Search

Highlights • Machine learning models are compared for simulation and forecasting of streamflows. • Support Vector Regression, Artificial Neural Networks and Random Forest are used. • A base flow separation method is coupled with machine learning models. • North Fork, Chehalis, Carson and Sacramento rivers in the United States are analysed. • Base flow separation method improved streamflow simulation to a certain degree. Abstract Efficient simulation of rainfall-runoff relationships is one of the most complex problems owing to the high number of interrelated hydrological processes. It is well-known that machine learning models could fail in simulating streamflows from only meteorological variables in the absence of antecedent streamflow values. The main reason for this could be low and lagged relationships between streamflow and meteorological variables. To overcome this inefficiency, for the first time, we developed a simulation framework by coupling a base flow separation method to three machine learning methods. It was demonstrated that separating streamflow into different components such as base flow and surface flow can be useful for improving simulation and forecasting capabilities of machine learning models. We simulated streamflow in four rivers in the United States with Support Vector Regression (SVR), Artificial Neural Networks (ANNs) and Random Forest (RF) as a function of precipitation (P), temperature (T) and potential evapotranspiration (PET). We concluded that the base flow separation method improved the simulation performances of the machine learning models to a certain degree. Apart from the simulation scheme, we also employed a forecasting scheme by using the antecedent observed discharge values in addition to P, T, and PET. We discussed performances of models in simulation and forecasting of streamflow regarding model types, input structures and catchment dynamics in detail. [ABSTRACT FROM AUTHOR]

• Precipitation δ 18O predicted using geostatistical and machine learning algorithms. • Prediction performance evaluated using ordinary and hybrid error metrics. • Machine learning algorithms performed better, particularly using scarcer datasets. • Two spatiotemporally different subcontinental datasets of precipitation δ 18O were used. Stable isotopes of precipitation are important natural tracers in hydrology, ecology, and forensics. The spatially explicit predictions of oxygen and hydrogen isotopes in precipitation are obtained through different interpolation techniques. In the present study we aim to examine the performance of various interpolation techniques when predicting the spatial distribution of precipitation stable isotopes. The efficiency of combined geostatistical tools (i.e. regression kriging; RK) and various machine learning methods (including regression enhanced random forest methods: MRRF, RERF) are compared in interpolating the spatial variability of precipitation stable oxygen isotope values from two different sampling networks in Europe. To assess the performance of the models, mean squared error (MSE), nonparametric Kling Gupta efficiency (KGE), absolute differences and relative mean absolute error metrics were employed. It was found that the combination of the different regression techniques with Random Forest can produce estimations with comparable accuracy in terms of descending order of overall average MSE, MRRF: 2.61, RK: 2.77, RERF: 2.99, RF: 3.08. The best performing combined random forest model variant (MRRF) outperformed regression kriging in terms of a hybrid error metric (KGE) by 7.5%. Sequential random rarefying the station networks showed that machine-learning methods are more capable of maintaining high prediction accuracy even with fewer input data. This can be a great advantage when a suitable method is needed to predict the stable isotope composition of precipitation for large spatial domains where the spatial density of data stations shows large differences. [ABSTRACT FROM AUTHOR]

• Forecasting the short-term daily streamflow is an essential and complex task in hydrology. • A novel stacked machine learning algorithm is compared to a powerful deep learning model. • The stacked model combines Random Forest and Multilayer Perceptron through Elastic Net. • The deep learning model is based on bidirectional Long Short-Term memory networks. • The two models showed comparable performance, strongly affected by the forecast horizon. Prediction of river flow rates is an essential task for both flood protection and optimal water resource management. The high uncertainty associated with basin characteristics, hydrological processes, and climatic factors affecting river flows make streamflow prediction a very challenging problem. These reasons, together with the increasingly wide availability of data relating to flow rates and rainfall, frequently lead to a preference for data-driven models over physically based or conceptual forecasting models. This study shows the results of an in-depth comparison between two different daily streamflow prediction models: a novel simpler model based on the stacking of the Random Forest and Multilayer Perceptron algorithms, using the Elastic Net algorithm as meta -learner, and a more complex model based on bi-directional Long Short-Term Memory (LSTM) networks. Bayesian optimization was employed in selecting hyperparameters. The two prediction models were compared through the analysis of four different case studies: the Bacchiglione River, the Raccoon River, the Wilson River, and the Trent River. The two models showed comparable forecasting capabilities. In the most favourable cases, both models demonstrated very high accuracy (R2 greater than 0.93, Mean absolute percentage error approximately equal to 10%). The stacked model outperformed the bi-directional LSTM network model in several cases in predicting peak flow rates but was less accurate in forecasting low flow rates. In addition, its computation times are significantly shorter. The prediction accuracy of both models decreased as the forecast horizon increased. The length of the time series plays an essential role in developing models with satisfactory forecasting capabilities. In this study, the effectiveness of forecasting models was not influenced by the river regime. However, a high variance of the input dataset and a large number of outliers in the time series can reduce the accuracy of prediction models. [ABSTRACT FROM AUTHOR]

• First reports on the ensemble learning (EL) algorithms for reference evapotranspiration (ETo) prediction in Bangladesh. • Random tree (RT) model outperforms other models to estimate daily ETo. • The RT and RF models have suggested as the promising potentials for daily ETo estimate. • Wind speed and solar radiation are the highest contributing variables influencing the observed and predicted ETo changes. Ensemble learning (EL), an alternative approach in the machine-learning field, offers an accurate reference evapotranspiration (ETo) prediction, which is of paramount significance for the hydrological studies and agricultural water practices. Although the FAO-56 Penman-Monteith (PM) equation is regarded as an ideal model for estimating ETo, its applicability is limited due to the absence of required climatic datasets in many regions of the world. Despite its significance, only a few studies use the EL algorithms for the ETo prediction perspective. In this study, we contribute to fill this gap from a two-fold way. First, we present the potential of new EL Random Tree (RT), Bagging and Random Subspace (RS) algorithms which were compared with two commonly used Random Forest (RF), and Support Vector Machine (SVM) algorithms for predicting daily ETo with the climatic data-limited humid region in Bangladesh during 1983–2017 using a five-fold cross-validation scheme. Second, we explore the role of contributing variables influencing ETo change at the regional scale. When a lack of climatic datasets, Tmax (maximum temperature), Tmin (minimum temperature) and Rs (solar radiation) as 3 input combinations obtained the reasonable precision for estimating ETo in all climatic regions except for south-central and south-western regions. Compared to other state-of-arts models, the RT model performed superior to predict ETo in all input combinations followed by the RF, Bagging, RS, and SVM. Considering less difficulty level, high prediction accuracy, more dependability and fewer computation costs of studied models, RT and RF models have been suggested as the promising potentials for daily ETo estimate in subtropical climate regions of Bangladesh and also may be applicable worldwide in the similar climates. The importance analysis from the RF model depicted that the wind speed (U2) and solar radiation (Rs) are the largest influential variables affecting the observed and predicted daily ETo changes in Bangladesh. [ABSTRACT FROM AUTHOR]