Hindcasting and Forecasting Total Suspended Sediment Concentrations Using a NARX Neural Network
TL;DR: In this article, a non-linear autoregressive exogenous neural network (NARX) was used for forecasting sediment concentrations at the exit of Francia Creek watershed (Valparaiso, Chile).
Abstract: Estimating and forecasting suspended sediments concentrations in streams constitutes a valuable asset for sustainable land management. This research presents the development of a non-linear autoregressive exogenous neural network (NARX) for forecasting sediment concentrations at the exit of Francia Creek watershed (Valparaiso, Chile). Details are presented on input data selection, data splitting, selection of model architecture, determination of model structure, NARX training (optimization of model parameters), and model validation (hindcasting and forecasting). The study explored if the developed artificial neural network model is valid for forecasting daily suspended sediment concentrations for a complete year, capturing seasonal trends, and maximum and baseflow concentrations. Francia Creek watershed covers approximately 3.24 km2. Land cover within the catchment consists mainly of native and exotic vegetation, eroded soil, and urban areas. Input data consisting of precipitation and stream flow time-series were fed to a NARX network for forecasting daily suspended sediments (SST) concentrations for years 2013–2014, and hindcasting for years 2008–2010. Training of the network was performed with daily SST, precipitation, and flow data from years 2012 and 2013. The resulting NARX net consisted of an open-loop, 12-node hidden layer, 100 iterations, using Bayesian regularization backpropagation. Hindcasting of daily and monthly SST concentrations for years 2008 through 2010 was successful. Daily SST concentrations for years 2013 and 2014 were forecasted successfully for baseflow conditions (R2 = 0.73, NS = 0.71, and Kling-Gupta efficiency index (K-G) = 0.84). Forecasting daily SST concentrations for year 2014 was within acceptable statistical fit and error margins (R2 = 0.53, NS = 0.47, K-G = 0.60, d = 0.82). Forecasting of monthly maximum SST concentrations for the two-year period (2013 and 2014) was also successful (R2 = 0.69, NS = 0.60, K-G = 0.54, d = 0.84).
Citations
More filters
13 Sep 2021
TL;DR: In this paper, a closed-loop autoregressive neural network with exogenous inputs was developed to estimate salinity concentrations at a coastal water quality station (BISCC4) in Biscayne Bay, Florida.
Abstract: Estimating salinity concentrations in coastal waters allows characterization of the spatial and temporal dynamics of the freshwater/saltwater interface. In Southeast Florida (USA) the saltwater interface is monitored and evaluated for potential impacts to public supply wellfields and biological communities. In this research, a closed-loop autoregressive neural network with exogenous inputs was developed to estimate salinity concentrations at a coastal water quality station (BISCC4) in Biscayne Bay, Florida. The neural network (ANN) is shown to successfully simulate hourly salinity concentrations for years 2015 through 2019. A statistical comparison of simulated concentrations versus observed data demonstrates that the ANN simulates salinity concentration values and trends within acceptable margin of errors (R2 = 0.59, K-G = 0.64, NSE = 0.33, d = 0.86, PBIAS = 1.5%, RSR = 0.82). In its current form, the ANN model performs better in simulating salinity concentrations and trends, than an existing hydrodynamic model. These results have the potential to be applied to other coastal locations in Biscayne Bay where freshwater inputs from inland streams and canals are affecting salinity concentrations.
2 citations
TL;DR: In this article , a hybrid model that combines two different algorithms to increase the accuracy of short-term berry yield prediction using only previous yield data is presented, where the model integrates both autoregressive integrated moving average (ARIMA) with Kalman filter refinement and neural network techniques, specifically support vector regression (SVR), and nonlinear auto-gressive (NAR) neural networks, to improve prediction accuracy by correcting the errors generated by the system.
Abstract: This study presents a novel hybrid model that combines two different algorithms to increase the accuracy of short-term berry yield prediction using only previous yield data. The model integrates both autoregressive integrated moving average (ARIMA) with Kalman filter refinement and neural network techniques, specifically support vector regression (SVR), and nonlinear autoregressive (NAR) neural networks, to improve prediction accuracy by correcting the errors generated by the system. In order to enhance the prediction performance of the ARIMA model, an innovative method is introduced that reduces randomness and incorporates only observed variables and system errors into the state-space system. The results indicate that the proposed hybrid models exhibit greater accuracy in predicting weekly production, with a goodness-of-fit value above 0.95 and lower root mean square error (RMSE) and mean absolute error (MAE) values compared with non-hybrid models. The study highlights several implications, including the potential for small growers to use digital strategies that offer crop forecasts to increase sales and promote loyalty in relationships with large food retail chains. Additionally, accurate yield forecasting can help berry growers plan their production schedules and optimize resource use, leading to increased efficiency and profitability. The proposed model may serve as a valuable information source for European food retailers, enabling growers to form strategic alliances with their customers.
01 Jan 2023
TL;DR: In this paper , a hybrid model that combines a new approach to the classical Kalman filter with machine learning techniques, i.e., support vector regression (SVR) and nonlinear autoregressive (NAR) neural networks, is proposed to improve the performance of existing predictive models.
Abstract: Accurate time series prediction techniques are becoming fundamental to modern decision support systems. As massive data processing develops in its practicality, machine learning (ML) techniques applied to time series can automate and improve prediction models. The radical novelty of this paper is the development of a hybrid model that combines a new approach to the classical Kalman filter with machine learning techniques, i.e., support vector regression (SVR) and nonlinear autoregressive (NAR) neural networks, to improve the performance of existing predictive models. The proposed hybrid model uses, on the one hand, an improved Kalman filter method that eliminates the convergence problems of time series data with large error variance and, on the other hand, an ML algorithm as a correction factor to predict the model error. The results reveal that our hybrid models obtain accurate predictions, substantially reducing the root mean square and absolute mean errors compared to the classical and alternative Kalman filter models and achieving a goodness of fit greater than 0.95. Furthermore, the generalization of this algorithm was confirmed by its validation in two different scenarios.
References
More filters
TL;DR: In this paper, the authors present guidelines for watershed model evaluation based on the review results and project-specific considerations, including single-event simulation, quality and quantity of measured data, model calibration procedure, evaluation time step, and project scope and magnitude.
Abstract: Watershed models are powerful tools for simulating the effect of watershed processes and management on soil and water resources. However, no comprehensive guidance is available to facilitate model evaluation in terms of the accuracy of simulated data compared to measured flow and constituent values. Thus, the objectives of this research were to: (1) determine recommended model evaluation techniques (statistical and graphical), (2) review reported ranges of values and corresponding performance ratings for the recommended statistics, and (3) establish guidelines for model evaluation based on the review results and project-specific considerations; all of these objectives focus on simulation of streamflow and transport of sediment and nutrients. These objectives were achieved with a thorough review of relevant literature on model application and recommended model evaluation methods. Based on this analysis, we recommend that three quantitative statistics, Nash-Sutcliffe efficiency (NSE), percent bias (PBIAS), and ratio of the root mean square error to the standard deviation of measured data (RSR), in addition to the graphical techniques, be used in model evaluation. The following model evaluation performance ratings were established for each recommended statistic. In general, model simulation can be judged as satisfactory if NSE > 0.50 and RSR < 0.70, and if PBIAS + 25% for streamflow, PBIAS + 55% for sediment, and PBIAS + 70% for N and P. For PBIAS, constituent-specific performance ratings were determined based on uncertainty of measured data. Additional considerations related to model evaluation guidelines are also discussed. These considerations include: single-event simulation, quality and quantity of measured data, model calibration procedure, evaluation time step, and project scope and magnitude. A case study illustrating the application of the model evaluation guidelines is also provided.
9,386 citations
TL;DR: A strong case can be made for moving away from ad hoc use of aggregated efficiency metrics and towards a framework based on purpose-dependent evaluation metrics and benchmarks that allows for more robust model adequacy assessment.
Abstract: . A traditional metric used in hydrology to summarize model
performance is the Nash–Sutcliffe efficiency (NSE). Increasingly an
alternative metric, the Kling–Gupta efficiency (KGE), is used instead. When
NSE is used, NSE = 0 corresponds to using the mean flow as a benchmark
predictor. The same reasoning is applied in various studies that use KGE as
a metric: negative KGE values are viewed as bad model performance, and only
positive values are seen as good model performance. Here we show that using
the mean flow as a predictor does not result in KGE = 0, but instead KGE = 1 - √ 2 ≈ - 0.41 . Thus, KGE values greater than −0.41
indicate that a model improves upon the mean flow benchmark – even if the
model's KGE value is negative. NSE and KGE values cannot be directly
compared, because their relationship is non-unique and depends in part on
the coefficient of variation of the observed time series. Therefore,
modellers who use the KGE metric should not let their understanding of NSE
values guide them in interpreting KGE values and instead develop new
understanding based on the constitutive parts of the KGE metric and the
explicit use of benchmark values to compare KGE scores against. More
generally, a strong case can be made for moving away from ad hoc use of
aggregated efficiency metrics and towards a framework based on
purpose-dependent evaluation metrics and benchmarks that allows for more
robust model adequacy assessment.
524 citations
TL;DR: Bayesian regularized artificial neural networks (BRANNs) as mentioned in this paper are more robust than standard back-propagation nets and can reduce or eliminate the need for lengthy cross-validation.
Abstract: Bayesian regularized artificial neural networks (BRANNs) are more robust than standard back-propagation nets and can reduce or eliminate the need for lengthy cross-validation. Bayesian regularization is a mathematical process that converts a nonlinear regression into a "well-posed" statistical problem in the manner of a ridge regression. The advantage of BRANNs is that the models are robust and the validation process, which scales as O(N2) in normal regression methods, such as back propagation, is unnecessary. These networks provide solutions to a number of problems that arise in QSAR modeling, such as choice of model, robustness of model, choice of validation set, size of validation effort, and optimization of network architecture. They are difficult to overtrain, since evidence procedures provide an objective Bayesian criterion for stopping training. They are also difficult to overfit, because the BRANN calculates and trains on a number of effective network parameters or weights, effectively turning off those that are not relevant. This effective number is usually considerably smaller than the number of weights in a standard fully connected back-propagation neural net. Automatic relevance determination (ARD) of the input variables can be used with BRANNs, and this allows the network to "estimate" the importance of each input. The ARD method ensures that irrelevant or highly correlated indices used in the modeling are neglected as well as showing which are the most important variables for modeling the activity data. This chapter outlines the equations that define the BRANN method plus a flowchart for producing a BRANN-QSAR model. Some results of the use of BRANNs on a number of data sets are illustrated and compared with other linear and nonlinear models.
482 citations
TL;DR: In this paper, a Nonlinear Autoregressive Exogenous (NARX) neural network was used to predict the solar radiation on a horizontal surface of a race sailboat.
Abstract: The solar photovoltaic (PV) energy has an important place among the renewable energy sources. Therefore, several researchers have been interested by its modelling and its prediction, in order to improve the management of the electrical systems which include PV arrays. Among the existing techniques, artificial neural networks have proved their performance in the prediction of the solar radiation. However, the existing neural network models don’t satisfy the requirements of certain specific situations such as the one analyzed in this paper. The aim of this research work is to supply, with electricity, a race sailboat using exclusively renewable sources. The developed solution predicts the direct solar radiation on a horizontal surface. For that, a Nonlinear Autoregressive Exogenous (NARX) neural network is used. All the specific conditions of the sailboat operation are taken into account. The results show that the best prediction performance is obtained when the training phase of the neural network is performed periodically.
238 citations
TL;DR: A systematic protocol for the development and documentation of ANN models is introduced and shows that model architecture selection is the best implemented step, while greater focus should be given to input selection considering input independence and model validation considering replicative and structural validity.
Abstract: The application of Artificial Neural Networks (ANNs) in the field of environmental and water resources modelling has become increasingly popular since early 1990s. Despite the recognition of the need for a consistent approach to the development of ANN models and the importance of providing adequate details of the model development process, there is no systematic protocol for the development and documentation of ANN models. In order to address this shortcoming, such a protocol is introduced in this paper. In addition, the protocol is used to critically review the quality of the ANN model development and reporting processes employed in 81 journal papers since 2000 in which ANNs have been used for drinking water quality modelling. The results show that model architecture selection is the best implemented step, while greater focus should be given to input selection considering input independence and model validation considering replicative and structural validity.
220 citations