Figure 1 - uploaded by Carolina Meruane
Content may be subject to copyright.
Source publication
The intensification of the hydrological cycle because of global warming raises concerns about future floods and their impact on large cities where exposure to these events has also increased. The development of adequate adaptation solutions such as early warning systems is crucial. Here, we used deep learning (DL) for weather-runoff forecasting in...
Contexts in source publication
Context 1
... illustrated in Figure 1, an ANN, also known as a multilayer perceptron, consists of an arrangement of input neurons known as the input layer, an arrangement of output neurons known as the output layer and a number of hidden layers. Each neuron receives a weighted sum from the neurons in the previous layer and gives an input to every neuron of the next layer. ...
Context 2
... evaluate the differences between the ANN and DL weather-runoff models, Figure 10 shows the direct comparison between observed and predicted maximum and average flow for the DL model ((a) and (b)) and the ANN model ((c) and (d)); using the test subset (the goodness of the fit is indicated in Table 5). The same figure for the rest of flow stations showed, in general, good agreement between predicted and observed maximum and average flows for both the DL and the ANN models, however, the performance of the DL weather-runoff model was better than the ANN model (see Figures S1-S8 in Supporting Information and Table 5). ...
Context 3
... evaluate the differences between the ANN and DL weather-runoff models, Figure 10 shows the direct comparison between observed and predicted maximum and average flow for the DL model ((a) and (b)) and the ANN model ((c) and (d)); using the test subset (the goodness of the fit is indicated in Table 5). The same figure for the rest of flow stations showed, in general, good agreement between predicted and observed maximum and average flows for both the DL and the ANN models, however, the performance of the DL weather-runoff model was better than the ANN model (see Figures S1-S8 in Supporting Information and Table 5). The normalized RMSE for the DL model were 5.9% and 4.3% for í µí± and í µí± , respectively; NSE and í µí° ¶ were very close to 1, and the RMSE was 7.0 m 3 /s and 4.3 m 3 /s for í µí± and Q , respectively (see Table 5). ...
Context 4
... normalized RMSE for the DL model were 5.9% and 4.3% for í µí± and í µí± , respectively; NSE and í µí° ¶ were very close to 1, and the RMSE was 7.0 m 3 /s and 4.3 m 3 /s for í µí± and Q , respectively (see Table 5). With respect to performance in predicting the entire time-series of the flow for the following 3 days, Figure 11 shows the comparison between the observed and predicted flow for different cases identified with open circles in Figure 10. These examples were chosen based on the cumulative frequency of the average observed flow, using percentiles of 99.9%, 99.6%, 99%, 98%, 97%, 96%, 95%, 94%, 93%, 92%, 91% and 90%, for panels (a) to (l), respectively. ...
Context 5
... normalized RMSE for the DL model were 5.9% and 4.3% for í µí± and í µí± , respectively; NSE and í µí° ¶ were very close to 1, and the RMSE was 7.0 m 3 /s and 4.3 m 3 /s for í µí± and Q , respectively (see Table 5). With respect to performance in predicting the entire time-series of the flow for the following 3 days, Figure 11 shows the comparison between the observed and predicted flow for different cases identified with open circles in Figure 10. These examples were chosen based on the cumulative frequency of the average observed flow, using percentiles of 99.9%, 99.6%, 99%, 98%, 97%, 96%, 95%, 94%, 93%, 92%, 91% and 90%, for panels (a) to (l), respectively. ...
Context 6
... examples were chosen based on the cumulative frequency of the average observed flow, using percentiles of 99.9%, 99.6%, 99%, 98%, 97%, 96%, 95%, 94%, 93%, 92%, 91% and 90%, for panels (a) to (l), respectively. Finally, Figure 11 compares predicted and observed flow time-series for the maximum flow event. Similar figures for the rest of the flow stations are found in the Supporting Information. ...
Context 7
... evaluate the differences between the ANN and DL weather-runoff models, Figure 10 shows the direct comparison between observed and predicted maximum and average flow for the DL model ((a) and (b)) and the ANN model ((c) and (d)); using the test subset (the goodness of the fit is indicated in Table 5). The same figure for the rest of flow stations showed, in general, good agreement between predicted and observed maximum and average flows for both the DL and the ANN models, however, the performance of the DL weather-runoff model was better than the ANN model (see Figures S1-S8 in Supporting Information and Table 5). ...
Context 8
... evaluate the differences between the ANN and DL weather-runoff models, Figure 10 shows the direct comparison between observed and predicted maximum and average flow for the DL model ((a) and (b)) and the ANN model ((c) and (d)); using the test subset (the goodness of the fit is indicated in Table 5). The same figure for the rest of flow stations showed, in general, good agreement between predicted and observed maximum and average flows for both the DL and the ANN models, however, the performance of the DL weather-runoff model was better than the ANN model (see Figures S1-S8 in Supporting Information and Table 5). The normalized RMSE for the DL model were 5.9% and 4.3% for Q avg and Q max , respectively; NSE and C xy were very close to 1, and the RMSE was 7.0 m 3 /s and 4.3 m 3 /s for Q max and Q avg , respectively (see Table 5). ...
Context 9
... normalized RMSE for the DL model were 5.9% and 4.3% for Q avg and Q max , respectively; NSE and C xy were very close to 1, and the RMSE was 7.0 m 3 /s and 4.3 m 3 /s for Q max and Q avg , respectively (see Table 5). With respect to performance in predicting the entire time-series of the flow for the following 3 days, Figure 11 shows the comparison between the observed and predicted flow for different cases identified with open circles in Figure 10. These examples were chosen based on the cumulative frequency of the average observed flow, using percentiles of 99.9%, 99.6%, 99%, 98%, 97%, 96%, 95%, 94%, 93%, 92%, 91% and 90%, for panels (a) to (l), respectively. ...
Context 10
... normalized RMSE for the DL model were 5.9% and 4.3% for Q avg and Q max , respectively; NSE and C xy were very close to 1, and the RMSE was 7.0 m 3 /s and 4.3 m 3 /s for Q max and Q avg , respectively (see Table 5). With respect to performance in predicting the entire time-series of the flow for the following 3 days, Figure 11 shows the comparison between the observed and predicted flow for different cases identified with open circles in Figure 10. These examples were chosen based on the cumulative frequency of the average observed flow, using percentiles of 99.9%, 99.6%, 99%, 98%, 97%, 96%, 95%, 94%, 93%, 92%, 91% and 90%, for panels (a) to (l), respectively. ...
Context 11
... examples were chosen based on the cumulative frequency of the average observed flow, using percentiles of 99.9%, 99.6%, 99%, 98%, 97%, 96%, 95%, 94%, 93%, 92%, 91% and 90%, for panels (a) to (l), respectively. Finally, Figure 11 compares predicted and observed flow time-series for the maximum flow event. Similar figures for the rest of the flow stations are found in the Supporting Information. ...
Context 12
... DL model, which incorporates LSTM cells, has a much better prediction performance for the time-flow series as well as for the maximum and average flow, demonstrating that the temporal capacity of LSTM-based algorithms allows a prediction of temporal changes. Figure 11, associated to percentiles of 99.9%, 99.6%, 99%, 98%, 97%, 96%, 95%, 94%, 93%, 92%, 91% and 90%, for panels (a-l), respectively. (m) Plots predicted and observed time-series for the event with maximum flow. ...
Context 13
... order to verify the early warning advantage of the DL weather-runoff model, two extreme events were analyzed in detail. These events correspond to floods that occurred in April of 2016 (with a peak flow of 1078.6 m 3 /s, Figure 12a), and in May of 2012 (with a maximum flow of 546.1 m 3 /s, Figure 12b). For each event, the DL weather-runoff forecast model was run several times, starting at different days before the time at which the maximum flow was observed. ...
Context 14
... order to verify the early warning advantage of the DL weather-runoff model, two extreme events were analyzed in detail. These events correspond to floods that occurred in April of 2016 (with a peak flow of 1078.6 m 3 /s, Figure 12a), and in May of 2012 (with a maximum flow of 546.1 m 3 /s, Figure 12b). For each event, the DL weather-runoff forecast model was run several times, starting at different days before the time at which the maximum flow was observed. ...
Context 15
... each event, the DL weather-runoff forecast model was run several times, starting at different days before the time at which the maximum flow was observed. As an example, Figure 12 shows three of these runs: One that starts 6 days before the peak flow and ends before the flow start to rise (black line that starts with the circle and ends with the black x); the second (blue simulation) that starts 4 days the peak flow, at the beginning of the storm, and ends before the peak flow was observed; and the third simulation starting on 2 days before the peak flow, in the middle of the storm, and ending after the peak flow has passed. Similarly, Table 6 shows the errors of time to peak ETp (Equation 13), and peak discharge, EQp (Equation 12), calculated for each one of the different simulations. ...
Context 16
... similar situation is observed for the May 2012 event. In terms of the errors EQp and ETp (Table 6), the relative error in predicting the maximum flow tends to be smaller for larger maximum flows, and it takes positive values; so that, in this case, the model predicts maximum flows Figure 11, associated to percentiles of 99.9%, 99.6%, 99%, 98%, 97%, 96%, 95%, 94%, 93%, 92%, 91% and 90%, for panels (a-l), respectively. (m) Plots predicted and observed time-series for the event with maximum flow. ...
Context 17
... order to verify the early warning advantage of the DL weather-runoff model, two extreme events were analyzed in detail. These events correspond to floods that occurred in April of 2016 (with a peak flow of 1078.6 m 3 /s, Figure 12a), and in May of 2012 (with a maximum flow of 546.1 m 3 /s, Figure 12b). For each event, the DL weather-runoff forecast model was run several times, starting at different days before the time at which the maximum flow was observed. ...
Context 18
... order to verify the early warning advantage of the DL weather-runoff model, two extreme events were analyzed in detail. These events correspond to floods that occurred in April of 2016 (with a peak flow of 1078.6 m 3 /s, Figure 12a), and in May of 2012 (with a maximum flow of 546.1 m 3 /s, Figure 12b). For each event, the DL weather-runoff forecast model was run several times, starting at different days before the time at which the maximum flow was observed. ...
Context 19
... each event, the DL weather-runoff forecast model was run several times, starting at different days before the time at which the maximum flow was observed. As an example, Figure 12 shows three of these runs: One that starts 6 days before the peak flow and ends before the flow start to rise (black line that starts with the circle and ends with the black x); the second (blue simulation) that starts 4 days the peak flow, at the beginning of the storm, and ends before the peak flow was observed; and the third simulation starting on 2 days before the peak flow, in the middle of the storm, and ending after the peak flow has passed. Similarly, Table 6 shows the errors of time to peak ET p (Equation (13)), and peak discharge, EQ p (Equation (12)), calculated for each one of the different simulations. ...
Context 20
... data-driven weather-runoff models were designed based on the following three central ideas: (i) The near future flow (3 days) in the studied flow stations responds to both the precipitation rate of the storm, but also to changes in the watershed area or rate of snow melt (see Figure 11f). Consequently, a rainfall-runoff scheme (e.g., [25]) is not enough for predicting near-future flow, which justifies the weather-runoff concept that also uses air temperature and relatively humidity and the 0 °C isotherm for predicting the near-future flow. ...
Context 21
... data-driven weather-runoff models were designed based on the following three central ideas: (i) The near future flow (3 days) in the studied flow stations responds to both the precipitation rate of the storm, but also to changes in the watershed area or rate of snow melt (see Figure 11f). Consequently, a rainfall-runoff scheme (e.g., [25]) is not enough for predicting near-future flow, which justifies the weather-runoff concept that also uses air temperature and relatively humidity and the 0 • C isotherm for predicting the near-future flow. ...
Context 22
... ANN do not have a temporal memory, they have difficulties in recognising temporal changes. This is reflected in greater error when predicting flow floods and therefore tend to underestimate the flow, as shown in Figure 11. In this context, one of the most successful techniques based on Recurrent Neural Networks (RNN) is the DL approach based on LSTM cells [24][25][26]. ...
Context 23
... another important feature of the proposed architecture of the DL weather-runoff forecast (Figure 5b) is that it is capable of predicting an output time-series with a finer temporal resolution (1 h) than for the input time-series (6 h), thus enabling the use of DL as a temporal downscaling technique. This allows to precisely locating the time at which the peak flow will occur, which gives the system an early warning advantage, as shown in Figure 12 and Table 6. The DL weather-runoff model is capable of capturing the peak flow, the time at which it will occur and the flow duration. ...
Similar publications
The evolution of several sub‐alpine alluvial fans SE of the Jostedalsbreen ice cap was investigated based on their geomorphology and Schmidt‐hammer exposure‐age dating (SHD) applied to 47 boulder deposits on the fan surfaces. A debris‐flood rather than debrisflow or water‐flow origin for the deposits was inferred from their morphology, consisting o...
Citations
... These include a range of machine learning techniques, such as support vector regression (SVR) (Bafitlhile and Li 2019;De Gregorio et al. 2018;Kisi 2015;Li et al. 2016) and other traditional machine learning techniques Zucco et al. 2015). Additionally, deep learning techniques, such as Recurrent Neural Networks (RNNs) (Chang et al. 2014;Chen et al. 2013), Long-Short Term Memory networks (LSTMs) (Cheng et al. 2015;de la Fuente et al. 2019;Ding et al. 2020;Gude et al. 2020;Hu et al. 2019;Liu et al. 2022;Moishin et al. 2021;Ruma et al. 2023;Won et al. 2022), and Convolutional Neural Networks (CNNs) (Shu et al. 2021;Wang et al. 2021). Compared with machine learning techniques, deep learning techniques often demonstrate superior performance and the capability to interpret complex patterns within hydrological time series data (Farahmand et al. 2023;Salman et al. 2015;Sankaranarayanan et al. 2020), which makes them excellent for the difficulties associated with flood prediction. ...
The rapid and devastating nature of flood events in small and medium basins presents considerable challenges to flood forecasting. Developing a robust and accurate flood forecasting method is crucial to mitigating flood effects. To this end, we propose the Time-distributed CNN-LSTM model (TD-CNN-LSTM), a hybrid deep learning framework based on a Time Distribution layer (TD), Convolutional Neural Networks (CNNs), and Long-Short Term Memory Networks (LSTMs). The TD-CNN-LSTM model efficiently captures spatiotemporal hydrological features within spatial dimensions using the Time-distributed local Features extractor, while the LSTM focuses on extracting spatiotemporal features in temporal dimensions. Our model effectively captures complex spatial and temporal relationship patterns within hydrological data from flood time series. Experimental results on the Tunxi and Changhua basins show the superior predictive capabilities of our model, particularly in forecasting flood occurrence time and peak. When compared with baseline models LSTM, CNN, ConvLSTM, STA-LSTM, and CNN-LSTM at the moment T + 9, TD-CNN-LSTM achieved a 6.7%, 10.14%, 8.5%, 6.3%, and 6.6% decreased in Root-Mean-Square Error (RMSE), respectively, a 6.5%, 9.8%, 10.9%, 6.8%, and 10.6% decreased in Mean Absolute Error (MAE), respectively, and a 7.4%, 31.6%, 23.5%, 18.8%, and 27.8% decreased in Mean Absolute Percentage Error (MAPE), respectively. In addition, the determination Coefficient (R²) increased by 3.6% and Nash-Sutcliffe Efficiency (NSE) increased by 4.8% .
... The time-series data must be reframed into a supervised learning format to anticipate river flow using meteorological data and the sliding window approach (Meng et al. 2022;Shin and Yi 2019). The sliding window approach has been integrated with a deep-learning runoff model-based meteorological forecast for a hydrological early warning system (Fuente et al. 2019). It has also been utilized in flood forecasting, which is employed to power a river routing model that generates river flow predictions several months in advance (Emerton et al. 2019). ...
The Yeşilırmak River Basin in northern Türkiye is crucial for the region’s water supply, agriculture, hydroelectric power generation, and clean drinking water. The primary goal of this study is to determine which modeling approach is most appropriate for various locations within the basin and how well meteorological data can predict river flow rates. Hydrological and meteorological forecasting both depend on the prediction of river flow rates. An artificial neural network (ANN), Univariate and Multivariate Long Short-Term Memory (LSTM) models have been utilized for streamflow forecasting. This research aims to determine the best model for several provinces in the basin area and give decision-makers a tool for reliable river flow rate estimates by combining LSTM and ANN models. According to research findings, the supervised multivariate LSTM model performed better than the unsupervised model in accuracy and precision. The sliding window methodology is suitable for estimating river flow based on meteorological datasets because it offers a primary method for reinterpreting time-series data in a supervised learning style. Compared to LSTM models, the ANN model that has been statistically optimized through experiments (DoE) design performs better in forecasting the river flow rate in the Yeşilırmak River basin (R² = 0.98, RMSE = 0.18). The study’s findings provided prospective cognitive models for the strategic management of water resources by forecasting future data from flow monitoring stations.
... This catastrophe devastated approximately 25% of the city due to the massive currents of water and mud, resulting in the loss of over 4000 lives and leaving around 10,000 individuals unaccounted for [8]. Human beings have proven through years of struggling with floods that combining engineering and non-engineering measures is an effective measure for solving the problem of floods [9][10][11]. In addition, among the usual non-engineering measures, continuous efforts are made to enhance the precision of hydrological forecasting, a strategy that has proven to be effective. ...
Artificial intelligence has undergone rapid development in the last thirty years and has been widely used in the fields of materials, new energy, medicine, and engineering. Similarly, a growing area of research is the use of deep learning (DL) methods in connection with hydrological time series to better comprehend and expose the changing rules in these time series. Consequently, we provide a review of the latest advancements in employing DL techniques for hydrological forecasting. First, we examine the application of convolutional neural networks (CNNs) and recurrent neural networks (RNNs) in hydrological forecasting, along with a comparison between them. Second, a comparison is made between the basic and enhanced long short-term memory (LSTM) methods for hydrological forecasting, analyzing their improvements, prediction accuracies, and computational costs. Third, the performance of GRUs, along with other models including generative adversarial networks (GANs), residual networks (ResNets), and graph neural networks (GNNs), is estimated for hydrological forecasting. Finally, this paper discusses the benefits and challenges associated with hydrological forecasting using DL techniques, including CNN, RNN, LSTM, GAN, ResNet, and GNN models. Additionally, it outlines the key issues that need to be addressed in the future.
... In recent decades, ML models have become popular among hydrologists due to their advantages, including shorter simulation times, the possibility of computationally inexpensive real-time operation, and less overfitting compared to models based on physical processes (Solomatine and Ostfeld, 2008;Muñoz et al., 2018;Kwon et al., 2020;Adnan et al., 2021;Huang and Lee, 2021;Moreido et al., 2021). The Random Forest (RF) algorithm and the Long Short-Term Memory (LSTM) networks are among the most commonly used ML techniques for hydrological time-series forecasting (Muñoz et al., 2018(Muñoz et al., , 2023de la Fuente et al., 2019;Campozano et al., 2020;Li et al., 2022;Zhou et al., 2023). Recent studies have shown the superiority of LSTM over RF models for sub-daily runoff forecasting (Zhou et al., 2023). ...
... The main LSTM hyperparameters are the number of layers, number of units, learning rate, dropout, epochs, batch size, and activation function. We initially set the transfer function of the output layer to linear as suggested by de la Fuente et al. (2019). The architecture and the search space used for the hyperparameterization task are shown in Table 1. ...
Introduction
In complex mountain basins, hydrological forecasting poses a formidable challenge due to the intricacies of runoff generation processes and the limitations of available data. This study explores the enhancement of short-term runoff forecasting models through the utilization of long short-term memory (LSTM) networks.
Methods
To achieve this, we employed feature engineering (FE) strategies, focusing on geographic data and the Soil Conservation Service Curve Number (SCS-CN) method. Our investigation was conducted in a 3,390 km ² basin, employing the GSMaP-NRT satellite precipitation product (SPP) to develop forecasting models with lead times of 1, 6, and 11 h. These lead times were selected to address the needs of near-real-time forecasting, flash flood prediction, and basin concentration time assessment, respectively.
Results and discussion
Our findings demonstrate an improvement in the efficiency of LSTM forecasting models across all lead times, as indicated by Nash-Sutcliffe efficiency values of 0.93 (1 h), 0.77 (6 h), and 0.67 (11 h). Notably, these results are on par with studies relying on ground-based precipitation data. This methodology not only showcases the potential for advanced data-driven runoff models but also underscores the importance of incorporating available geographic information into precipitation-ungauged hydrological systems. The insights derived from this study offer valuable tools for hydrologists and researchers seeking to enhance the accuracy of hydrological forecasting in complex mountain basins.
... On the other hand, NN models adequately address the flood risk by locating the peak flow rates with sufficient temporal resolution (de la Fuente et al., 2019). Timely information gathering and real-time simulation can be expressed through the perfect flood forecasting models. ...
... In the same way, the modeling approach should be as simple as possible according to the parsimony principle (Serinaldi & Kilsby, 2015). In this context, hybrid dynamical-statistical approaches are particularly useful for hydroclimatic extremes (Slater et al., 2021), and hybrid dynamical-data driven approaches have shown better results and should be used (de la Fuente et al., 2019). The advantages of these new approaches are that they allow the development of models with low error, as well as the modeling of ensembles of future projections due to low calculation costs. ...
... large amount of data to generate several hydroclimatic projections without significant computational requirements or a deep understanding of the physical processes that explain the catchment response. The CMIP6 climate models for the worst-case scenario SSP5-RCP8.5 were coupled to these hydrological deep learning models that were based on LSTM layers, which have the advantage of capturing temporal correlations among the model input variables (de la Fuente et al., 2019;Hu et al., 2018;Ma et al., 2021;Shen, 2018). The risk is, however, overfitting models that lack the flexibility to predict the catchment response under conditions that are observed in the training dataset. ...
Extreme hydrological events have been recorded around the world, confirming the changing hydroclimate and therefore threatening the reliability over a planning horizon of the existing and projected hydraulic infrastructure. Unfortunately, there is a gap between climate science and engineering planning and design, which is partially filled in this study, whose aim is to evaluate the expected reliability over a planning horizon of a hydraulic infrastructure and its uncertainty for hydroclimatic projections. To achieve this goal, we implemented hydrological deep learning models that are driven by the Coupled Model Intercomparison Project Phase 6 (CMIP6). After validating and testing the hydrological deep learning models, sample series of the annual maximum flood (AMF) were generated for different mountain rivers of Chile and used for a standard frequency analysis that updates the exceedance probability of the flood events used for the hydraulic design. We found that the mean and the standard deviation of the sample series of the AMF are the key parameters that enable quantifying temporal changes in the exceedance probability of the design flow. The cumulative effects of these changes are summarized by the projected reliability over the planning horizon of the infrastructure, and the age of the infrastructure at which this projected reliability equates to the design reliability, quantifies the impact of climate change on the expected security of the hydraulic project. We show that the infrastructure that was designed for the smaller exceedance probability is more vulnerable to climate change, where the updated planning horizon can be as small as 45% of the planning horizon used for the design.
... These results contradict those of past studies (e.g. de la Fuente et al. 2019, Hu et al. 2020, Wegayehu and Muluneh 2022 wherein the superiority of the LSTM over the traditional ANN is advocated for lead time flood forecasting. As the LSTM networks possess better inherent memory integrated with the feedback structure (Karimi et al. 2019), the reason behind the comparatively lower efficiency of the LSTM model tested herein could be the insufficient size of input data. ...
A novel smoothing-based long short-term memory (Smooth-LSTM) framework for up to 5 days ahead flood forecasting is proposed, and compared with the benchmark LSTM (LSTM) model, an ANN model, and the conceptual MIKE11 NAM-HD (MIKE) hydrological model. This framework was tested in the typical middle Mahanadi River basin (India) having a tropical monsoon-type climatology. Variation of training loss indicated that the LSTM network has a higher learning ability at smaller network and batch sizes. The Smooth-LSTM model could predict the streamflow with higher Nash-Sutcliffe efficiency of 0.82–0.87 up to 5 days lead-time with a better reproduction of the observed crucial high peak floods; whereas the corresponding MIKE, ANN and LSTM model-based forecasts were acceptable only up to 4, 3 and 1-day lead-times, respectively. Overall, the Smooth-LSTM model is found to be robust in operational flood forecasting characterized with lower uncertainty and the least sensitivity to any redundant input information.
... The GMDH, by finding and sorting different solutions and also due to its self-organizing approach, can find the relationship between each meteorological variable and runoff process, and was reported as an effective tool for rainfall-runoff simulation. In addition, the current study proposed a coupled model by using hydrological physically based models and DL models, and this finding is concurrent with several previous studies in which machine learning/DL models are applied to support hydrological models [7,8,35,[60][61][62]. ...
Accurate streamflow simulation is crucial for many applications, such as optimal reservoir operation and irrigation. Conceptual techniques employ physical ideas and are suitable for representing the physics of the hydrologic model, but they might fail in competition with their more advanced counterparts. In contrast, deep learning (DL) approaches provide a great computational capability for streamflow simulation, but they rely on data characteristics and the physics of the issue cannot be fully understood. To overcome these limitations, the current study provided a novel framework based on a combination of conceptual and DL techniques for enhancing the accuracy of streamflow simulation in a snow-covered basin. In this regard, the current study simulated daily streamflow in the Kalixälven river basin in northern Sweden by integrating a snow-based conceptual hydrological model (MISD) with a DL model. Daily precipitation, air temperature (average, minimum, and maximum), dew point temperature, evapotranspiration, relative humidity, sunshine duration, global solar radiation, and atmospheric pressure data were used as inputs for the DL model to examine the effect of each meteorological variable on the streamflow simulation. Results proved that adding meteorological variables to the conceptual hydrological model underframe of parallel settings can improve the accuracy of streamflow simulating by the DL model. The MISD model simulated streamflow had an MAE = 8.33 (cms), r = 0.88, and NSE = 0.77 for the validation phase. The proposed deep-conceptual learning-based framework also performed better than the standalone MISD model; the DL method had an MAE = 7.89 (cms), r = 0.90, and NSE = 0.80 for the validation phase when meteorological variables and MISD results were combined as inputs for the DL model. The integrated rainfall-runoff model proposed in this research is a new concept in rainfall runoff modeling which can be used for accurate streamflow simulations.
... However, they have shown high accuracy in correspondence of 1D time series data, such as those used for speech recognition [17] and text classification [18]. Ragab et al. [19] showed that the accuracy of CNNs increased faster than other neural prediction approaches, such as LSTM, due to their low computational requirements in real-time and low-cost implementations; however, other recent studies (e.g., [20][21][22]) concluded that LSTM represented the most advisable choice. ...
Accurate flow forecasting may support responsible institutions in managing river systems and limiting damages due to high water levels. Machine-learning models are known to describe many nonlinear hydrological phenomena, but up to now, they have mainly provided a single future value with a fixed information structure. This study trains and tests multi-step deep neural networks with different inputs to forecast the water stage of two sub-alpine urbanized catchments. They prove effective for one hour ahead flood stage values and occurrences. Convolutional neural networks (CNNs) perform better when only past information on the water stage is used. Long short-term memory nets (LSTMs) are more suited to exploit the data coming from the rain gauges. Predicting a set of water stages over the following hour rather than just a single future value may help concerned agencies take the most urgent actions. The paper also shows that the architecture developed for one catchment can be adapted to similar ones maintaining high accuracy.
... Due to the complex structure of the hydrological model and the complex evolution characteristics and spatiotemporal evolution trend of a basin water cycle system, in order to accurately describe the hydrological cycle process of the basin, most hydrological model parameters are difficult to determine. How to determine the runoff parameters that can adapt to the heterogeneity of the underlying surface and the size of runoff level of different basins is the key to obtain high-precision short-term runoff forecast information of the basin [23,24]. Including basin spatial data, basin meteorological data, runoff simulation data, and other information, through data collection, data processing, parameter estimation, and optimization, hydrological runoff relationship modeling and analysis are realized [25]. ...
In actual engineering fields, the bearing capacity of a rock is closely related to the pore water pressure in the rock. Studies have shown that the pore water in the rock has a great relationship with the change in runoff. Thus, it has crucial meaning to accurately evaluate and quantitate the property of the rainfall–runoff, and many traditional classic models are proposed to study the characteristic of rainfall–runoff. While considering the high uncertainty and randomness of the rainfall–runoff property, more and more artificial neural networks (ANN) are used for the rainfall–runoff modeling as well as other fields. Among them, the long short-term memory (LSTM), which can be trained for sequence generation by processing real data sequences one step at a time and has good prediction results in other engineering fields, is adopted in this study to investigate the changes of rainfall–runoff values and make a prediction. In order to ensure the accuracy of the trained model, the cross-validation method is used in this study. The training data set is divided into 12 parts. The monthly forecast results from 2014 to 2015 show that the model can well reflect the peaks and troughs. In a recent study, the relationship between the rainfall–runoff and discharge are commonly based on the current measured data, while the prediction results are adopted to analyze the relation of these parameters, and considering that the existing methods have fuzzy relationship between runoff and discharge, which leads to a high risk of forecasting and dispatching. A method of modeling analysis and parameter estimation of hydrological runoff and discharge relationship based on machine learning is designed. From the experimental results, the average risk of this method is 61.23%, which is 15.104% and 13.397% less than that of the other two existing methods, respectively. It proves that the method of hydrological runoff relationship modeling and parameter estimation integrated with machine learning has better practical application effect.
