Daily reservoir inflow forecasting using multiscale deep feature learning with hybrid models

[1] Input variable selection is an important issue associated with the development of several hydrological applications. Determining the optimal input vector from a large set of candidates to characterize a preselected output might result in a more accurate, parsimonious, and, possibly, physically interpretable model of the natural process. In the hydrological context, the modeled system often exhibits nonlinear dynamics and multiple interrelated variables. Moreover, the number of candidate inputs can be very large and redundant, especially when the model reproduces the spatial variability of the physical process. The ideal input selection algorithm should therefore provide modeling flexibility, computational efficiency in dealing with high dimension data set, scalability with respect to input dimensionality and minimum redundancy. In this paper, we propose the tree-based iterative input variable selection algorithm, a novel hybrid model-based/model-free approach specifically designed to fulfill these four requirements. The algorithm structure provides robustness against redundancy, while the tree-based nature of the underlying model ensures the other key properties. The approach is first tested on a well-known benchmark case study to validate its accuracy and subsequently applied to a real-world streamflow prediction problem in the upper Ticino River Basin (Switzerland). Results indicate that the algorithm is capable of selecting the most significant and nonredundant inputs in different testing conditions, including the real-world large data set characterized by the presence of several redundant variables. This permits one to identify a compact representation of the observational data set, which is key to improving the model performance and assisting with the interpretation of the underlying physical processes.

https://agupubs.onlinelibrary.wiley.com/doi/pdf/10.1002/wrcr.20339

Tree‐based iterative input variable selection for hydrological modeling

. Combining randomization methods with ensemble prediction is emerging as an effective option to balance accuracy and computational efficiency in data-driven modelling. In this paper, we investigate the prediction capability of extremely randomized trees (Extra-Trees), in terms of accuracy, explanation ability and computational efficiency, in a streamflow modelling exercise. Extra-Trees are a totally randomized tree-based ensemble method that (i) alleviates the poor generalisation property and tendency to overfitting of traditional standalone decision trees (e.g. CART); (ii) is computationally efficient; and, (iii) allows to infer the relative importance of the input variables, which might help in the ex-post physical interpretation of the model. The Extra-Trees potential is analysed on two real-world case studies – Marina catchment (Singapore) and Canning River (Western Australia) – representing two different morphoclimatic contexts. The evaluation is performed against other tree-based methods (CART and M5) and parametric data-driven approaches (ANNs and multiple linear regression). Results show that Extra-Trees perform comparatively well to the best of the benchmarks (i.e. M5) in both the watersheds, while outperforming the other approaches in terms of computational requirement when adopted on large datasets. In addition, the ranking of the input variable provided can be given a physically meaningful interpretation.

/pdf/assessing-the-predictive-capability-of-randomized-tree-based-1t2ntqaddn.pdf

Assessing the predictive capability of randomized tree-based ensembles in streamflow modelling

Performance evaluation of artificial neural network approaches in forecasting reservoir inflow

Multi-time-step ahead daily and hourly intermittent reservoir inflow prediction by artificial intelligent techniques using lumped and distributed data

This study reports the performance of an M5 model tree (MT) and the effects of pruning and smoothing applied to reservoir inflow prediction. The full year and seasonal monthly time step MT predictions were compared with conventional univariate autoregressive integrated moving average (stochastic) models. It was found that stochastic models could not predict the future inflows in a better way, because the observed series had not followed any particular distribution. However, it was found that the stochastic models showed better improvement using a logarithmic-transformed series, but the logarithmic-transformed MT results showed otherwise. The model validation was performed using the comparison of goodness of fit measures, standard statistics, time series, and scatter plots of predicted inflows with observed inflows. The effect of pruning each leaf in the MT model was also studied. Instead of pruning all the leaves, leading to lesser predictive accuracy, selective pruning was carried out based on the import...

Effect of Pruning and Smoothing while Using M5 Model Tree Technique for Reservoir Inflow Prediction

The effect of data pre-processing while developing artificial intelligence (AI) -based data-driven techniques, such as artificial neural networks (ANN), model trees (MT) and linear genetic programming (LGP), is studied for Pawana Reservoir in Maharashtra, India. The daily one-step-ahead inflow forecasts are compared with flows generated from a univariate autoregressive integrated moving average (ARIMA) model. For the full-year data series, a large error is found mainly due to the occurrence of zero values, since the reservoir is located in an intermittent river. Hence, all the techniques are evaluated using two data series: 18 years of daily full-year inflow data (from 1 January to 31 December); and 18 years of daily monsoon season inflow data (from 1 June to 31 October) to take into account the intermittent nature of the data. The relevant range of inputs for each category is selected based on autocorrelation and partial autocorrelation analyses of the inflow series. Conventional pre-processing ...

Improving the performance of data-driven techniques through data pre-processing for modelling daily reservoir inflow

This study evaluates the performance of two modeling approaches for an intermittent reservoir in series. In the first approach, artificial neural network (ANN) model has been developed and applied to the observed time series of Yedgaon reservoir in Upper Bhima River Basin, India. The ANN model results are compared with conventional univariate auto regressive, auto regressive moving average, and auto regressive integrated moving average models. The number of input nodes in ANN model is varied and found that all the training algorithms performed better with four nodes. Back propagation through time algorithm, conjugate gradient algorithm and radial basis function are employed with combination of activation functions such as sigmoid, hyperbolic tangent, step, quick prop, delta rule delta to achieve the best training network. Principal statistic criteria of the predicted inflow time series are compared with observed inflows. Eleven goodnessof- fit measures are also employed to select the best model in each case. The prediction of inflow using time lagged recurrent network with time delay neural network has resulted in a better scenario. However the high and low inflows are not captured well by the ANN network. In the second approach, the network is modified and adopted for predicting seasonal (monsoon) reservoir inflows and the results show accurate mapping of high and low inflows.

Alka S. Kote

Papers

Effect of Pruning and Smoothing while Using M5 Model Tree Technique for Reservoir Inflow Prediction

Improving the performance of data-driven techniques through data pre-processing for modelling daily reservoir inflow

Monthly Reservoir Inflow Modeling Using Time Lagged Recurrent Networks