Rainfall‐runoff modeling is a complex nonlinear time series problem. While there is still room for improvement, researchers have been developing physical and machine learning models for decades to ...predict runoff using rainfall data sets. With the advancement of computational hardware resources and algorithms, deep learning methods such as the long short‐term memory (LSTM) model and sequence‐to‐sequence (seq2seq) modeling have shown a good deal of promise in dealing with time series problems by considering long‐term dependencies and multiple outputs. This study presents an application of a prediction model based on LSTM and the seq2seq structure to estimate hourly rainfall‐runoff. Focusing on two Midwestern watersheds, namely, Clear Creek and Upper Wapsipinicon River in Iowa, these models were used to predict hourly runoff for a 24‐hr period using rainfall observation, rainfall forecast, runoff observation, and empirical monthly evapotranspiration data from all stations in these two watersheds. The models were evaluated using the Nash‐Sutcliffe efficiency coefficient, the correlation coefficient, statistical bias, and the normalized root‐mean‐square error. The results show that the LSTM‐seq2seq model outperforms linear regression, Lasso regression, Ridge regression, support vector regression, Gaussian processes regression, and LSTM in all stations from these two watersheds. The LSTM‐seq2seq model shows sufficient predictive power and could be used to improve forecast accuracy in short‐term flood forecast applications. In addition, the seq2seq method was demonstrated to be an effective method for time series predictions in hydrology.
Key Points
An hourly runoff model was developed using the LSTM sequence‐to‐sequence learning method for 24‐hr predictions on USGS stations
The proposed model shows better performance than traditional data‐driven models and is applicable to different watersheds
The advantages and limitations of seq2seq models and how this model structure could work on the rainfall‐runoff modeling is presented
•A new seq2seq rainfall-runoff model named LSTM-MSV-S2S is proposed.•LSTM-MSV-S2S has promising performance on the CAMELS data set.•LSTM-MSV-S2S is more appropriate for multi-day-ahead runoff ...predictions.
Rainfall-runoff modeling is a challenging and important nonlinear time series problem in hydrological sciences. Recently, among the data-driven rainfall-runoff models, those ones based on the long short-term memory (LSTM) network show good performance. Furthermore, LSTM-based sequence-to-sequence (LSTM-S2S) models achieve promising performance for multi-step-ahead runoff predictions. In this paper, for multi-day-ahead runoff predictions, we propose a novel data-driven model named LSTM-based multi-state-vector sequence-to-sequence (LSTM-MSV-S2S) rainfall-runoff model, which contains m multiple state vectors for m-step-ahead runoff predictions. It differs from the existing LSTM-S2S rainfall-runoff models using only one state vector and is more appropriate for multi-day-ahead runoff predictions. To show its performance and advantages, we compare it with two LSTM-S2S models by testing them on 673 basins of the Catchment Attributes and Meteorology for Large-Sample Studies (CAMELS) data set. The results show that our LSTM-MSV-S2S model has better performance in general and thus using multiple state vectors is more appropriate for multi-day-ahead runoff predictions.
► BMA and GLUE is used to define uncertainties in climate change impact studies. ► The uncertainty envelop derived from BMA are wider than that estimated from GLUE. ► Hydrological model uncertainty ...has a significant role in impact studies. ► Future stream flow for Irish basin shows progressive increases of winter flow.
This study attempts to assess the uncertainty in the hydrological impacts of climate change using a multi-model approach combining multiple emission scenarios, GCMs and conceptual rainfall–runoff models to quantify uncertainty in future impacts at the catchment scale. The uncertainties associated with hydrological models have traditionally been given less attention in impact assessments until relatively recently. In order to examine the role of hydrological model uncertainty (parameter and structural uncertainty) in climate change impact studies a multi-model approach based on the Generalised Likelihood Uncertainty Estimation (GLUE) and Bayesian Model Averaging (BMA) methods is presented. Six sets of regionalised climate scenarios derived from three GCMs, two emission scenarios, and four conceptual hydrological models were used within the GLUE framework to define the uncertainty envelop for future estimates of stream flow, while the GLUE output is also post processed using BMA, where the probability density function from each model at any given time is modelled by a gamma distribution with heteroscedastic variance. The investigation on four Irish catchments shows that the role of hydrological model uncertainty is remarkably high and should therefore be routinely considered in impact studies. Although, the GLUE and BMA approaches used here differ fundamentally in their underlying philosophy and representation of error, both methods show comparable performance in terms of ensemble spread and predictive coverage. Moreover, the median prediction for future stream flow shows progressive increases of winter discharge and progressive decreases in summer discharge over the coming century.
•We answer ten basic questions on use of global sensitivity analysis on RR models.•SA can be performed easily using the free Hydromad and Sensitivity packages in R.•SA can help improve ...identifiability of a model by justifying fixing of parameters.•With daily data a minimum duration for obtaining reliable SA results is five years.•Simpler models have better-identified parameters but varying sensitivity.
Sensitivity analysis (SA) is generally recognized as a worthwhile step to diagnose and remedy difficulties in identifying model parameters, and indeed in discriminating between model structures. An analysis of papers in three journals indicates that SA is a standard omission in hydrological modeling exercises. We provide some answers to ten reasonably generic questions using the Morris and Sobol SA methods, including to what extent sensitivities are dependent on parameter ranges selected, length of data period, catchment response type, model structures assumed and climatic forcing. Results presented demonstrate the sensitivity of four target functions to parameter variations of four rainfall–runoff models of varying complexity (4–13 parameters). Daily rainfall, streamflow and pan evaporation data are used from four 10-year data sets and from five catchments in the Australian Capital Territory (ACT) region. Similar results are obtained using the Morris and Sobol methods. It is shown how modelers can easily identify parameters that are insensitive, and how they might improve identifiability. Using a more complex objective function, however, may not result in all parameters becoming sensitive. Crucially, the results of the SA can be influenced by the parameter ranges selected. The length of data period required to characterize the sensitivities assuredly is a minimum of five years. The results confirm that only the simpler models have well-identified parameters, but parameter sensitivities vary between catchments. Answering these ten questions in other case studies is relatively easy using freely available software with the Hydromad and Sensitivity packages in R.
To assess the influence of storage dynamics and nonlinearities in hydrological connectivity on time‐variant stream water ages, we used a new long‐term record of daily isotope measurements in ...precipitation and streamflow to calibrate and test a parsimonious tracer‐aided runoff model. This can track tracers and the ages of water fluxes through and between conceptual stores in steeper hillslopes, dynamically saturated riparian peatlands, and deeper groundwater; these represent the main landscape units involved in runoff generation. Storage volumes are largest in groundwater and on the hillslopes, though most dynamic mixing occurs in the smaller stores in riparian peat. Both streamflow and isotope variations are generally well captured by the model, and the simulated storage and tracer dynamics in the main landscape units are consistent with independent measurements. The model predicts that the average age of stream water is ∼1.8 years. On a daily basis, this varies between ∼1 month in storm events, when younger waters draining the hillslope and riparian peatland dominates, to around 4 years in dry periods when groundwater sustains flow. This variability reflects the integration of differently aged water fluxes from the main landscape units and their mixing in riparian wetlands. The connectivity between these spatial units varies in a nonlinear way with storage that depends upon precipitation characteristics and antecedent conditions. This, in turn, determines the spatial distribution of flow paths and the integration of their contrasting nonstationary ages. This approach is well suited for constraining process‐based modeling in a range of northern temperate and boreal environments.
Key Points:
Stream water age is estimated by modeling high‐resolution isotope time series
Storage dynamics drive nonlinear spatial patterns of hydrologic connectivity
Connectivity variations determine nonstationary stream water ages
•Synthetic samples are added to LSTM by previously undiscussed physical mechanisms.•Using extreme events to improve flood peaks and avoid negative streamflow.•Proposed PHY-LSTM outperforms ...conventional one both in local and regional models.•Physics-based monotonic relationships are upheld in the PHY-LSTM.
Deep learning methods have recently shown a broad application prospect in rainfall-runoff modeling. However, the lack of physical mechanism becomes a major limitation in using machine learning methods that rely on the available labeled observations. To address this issue, the study proposes that synthetic samples are added to train the deep learning network by using three previously undiscussed physical mechanisms as follows: (1) extreme heavy rainfalls when the soil water is saturated, (2) long-duration rainless events when soil water is exhausted, and (3) the monotonic relationship between rainfall and runoff. A physics-guided Long Short-Term Memory (LSTM) network, shortly named PHY-LSTM, is then formulated. PHY-LSTM network is trained on 531 basins of the Catchment Attributes and Meteorology for Large-sample Studies (CAMELS) dataset, indicating that the performance is significantly improved compared to conventional LSTM. Specifically, the mean Nash-Sutcliffe Efficiency (NSE) improves from 0.52 to 0.61 from the daily simulations during the testing period in local models. It is demonstrated that synthetic samples can effectively improve the simulation of flood peaks and reduce the number of negative streamflow, and strong monotonicity is still maintained even if a slight disturbance is involved in the training dataset. The proposed PHY-LSTM shows that physical mechanisms are very useful to improve efficiencies of the data-driven rainfall-runoff model.