learn the mapping between streamflow time series of 2 or 3 days ahead more than support vector machine (SVM) and gradient boosting (GB). LSTM produces higher F1 score in various rivers (by 5% in Johor, 2% in Kelantan and Melaka and Selangor, 4% in Perlis) in 2 days ahead scenario. Furthermore, the ensemble stacking of the SVM and GB achieves high performance in terms of F1 score and quadratic weighted kappa. Ensemble stacking gives 3% higher F1 score in Perak river compared to SVM and gradient boosting.

STL-decomposition ensemble deep learning models for daily reservoir inflow forecast for hydroelectricity production

Accurate reservoir inflow forecasting is crucial for efficient water management. In this study, different deep learning models, including Dense, Long short-term memory (LSTM), and one-dimensional convolutional neural networks (Conv1D), were used to build ensembles. Seasonal-trend decomposition using loess (STL) was applied to decompose reservoir inflows and precipitations into random, seasonal, and trend components. Seven ensemble models, namely STL-Dense, STL-Conv1D, STL-LSTM, STL-Dense-LSTM-Conv1D, STL-Dense multivariate, STL-LSTM multivariate, and STL-Conv1D multivariate, were proposed and evaluated using daily inflows and precipitation decomposed data from the Lom Pangar reservoir from 2015 to 2020. Evaluation metrics, such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and Nash Sutcliff Efficiency (NSE), were applied to assess model performance. Results showed that the STL-Dense multivariate model was the best ensemble among the thirteen models with MAE of 14.636 m³/s, RMSE of 20.841 m³/s, MAPE of 6.622%, and NSE of 0.988. These findings stress the importance of considering multiple inputs and models for accurate reservoir inflow forecasting and optimal water management. Not all ensemble models were good for Lom pangar inflow forecast as the Dense, Conv1D, and LSTM models performed better than their proposed STL monovariate ensemble models.