Listen

This article explores the forecasting capabilities of three classic linear and nonlinear autoregressive modeling techniques and proposes a new ensemble evolutionary time series approach to model and forecast daily dynamics in stream dissolved organic carbon (DOC). The model used data from the Oulankajoki River basin, a boreal catchment in Northern Finland. The models that were evolved used both accuracy and parsimony measures including autoregressive (AR), vector autoregressive (VAR), and self-exciting threshold autoregressive (SETAR). The new method, called genetic-based SETAR (GTAR), evolved through the integration of state-of-the-art genetic programming with SETAR. To develop the models, high-resolution DOC concentration and daily streamflow (as the external input for VAR) were measured at the same gauging station throughout the ice free season. The results showed that all the models characterize the DOC dynamics with an acceptable 1-day-ahead forecasting accuracy. Use of the streamflow time series as an exogenous variable did not increase the predictive accuracy of AR models. Moreover, the hybrid GTAR provided the best accuracy for the holdout testing data and proved to be a suitable approach for predicting DOC in boreal conditions.

Listen

  • A new ensemble model called GTAR, was introduced for DOC time series modeling and prediction.

  • GTAR integrates self-exciting threshold autoregressive (SETAR) and genetic programming.

  • GTAR produces more accurate predictions than AR, VAR, and SETAR.

Graphical Abstract

Graphical Abstract
Graphical Abstract
View largeDownload slide
Graphical Abstract
Graphical Abstract
View largeDownload slide
dissolved organic carbon, exogenous input, genetic programming, Oulankajoki river, streamflow, threshold autoregressive modeling
Listen

Dissolved organic carbon (DOC) is an important water quality variable in aquatic systems that are controlled by a variety of in-stream and catchment processes. DOC is critically important to streams’ biotic and abiotic health as it provides sustenance to microbes at the base of aquatic food webs, controls the light penetration in water, and initiates photochemical reactions (Spencer et al. 2009; Moran et al. 2016). Additionally, DOC influences drinking water quality and purification processes, particularly in communities that rely on rivers and lakes as freshwater resources (Ledesma et al. 2012; Gough et al. 2014). However, DOC transport, concentrations, and loading patterns are being altered locally by land features and storm patterns (Mancinelli et al. 2018) as well as globally by climate change (Liu & Wang 2022) and land-use change (Seybold et al. 2019). DOC dynamics in the northern latitudes are particularly vulnerable to climatic-related changes due to strong seasonality in DOC (Tiwari et al. 2018). These changes are especially notable in boreal areas due to longer growing seasons and increasing precipitation, while intensive forestry and agriculture have already increased DOC loading (Asmala et al. 2019). Due to the significant role of DOC in the riverine systems, many studies have focused on the process understanding of DOC transport with intensive measurements and process-based models (Jones et al. 2014; Rawlins et al. 2021). However, many current approaches are time-consuming to apply to new locations, and thus also, traditional time series analyses are needed to further develop DOC prediction and forecast.

In previous studies, several linear regression methods, such as multiple linear regression (MLR) and partial least-squares regression (PLSR), were applied for the direct prediction of DOC concentration (e.g., Escalas et al. 2003; Marhaba et al. 2003; Ågren et al. 2010; Etheridge et al. 2014; Hestir et al. 2015). For example, Ågren et al. (2010) implemented linear regression to model and predict DOC concentrations in Swedish catchments during a spring flood. Comparing different easily available landscape data and base flow chemistry, the authors demonstrated that the best regression model included base flow, DOC, mean annual runoff, and wetland coverage as the dominant predictors. Etheridge et al. (2014) implemented PLSR, lasso, and stepwise regression techniques to find a relationship between the absorption spectra, collected by ultraviolet-visual spectrometers, and multiple water quality parameters in a brackish tidal marsh. The results showed a high covariance between measured optical properties and some material concentrations such as nitrate, total nitrogen, and DOC if independent calibration is developed for each site using the PLSR technique. Recent studies have demonstrated the partial success of linear regression techniques, particularly when the sample size and range are limited (Etheridge et al. 2014; Zerafati et al. 2022). Moreover, site-specific measurement and analysis become costly when temporal resolution boosts. To address these issues, nonlinear methods using machine learning (ML) approaches such as artificial neural networks (ANNs), random forest (RF), and support vector regression (SVR), have been implemented (Sharghi et al. 2018; Elkiran et al. 2019; Zerafati et al. 2022; Toming et al. 2020; Gul et al. 2021). For instance, Zerafati et al. (2022) demonstrated that DOC dynamics in Groves Creek, a salt marsh creek in eastern Georgia, USA, can ideally be modeled using SVR or RF paired with cost-free auxiliary predictors such as online rainfall, streamflow, and tidal stage data.

Despite the complexity of RF or SVR calibration, a main drawback is their implicit structure so that the modeler can scarcely distill insight from the model, albeit it provides accurate predictions. In this technical study, we therefore, aimed at developing prediction models for DOC that not only are precise but also may provide explicit structures. Thus, they would be simpler to be interpreted and implemented in practical applications for DOC prediction challenges. To this end, we explored the efficiency of two linear models namely autoregressive (AR) and vector autoregressive (VAR), as well as a nonlinear autoregressive model, called self-exciting threshold autoregressive (SETAR), for 1-day-ahead forecasting of DOC series in a boreal river catchment. In addition, we introduced a new hybrid ML model in which capabilities of evolutionary programming are integrated with SETAR solutions to increase the generalization abilities of a classic SETAR approach.

In boreal catchments dissolved organic matter typically comes from waterlogged areas (wetlands/peatlands) and forest areas (litter, surface soil). In our case study area, the Oulankariver basin, DOC transport is controlled by the processes in the peatlands (source). However, hydrological connectivity (= runoff) is controlling the transportation of DOC from wetlands to downstream. This study uses instream measurements for predicting DOC time series, thus spatial analysis (hydrological connectivity) is not in current paper targets.

Listen

Case study area and data

Listen
The present study was conducted in the Oulankajoki (66°2′ N, 29°19′ E), a large river basin in northern Finland (Figure 1). Most of the main rivers and tributaries are ice-covered from mid-November to early May, and the water temperature is highest (up to 21–23 °C) in July (Saraniemi et al. 2008). The annual mean flow is 25.5 m3 s−1, and minimum flows down to 3 m3 s−1 occur in late winter, and peak flows soon after snowmelt (in May–June, up to 249 m3 s−1). The Oulankajoki River is partly located in Oulanka National Park, and its upper parts contain forestry and peatland drainage activities and rural settlements. The landscape is peatland-dominated and forest typical boreal forest with pine and spruce species.
Figure 1
The Oulankajoki River basin and location of gauging station.
View largeDownload slide

The Oulankajoki River basin and location of gauging station.

Figure 1
The Oulankajoki River basin and location of gauging station.
View largeDownload slide

The Oulankajoki River basin and location of gauging station.

Close modal

Discharge is measured at the Oulanka Research Station location (Figure 1) using a gauging station. Discharge data were obtained from Finnish Environment Institute's open access Hertta-data archive. The DOC was measured with 15 min intervals using TriOS Opus UV spectral sensor (TriOS, Germany), which was calibrated and validated against measured DOC water samples in Oulanka research station laboratory. The purpose of the paper was to develop and test new prediction tool to predict DOC concentration in daily resolution, thus data from 5 December 2020 to 31 December 2020 were aggregated to mean daily resolution for the purposes of this study. Our approach could be applied also in higher resolutions, however, in larger river systems DOC processes are slower and thus daily resolution is justified. In smaller headwater systems, higher hourly or even 15 min resolution is reasonable.

The daily discharge and DOC series as a function of time (Qt and DOCt, respectively) are depicted in Figure 2, and their associated statistical features are summarized in Table 1. To account for the observed negative trend as well as the hidden cyclic oscillations of both Qt and DOCt time series, they were made stationary via differencing before the evolution of the linear AR and VAR models.
Table 1

Statistical features of the observed streamflow at Oulankajoki gauging station

VariableMeanMinimumMaximumStd. dev.Skewness
Discharge (m3 s−1) 45.73 5.47 300.0 61.79 1.35 
DOC (mg l−1) 9.52 6.10 13.22 2.05 0.22 
VariableMeanMinimumMaximumStd. dev.Skewness
Discharge (m3 s−1) 45.73 5.47 300.0 61.79 1.35 
DOC (mg l−1) 9.52 6.10 13.22 2.05 0.22 
View Large
Figure 2
Measured discharge and DOC at Oulankajoki gauging station from 5 December 2020 to 31 December 2020.
View largeDownload slide

Measured discharge and DOC at Oulankajoki gauging station from 5 December 2020 to 31 December 2020.

Figure 2
Measured discharge and DOC at Oulankajoki gauging station from 5 December 2020 to 31 December 2020.
View largeDownload slide

Measured discharge and DOC at Oulankajoki gauging station from 5 December 2020 to 31 December 2020.

Close modal

Input selection

Listen
Autocorrelation functions (ACF) of the Qt and DOCt time series (Figure 3) demonstrate high correlation in their sequential values so that the future magnitude can be projected from their previous values. Figure 3 implies that linear dependency among daily DOC values decreases slower than discharge. Thus, a univariate DOC time series model, which uses historical DOC values, could be developed for DOC prediction for the study area. In this study, the daily DOCt time series was modeled using two linear (AR and VAR) and two nonlinear (SETAR and genetic-based SETAR (hereafter GTAR)) approaches. The models were trained and tested using the first 70% and the last 30% of the observations, respectively. Indeed, the last 30% of data was held out as unseen values to test the models’ generalization skills. The best model for each scheme was selected based on the Akaike information criterion (AIC) that considers both accuracy and parsimony measures.
Figure 3
Sample autocorrelation functions of DOC and discharge measurements at Oulankajoki gauging station.
View largeDownload slide

Sample autocorrelation functions of DOC and discharge measurements at Oulankajoki gauging station.

Figure 3
Sample autocorrelation functions of DOC and discharge measurements at Oulankajoki gauging station.
View largeDownload slide

Sample autocorrelation functions of DOC and discharge measurements at Oulankajoki gauging station.

Close modal

Autoregressive models

Listen
AR is a univariate time series modeling approach that allows the modeler to forecast a stationary time series using its historical values. The number of predictors, also known as regressors that are immediately previous values of the response variable (here ⁠) is called the order (p) of the autoregressive (AR) model. For example, an AR model to predict DOCt based on its observed values in the preceding 2 days can be expressed as:
(1)
where the is the model constant, is the coefficient of the AR model, and the is the error term.

The best order for an AR model is commonly determined by assessing the partial autocorrelation sequence (Partial ACF) of the desired time series. However, Partial ACF does not meet the parsimony conditions. To cope with this problem, we considered AIC as suggested in the literature (Tür 2020).

VAR models

Listen
VAR is a bivariate autoregressive time series modeling approach that allows the modeler to forecast the current values of a response variable (here ⁠) using weighted combinations of its previous values and weighted combinations of historical values of an external variable known as the exogenous variable (here ⁠). The number of delays, i.e., regressors, that are immediately previous values of the response variable (here ⁠) is called the order (p) of VAR model. For example, a VAR model with the lag order 4 (i.e., VAR (4)) to predict based on DOC and Q values can be expressed as follows:
(2)
where the is the model constant, are the coefficients of the VAR model. Like AR, the optimum number of delays is selected through minimizing information criteria such as AIC and Bayesian information criterion (BIC).

Self-exciting threshold autoregressive (SETAR)

Listen

Threshold AutoRegressive (TAR) is a popular class of autoregressive models that was introduced as a simple and parsimonious method for nonlinear time series modeling. Despite the simplicity, TAR modeling suffers from a variety of parameters/variables that must be estimated or defined by modelers. To tackle these shortcomings, Tong (2012) proposed self-exciting TAR (SETAR) as a superior type of TAR that has been applied to model a variety of nonlinear time series comprising limit cycles, chaos, harmonic distortion, jumps, and time irreversibility features (e.g., Gonzalo & Wolf 2005; Amiri 2015).

In comparison to TAR, SETAR (Tiao & Tsay 1994) allows a higher degree of flexibility in model parameters through a regime-switching behavior and more accurate forecasts if the time series contains threshold-type nonlinearity in which behaviors (mean, variance, and autocorrelation) of the series change once the series enters a different regime. Given a time series, the number of thresholds (r) determines the number of regimes. It should be set so that each regime has at least a fraction of observations that are enough to produce reliable estimates of the associated AR model's parameter (the order p). Following Gonzalo & Pitarakis (2002), to attain the best SETAR model, we used an AIC-based parameter selection criterion. Considering a time series of ⁠, a two-regime SETAR model can be expressed as follow:
(3)
(4)
where = model intercepts and are the AR models’ coefficients, p1 and p2 are the orders of AR models for lower and upper regimes, and is the residuals time series. The residual value at time t is the value d at a lag or delay time t.

The new genetic-based SETAR model

Listen
GTAR is an ensemble time series modeling approach that integrates the evolutionary prediction skill of the state-of-the-art genetic programming (GP) with the threshold-based time series modeling capability of SETAR. GP is a symbolic regression technique that employs evolutionary operations (i.e., reproduction, crossover, and mutation) to find an optimum structure (linear or nonlinear) between the regressors and a desired target variable. As illustrated in Figure 4, in the first stage, the best SETAR model is trained with respect to the optimum threshold and lagged values determined through AIC. Then, the SETAR outputs are divided into the same train and unseen test subsets and are entered into a GP engine together with those of the optimum DOC lags (predictors). In this stage, we used GPdotNet engine as suggested by Danandeh Mehr et al. (2022) for evolutionary time series modeling. After the evolutionary modeling, the model output performance and generalization ability are controlled by considering statistical indicators (explained later in Section 3.5), and the best model is reported. For details on evolutionary modeling and GP engine alternatives, the reader is referred to Hrnjica & Danandeh Mehr (2018).
Figure 4
Flowchart of the genetic-based SETAR (GTAR) model proposed for DOC timeseries forecasting.
View largeDownload slide

Flowchart of the genetic-based SETAR (GTAR) model proposed for DOC timeseries forecasting.

Figure 4
Flowchart of the genetic-based SETAR (GTAR) model proposed for DOC timeseries forecasting.
View largeDownload slide

Flowchart of the genetic-based SETAR (GTAR) model proposed for DOC timeseries forecasting.

Close modal

Model selection

Listen
The models’ performances are compared using two statistical indicators: Nash–Sutcliff Efficiency (NSE) and root mean square error (RMSE). While the former presents a normalized relative error, the latter reflects the models’ average error. As previously mentioned, two information criteria, namely AIC and BIC, were also implemented to determine the optimum number of delays in univariate and bivariate models, respectively.
(5)
(6)
(7)
(8)
where is the measured DOC, is the DOC value attained by the models, and n is the number of measurements in the training (12 May 2020 to 22 October 2020) and the testing (23 October 2020 to 31 December 2020) periods. The SSR is the sum of squared residuals, and k denotes the number of independently adjusted parameters within the model.
Listen

Preprocessing the data

Listen
Stationarity is an essential feature in time series modeling when linear AR and VAR methods are used. Particularly, the window trend also needs to be carefully considered in the case of modeling via VAR (Hill 2007). Thus, an augmented Dicky Fuller test on the standardized Qt and DOCt time series was done to test if a unit root is present in the time series (null hypothesis). The test statistic revealed that we could not reject the null hypothesis of the test, albeit the negative trend was alleviated through the initial standardization of the data. Thus, differencing technique with the orders of one and two was applied to both series, and the results were controlled again using the same test. The results showed that the first-order differencing failed to remove the short-window trend in the discharge measurements, which is required for VAR modeling. Therefore, the second-order differencing was adopted to achieve stationary series (Figure 5). To develop nonlinear models, the use of stationary time series is not a must; however, we normalized the DOCt time series using min–max normalization to alleviate the sharp difference in their standard deviation and abolish the dimensional inconsistency of the hybrid model.
Figure 5
Stationary state of discharge and DOC timeseries.
View largeDownload slide

Stationary state of discharge and DOC timeseries.

Figure 5
Stationary state of discharge and DOC timeseries.
View largeDownload slide

Stationary state of discharge and DOC timeseries.

Close modal

AR, VAR, and SETAR models

Listen
As illustrated in Figure 6, the combination of partial ACF of the stationary DOCt series with AIC of the AR model in the training period (12 May 2020 to 22 October 2020; 164 days) and RMSE in the testing period (23 October 2020 to 31 December 2020; 70 days) were utilized to determine the order of the best AR model. The figure shows that a significant negative correlation exists within the first five lags of DOC. Considering both model parsimony and accuracy, we selected an order of four, and the associated model was expressed as Equation (9). As we used the second-order of differencing to meet the stationary condition of AR-based modeling, the evolved AR (4) model could be considered the same as an ARIMA (4,2,0) model developed by row DOC time series. Equation (10) expresses the AR model attained for DOC prediction in the study area.
(9)
Figure 6
Correlogram of DOC and AIC utilized to select the best order of AR model.
View largeDownload slide

Correlogram of DOC and AIC utilized to select the best order of AR model.

Figure 6
Correlogram of DOC and AIC utilized to select the best order of AR model.
View largeDownload slide

Correlogram of DOC and AIC utilized to select the best order of AR model.

Close modal
To achieve a parsimonious VAR model, AIC was calculated for different delays varying from 1 to 10. The results showed that the mere use of AIC to favor parsimony does not lead to a clear optimum delay as AIC values change in the limited range of 0.55–0.60. Thus, BIC (see Equation (9)), accompanied by AIC was used to determine the optimum number of delays, i.e., the order of VAR. As shown in Figure 7, adding an extra parameter (delay) in the VAR model increases the penalty term in BIC considerably, which ensures the selection of a parsimonious model. While AIC suggests the use of two or five delays, merely one delay would be the optimum delay based on BIC. This suggests that the previous day's DOC conditions predict rather well DOC transport in the Oulankajoki river catchment. DOC transport to river systems is activated when catchment connectivity from soil profiles to fluvial systems gets activated during rainfall periods. The possibility to use delay function thus means also delays in DOC transport from headwaters to the main river system, and thus enables statistical autocorrelation possibility to predict DOC concentration. Accordingly, VAR (1) was adopted as the best bivariate model and expressed in Equation (10).
(10)
Figure 7
Variations of AIC and BIC were used to select the best order of the VAR model.
View largeDownload slide

Variations of AIC and BIC were used to select the best order of the VAR model.

Figure 7
Variations of AIC and BIC were used to select the best order of the VAR model.
View largeDownload slide

Variations of AIC and BIC were used to select the best order of the VAR model.

Close modal
To develop nonlinear SETAR, TSA packages from the R library were used. As for the AR and VAR models, the least-squares method was used to optimize the models’ parameters in which modeling residuals are minimized using merely the training dataset. As previously discussed, the optimum threshold and the order of the associated AR models were determined based on AIC. As a result, the best orders p1 = 3, p2 = 1 and d = 2, and r = 6.816 were obtained. Thus, the optimum SETAR (2, 3, 1) was attained for the DOC time series as expressed below:
(11)

GTAR modeling

Listen
As illustrated in Figure 4, the best SETAR outputs together with the first three delays of the DOC time series were used as the predictors to train the GTAR. Like the nonevolutionary models, the output label was the DOC values for the next day. The implemented training tool (i.e., GPdotNet) uses arithmetic functions (+, ⁠,⁠, /) to link the inputs and RMSE as a cost function to attain the best GTAR structure. The population of GP solutions with an initial tree depth of four and maximum depth of six initialized by half and half process and iterated up to 500 generations with crossover and mutation rates of 0.9 and 0.05, respectively. A list of five random constants within the range 0.0–1.0, which is consistent with the range of input variables, was also used as the elements of the terminal set to increase the tool flexibility to evolve the best fit. After 10 runs to model a total of 161 days of DOC observation, a solution that provides the least RMSE for the holdout DOC series in the test set was chosen as the best GTAR model. Figure 8 shows the best GTAR tree and its equivalent binary function was given in Equation (12). The figure indicates that the best model was configured by a linear combination of SETAR output and streamflow.
(12)
Figure 8
The GTAR model for DOC preciction in the study area.
View largeDownload slide

The GTAR model for DOC preciction in the study area.

Figure 8
The GTAR model for DOC preciction in the study area.
View largeDownload slide

The GTAR model for DOC preciction in the study area.

Close modal

Comparison of the models’ performance

Listen
To further assess the models’ performance, their time series were depicted against observed DOC series (Figure 9) and the accuracy measures were summarized in Table 2. The figure indicates that the models well captured the observed DOC fluctuations; however, they slightly suffer from lagged prediction (timing) error. Between the linear models, VAR shows better performance than AR during the training phase; however, the AR outperformed VAR in the testing phase. This result indicates that the use of discharge as the exogenous variable may increase the accuracy of DOC time series models, but the model's generalization ability could be reduced as long as a linear combination is considered. However, the process is highly seasonally variable and the quality of the developed prediction models should further be tested also in other seasons and years.
Table 2

The predictive performance of the models developed for 1-day-ahead DOC forecasting

MethodModelTraining phase
Testing phase
RMSENSERMSENSE
Linear AR (4) 0.349 0.972 0.268 0.958 
VAR (1) 0.312 0.978 0.289 0.951 
Nonlinear SETAR 0.328 0.976 0.290 0.951 
GTAR 0.311 0.978 0.212 0.974 
MethodModelTraining phase
Testing phase
RMSENSERMSENSE
Linear AR (4) 0.349 0.972 0.268 0.958 
VAR (1) 0.312 0.978 0.289 0.951 
Nonlinear SETAR 0.328 0.976 0.290 0.951 
GTAR 0.311 0.978 0.212 0.974 
View Large
Figure 9
Time series plots of the models’ estimation vs. observed DOC (mg l−1) data sets at training (top panel) and the testing (bottom panel) periods.
View largeDownload slide

Time series plots of the models’ estimation vs. observed DOC (mg l−1) data sets at training (top panel) and the testing (bottom panel) periods.

Figure 9
Time series plots of the models’ estimation vs. observed DOC (mg l−1) data sets at training (top panel) and the testing (bottom panel) periods.
View largeDownload slide

Time series plots of the models’ estimation vs. observed DOC (mg l−1) data sets at training (top panel) and the testing (bottom panel) periods.

Close modal

Considering the nonlinear models, no change was observed in the models’ accuracy during the training period. While the SETAR performance is close to the VAR during the testing period, the GTAR provides a significant increase in the models’ accuracy during the testing period. Regarding the model's performance in the testing phase, the table proves the superiority of GTAR over its counterparts.

Eventually, scatter plots and Taylor scatter of the observed and predicted DOC values together with violin plots of the prediction errors during the testing phase were depicted in Figures 10 and 11. The scatter plots also provide a linear regression line and their associated correlation coefficients (R2) fitted on each data. Figure 10 shows the AR, VAR, SETAR, and GTAR model scatter plots with correlation of 0.965, 0.957, 0.956, and 0.981, respectively. Overall, the GTAR model is 2.44, 2.37, and 1.57% better than SETAR, VAR, and AR models in terms of the correlation coefficient, respectively. The Taylor diagram provides a concise statistical summary of how well the models’ prediction and the observed data match each other in terms of their correlation, their RMSE and the ratio of their variances. The figure shows that the GTAR with acorrelation of 0.98 is in a closer position to the observation than SETAR, VAR, and AR models. Like Figure 10(b) and 10(c), the VAR and SETAR models are in the same location indicating identical correlation and standard deviation. Figure 11(b) showed that the frequency of prediction error generated by the GTAR has the highest likelihood of being zero. Besides, the uncertainties from the GTAR are highly concentrated around the median showing wide in the middle and narrow onboth sides.
Figure 10
Scatter plots of the models’ estimation vs. observed DOC data sets: (a) AR, (b) VAR, (c) SETAR, and (d) GTAR.
View largeDownload slide

Scatter plots of the models’ estimation vs. observed DOC data sets: (a) AR, (b) VAR, (c) SETAR, and (d) GTAR.

Figure 10
Scatter plots of the models’ estimation vs. observed DOC data sets: (a) AR, (b) VAR, (c) SETAR, and (d) GTAR.
View largeDownload slide

Scatter plots of the models’ estimation vs. observed DOC data sets: (a) AR, (b) VAR, (c) SETAR, and (d) GTAR.

Close modal
Figure 11
Taylor diagram (a) and violin plots (b) of the models’ estimation vs. observed DOC.
View largeDownload slide

Taylor diagram (a) and violin plots (b) of the models’ estimation vs. observed DOC.

Figure 11
Taylor diagram (a) and violin plots (b) of the models’ estimation vs. observed DOC.
View largeDownload slide

Taylor diagram (a) and violin plots (b) of the models’ estimation vs. observed DOC.

Close modal
Listen

DOC concentration cannot always be directly measured in situ (expensive sensors), and prohibitively expensive laboratory-based analysis is required to capture high-resolution dynamics of DOC variation in different aquatic systems. Thus, the data-driven models possessing high generalization ability are of paramount importance and required for ecological modeling and water quality management. Previous DOC model development endeavors did not fully consider linear-nonlinear fusion models, so this paper carried out research on the comparison and combination of TAR models with the state-of-the-art GP in the Oulankajoki River basin. The experimental results showed that the proposed fusion model, called GTAR, had the best prediction accuracy and the strongest correlation with the observed DOC compared with the AR, VAR, and SETAR. Our modeling study shows clearly that autocorrelation statistical modeling can be efficiently used for 1-day-ahead forecasting of DOC concentration. This allows methodology to be improved for example drinking water purification processes or predicting ecological conditions in the river systems. Our approach should however further be tested with longer time series, variable seasons, and also in a different kind for river systems.

Listen

This research was funded by HYDRO-RDI network, HYDRO-RI-platform and Green-Digi-Basin, Academy of Finland (337523, 346163, 347704) and Freshwater Competence Centre (FWCC).

Listen

Data cannot be made publicly available; readers should contact the corresponding author for details.

Listen

The authors declare there is no conflict.

Ågren
A.
,
Buffam
I.
,
Bishop
K.
 &
Laudon
H.
2010
Modeling stream dissolved organic carbon concentrations during spring flood in the boreal forest: a simple empirical approach for regional predictions
.
Journal of Geophysical Research: Biogeosciences
115
(
G1
).
doi:10.1029/2009JG001013
.
Google Scholar
 
Amiri
E.
2015
Forecasting daily river flows using nonlinear time series models
.
Journal of Hydrology
527
,
1054
–
1072
.
doi:10.1016/j.jhydrol.2015.05.048
.
Google Scholar
Crossref
Search ADS
 
Asmala
E.
,
Carstensen
J.
 &
Räike
A.
2019
Multiple anthropogenic drivers behind upward trends in organic carbon concentrations in boreal rivers
.
Environmental Research Letters
14
(
12
),
124018
.
doi:10.1088/1748-9326/ab4fa9
.
Google Scholar
Crossref
Search ADS
 
Danandeh Mehr
A.
,
Ghadimi
S.
,
Marttila
H.
 &
Torabi Haghighi
A.
2022
A new evolutionary time series model for streamflow forecasting in boreal lake-river systems
.
Theoretical and Applied Climatology
148
(
1
),
255
–
268
.
doi:10.1007/s00704-022-03939-3
.
Google Scholar
 
Elkiran
G.
,
Nourani
V.
 &
Abba
S. I.
2019
Multi-step ahead modelling of river water quality parameters using ensemble artificial intelligence-based approach
.
Journal of Hydrology
577
,
123962
.
doi:10.1016/j.jhydrol.2019.123962
.
Google Scholar
Crossref
Search ADS
 
Escalas
A.
,
Droguet
M.
,
Guadayol
J. M.
 &
Caixach
J.
2003
Estimating DOC regime in a wastewater treatment plant by UV deconvolution
.
Water Research
37
(
11
),
2627
–
2635
.
doi:10.1016/S0043-1354(03)00063-0
.
Google Scholar
Crossref
Search ADS
PubMed
 
Etheridge
J. R.
,
Birgand
F.
,
Osborne
J. A.
,
Osburn
C. L.
,
Burchell
M. R.
 &
Irving
J.
2014
Using in situ ultraviolet-visual spectroscopy to measure nitrogen, carbon, phosphorus, and suspended solids concentrations at a high frequency in a brackish tidal marsh
.
Limnology and Oceanography: Methods
12
(
1
),
10
–
22
.
doi:10.4319/lom.2014.12.10
.
Google Scholar
 
Gonzalo
J.
 &
Pitarakis
J. Y.
2002
Estimation and model selection based inference in single and multiple threshold models
.
Journal of Econometrics
110
(
2
),
319
–
352
.
doi:10.1016/S0304-4076(02)00098-2
.
Google Scholar
Crossref
Search ADS
 
Gonzalo
J.
 &
Wolf
M.
2005
Subsampling inference in threshold autoregressive models
.
Journal of Econometrics
127
(
2
),
201
–
224
.
doi:10.1016/j.jeconom.2004.08.004
.
Google Scholar
Crossref
Search ADS
 
Gough
R.
,
Holliman
P. J.
,
Heard
T. R.
 &
Freeman
C.
2014
Dissolved organic carbon and trihalomethane formation potential removal during coagulation of a typical UK upland water with alum, PAX-18 and PIX-322
.
Journal of Water Supply: Research and Technology – AQUA
63
(
8
),
650
–
660
.
doi:10.2166/aqua.2014.007
.
Google Scholar
Crossref
Search ADS
 
Gul
E.
,
Safari
M. J. S.
,
Torabi Haghighi
A.
 &
Danandeh Mehr
A.
2021
Sediment transport modeling in non-deposition with clean bed condition using different tree-based algorithms
.
PLoS One
16
(
10
),
e0258125
.
doi:10.1371/journal.pone.0258125
.
Google Scholar
Crossref
Search ADS
PubMed
 
Hestir
E. L.
,
Brando
V.
,
Campbell
G.
,
Dekker
A.
 &
Malthus
T.
2015
The relationship between dissolved organic matter absorption and dissolved organic carbon in reservoirs along a temperate to tropical gradient
.
Remote Sensing of Environment
156
,
395
–
402
.
doi:10.1016/j.rse.2014.09.022
.
Google Scholar
Crossref
Search ADS
 
Hrnjica
B.
 &
Danandeh Mehr
A.
2018
Optimized Genetic Programming Applications: Emerging Research and Opportunities
.
IGI Global
, PA.
doi:10.4018/978-1-5225-6005-0
.
Google Scholar
 
Jones
T. D.
,
Chappell
N. A.
 &
Tych
W.
2014
First dynamic model of dissolved organic carbon derived directly from high-frequency observations through contiguous storms
.
Environmental Science & Technology
48
(
22
),
13289
–
13297
.
doi:10.1021/es503506m
.
Google Scholar
Crossref
Search ADS
 
Ledesma
J. L.
,
Köhler
S. J.
 &
Futter
M. N.
2012
Long-term dynamics of dissolved organic carbon: implications for drinking water supply
.
Science of the Total Environment
432
,
1
–
11
.
doi:10.1016/j.scitotenv.2012.05.071
.
Google Scholar
Crossref
Search ADS
PubMed
 
Liu
F.
 &
Wang
D.
2022
Dissolved organic carbon concentration and biodegradability across the global rivers: a meta-analysis
.
Science of the Total Environment
818
,
151828
.
doi:10.1016/j.scitotenv.2021.151828
.
Google Scholar
Crossref
Search ADS
PubMed
 
Mancinelli
E.
,
Baltrėnaitė
E.
,
Baltrėnas
P.
,
Grubliauskas
R.
,
Marčiulaitienė
E.
,
Passerini
G.
 &
Pranskevičius
M.
2018
Influence of biochar from slow pyrolysis on dissolved organic carbon and total dissolved nitrogen levels of urban storm-water runoff
.
Water, Air, & Soil Pollution
229
(
8
),
1
–
18
.
doi:10.1007/s11270-018-3896-9
.
Google Scholar
 
Marhaba
T. F.
,
Bengraïne
K.
,
Pu
Y.
 &
Aragó
J.
2003
Spectral fluorescence signatures and partial least squares regression: model to predict dissolved organic carbon in water
.
Journal of Hazardous Materials
97
(
1-3
),
83
–
97
.
doi:10.1016/S0304-3894(02)00246-7
.
Google Scholar
Crossref
Search ADS
PubMed
 
Moran
M. A.
,
Kujawinski
E. B.
,
Stubbins
A.
,
Fatland
R.
,
Aluwihare
L. I.
,
Buchan
A.
,
Crump
B. C.
,
Dorrestein
P. C.
,
Dyhrman
S. T.
,
Hess
N. J.
 &
Howe
B.
2016
Deciphering ocean carbon in a changing world
.
Proceedings of the National Academy of Sciences
113
(
12
),
3143
–
3151
.
doi:10.1073/pnas.1514645113
.
Google Scholar
Crossref
Search ADS
 
Rawlins
M. A.
,
Connolly
C. T.
 &
McClelland
J. W.
2021
Modeling terrestrial dissolved organic carbon loading to western Arctic rivers
.
Journal of Geophysical Research: Biogeosciences
126
(
10
),
e2021JG006420
.
doi:10.1029/2021JG006420
.
Google Scholar
PubMed
 
Saraniemi
M.
,
Huusko
A.
 &
Tahkola
H.
2008
Spawning migration and habitat use of adfluvial brown trout, salmo trutta, in a strongly seasonal boreal river
. Boreal Environment Research
13
,
121
–
132
.
Google Scholar
 
Seybold
E.
,
Gold
A. J.
,
Inamdar
S. P.
,
Adair
C.
,
Bowden
W. B.
,
Vaughan
M. C.
,
Pradhanang
S. M.
,
Addy
K.
,
Shanley
J. B.
,
Vermilyea
A.
 &
Levia
D. F.
2019
Influence of land use and hydrologic variability on seasonal dissolved organic carbon and nitrate export: insights from a multi-year regional analysis for the northeastern USA
.
Biogeochemistry
146
(
1
),
31
–
49
.
doi:10.1007/s10533-019-00609-x
.
Google Scholar
Crossref
Search ADS
 
Sharghi
E.
,
Nourani
V.
 &
Behfar
N.
2018
Earthfill dam seepage analysis using ensemble artificial intelligence based modeling
.
Journal of Hydroinformatics
20
(
5
),
1071
–
1084
.
DOI: 10.2166/hydro.2018.151
.
Google Scholar
Crossref
Search ADS
 
Spencer
R. G.
,
Stubbins
A.
,
Hernes
P. J.
,
Baker
A.
,
Mopper
K.
,
Aufdenkampe
A. K.
,
Dyda
R. Y.
,
Mwamba
V. L.
,
Mangangu
A. M.
,
Wabakanghanzi
J. N.
 &
Six
J.
2009
Photochemical degradation of dissolved organic matter and dissolved lignin phenols from the Congo river
.
Journal of Geophysical Research: Biogeosciences
114
(
G3
).
doi:10.1029/2009JG000968
.
Google Scholar
 
Tiao
G. C.
 &
Tsay
R. S.
1994
Some advances in non-linear and adaptive modelling in time-series
.
Journal of Forecasting
13
(
2
),
109
–
131
.
doi:10.1002/for.3980130206
.
Google Scholar
Crossref
Search ADS
 
Tiwari
T.
,
Sponseller
R. A.
 &
Laudon
H.
2018
Extreme climate effects on dissolved organic carbon concentrations during snowmelt
.
Journal of Geophysical Research: Biogeosciences
123
(
4
),
1277
–
1288
.
doi:10.1002/2017JG004272
.
Google Scholar
Crossref
Search ADS
 
Toming
K.
,
Kotta
J.
,
Uuemaa
E.
,
Sobek
S.
,
Kutser
T.
 &
Tranvik
L. J.
2020
Predicting lake dissolved organic carbon at a global scale
.
Scientific Reports
10
(
1
),
1
–
8
.
Available from: https://www.nature.com/articles/s41598-020-65010-3
Google Scholar
Crossref
Search ADS
PubMed
 
Tür, R. 2020 Maximum wave height hindcasting using ensemble linear-nonlinear models. Theoretical and Applied Climatology 141 (3–4), 1151–1163. doi:10.1007/s00704-020-03272-7
Zerafati, M. E., Bozorgi-Amiri, A., Golmohammadi, A. M. & Jolai, F. 2022 A multi-objective mixed integer linear programming model proposed to optimize a supply chain network for microalgae-based biofuels and co-products: a case study in Iran. Environmental Science and Pollution Research 1–23. doi: 10.1007/s11356-022-19465-8.
© 2023 The Authors
This is an Open Access article distributed under the terms of the Creative Commons Attribution Licence (CC BY 4.0), which permits copying, adaptation and redistribution, provided the original work is properly cited (http://creativecommons.org/licenses/by/4.0/).