Available at www.sciencedirect.com INFORMATION PROCESSING IN AGRICULTURE xxx (xxxx) xxx journal homepage: www.elsevier.com/locate/inpa
Dissolved oxygen prediction in prawn ponds from a group of one step predictors Ashfaqur Rahman *, Joel Dabrowski, John McCulloch Data61, CSIRO, Australia
A R T I C L E I N F O
A B S T R A C T
Article history:
In this paper we have presented a novel approach to predict dissolved oxygen in prawn
Received 3 June 2019
ponds. It is necessary to maintain dissolved oxygen above a certain level in the ponds
Received in revised form
for expected growth and survival of the prawns. An accurate prediction of dissolved oxygen
5 August 2019
can assist farmers to take necessary measures to maintain dissolved oxygen levels ideal for
Accepted 20 August 2019
prawn growth. Existing approaches to dissolved oxygen prediction performs well on short
Available online xxxx
term, however incurs high error on long term prediction. We propose a new approach where a group of predictors are developed where each model predicts a certain time stamps ahead. Each predictor is trained on sampled data so that it predicts a step ahead prediction only, however, the sampling process decides on the actual number of time stamp ahead prediction. Since step ahead predictor acts like a short term predictor, it incurs small error even at higher time stamp ahead prediction. Experimental results demonstrate that the proposed approach achieves significantly lower error on long term prediction compared to other existing approaches. Ó 2019 China Agricultural University. Production and hosting by Elsevier B.V. on behalf of KeAi. This is an open access article under the CC BY-NC-ND license (http://creativecommons. org/licenses/by-nc-nd/4.0/).
1.
Introduction
Dissolved oxygen (DO) plays an important role in aquaculture production. Stress is signalled at 1.4 mg/L [1]. Dissolved oxygen level at 4 or 5 mg/L or higher is considered as ideal in aquaculture production [2,3]. In general, values below 2.0 mg/L are associated with restricted growth and high mortality risk [4]. Low dissolved oxygen in the water can lead to anoxia, slow growth, and death in shrimp and fish. Low level of oxygen concentration can be caused by a number of reasons [2,5]: (i) Dissolved oxygen decreases with increase of temperature and salinity; (ii) Aquatic plants die due to excessive usage of herbicides [4,5] and this can lead to shortage of
dissolved oxygen; (iii) Dissolved oxygen normally decreases during night time. During the day, algae and plankton photosynthesize (with sunlight) and create oxygen dissolved in water. Absence of sunlight at night prohibits photosynthesis activities. Algae and aquatic plant’s respiration continue at night despite the lack of photosynthesis. This leads to reduced oxygen level. Similar situation can occur on days without sunlight or with overcast weather and rain. If caused by reason (iii), normally dissolved oxygen is increased by using paddlewheel aerator or aeration blower. In case of reason (ii) Water exchange is another common practice. As mentioned in [6], ‘‘dissolved oxygen level of incoming water can be enhanced if ripples are built into gravity inflow channels and water is injected into the ponds above water level”. An accurate predictor for dissolved oxygen in aquaculture ponds can act as a vital tool for the farmers as this can save them from huge loss (because of low growth or potential death of prawns). If farmers get alerts significant time ahead,
* Corresponding author. E-mail address:
[email protected] (A. Rahman). Peer review under responsibility of China Agricultural University. https://doi.org/10.1016/j.inpa.2019.08.002 2214-3173 Ó 2019 China Agricultural University. Production and hosting by Elsevier B.V. on behalf of KeAi. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Please cite this article as: A. Rahman, J. Dabrowski and J. McCulloch, Dissolved oxygen prediction in prawn ponds from a group of one step predictors, Information Processing in Agriculture, https://doi.org/10.1016/j.inpa.2019.08.002
2
Information Processing in Agriculture
they can consider appropriate action to help increase the DO level in the ponds. There exists a rich set of approaches to predict DO in aquaculture ponds. The current set of autoregressive approaches to DO prediction can be broadly grouped into three groups: (a) Fixed window approach that uses a fixed size window of past observations to predict future DO levels, (b) Variable window approach designed to predict one time stamp ahead only and use the predicted DO levels along with past observations to predict future DO levels, and (c) Recurrent neural network approach that maintains internal states for accurate prediction of future DO levels. Our experimental results reveal that existing approaches are suitable for short term (15 min) predictions while they struggle with long term (12 h) predictions. Short term predictions do not provide farmers adequate time to take preventative measures (like usage of paddlewheel aerator or aeration blower). We are thus motivated to design a DO prediction approach that is suitable for long term DO prediction. We observe that one step ahead prediction achieves maximum accuracy and it decreases as we predict multiple time stamp ahead. We have thus designed a group prediction approach where each predictor in the group is designed to predict a certain time stamp ahead. Combined the group of predictors can predict a time series of multiple time stamps ahead prediction. Each predictor is designed in a way so that in predicts one step ahead only but in disguise it is predicting several time stamps ahead. Because it’s predicting one step ahead only, it produces highly accurate prediction. At the same time, it’s producing several time stamps ahead (in disguise) thus serving the purpose of long-term prediction. The approach is detailed in Section 3. We have validated our proposed method on two prawn ponds from Bribie Island in Queensland Tasmania (Fig. 5). Historical time series data of dissolved oxygen from the two ponds were used to train and test the proposed prediction methods. Experimental results demonstrate considerable improvement over previous methods for long term predictions. Note that the proposed machine learning based prediction method is generic and can be trained and tested on any pond that has historical time series dissolved oxygen data available. The model parameters need to be tuned for the time series data of the specific pond. We restrict ourselves to the data of the two ponds from Bribie Island that we have access to and permission to publish. Also note that we confine ourselves to the approaches that are autoregressive in nature. Autoregression is a time series model that uses observations from previous time stamps as input to a regression equation to predict the value at the next time stamps. Given that DO oscillates to different peaks following a diurnal pattern and have some form of seasonality, such approaches are ideal for prediction [7].
2.
Background and current approaches
Several Machine Learning (ML) methods can be observed in the literature [8–26] for predicting DO. The most commonly used underlying ML methods across the literature include: Linear Regression (LR) [8–10], Neural Network (NN) Regression [11–18], and Support Vector Machine (SVM) Regression [19– 21]. We briefly explain the working methods of all these three
xxx (xxxx) xxx
approaches as follows. In all these modelling approaches, a window of past DO observations are considered as input to the model and future DO levels in upcoming time stamps are predicted by the model.
Linear Regression: In linear regression, DO is modelled as a weighted linear combination of its past observations. The weights are learned from historical data using either correlation or least square approximation methods. Neural Network Regression: Neural network assumes a nonlinear relationship between current DO level and its previous observations. A traditional neural network consists of a set of layers. The first layer is the input layer where the window of past DO observations are fed. The final layer is the prediction layer that produces future DO levels. The intermediate layers are called hidden layers. The hidden layers transform the input data in a non-linear fashion using activation functions. The series of non-linear transformations maps the input window (i.e. past DO observations) into future DO levels. The historical data is used to train the model parameters (called weights) using a method called back propagation. Different variants of neural network [11–18] are observed in the literature for predicting DO. Support Vector Regression: Support Vector methods also assume a nonlinear relationship like neural network. In this method, the input window of past DO observations are first transformed into higher dimension using non-linear kernel transformation functions. The transformed data is then mapped to the target future DO levels using a linear regression method. This is a combination of linear and nonlinear transformation methods depending on the nature of kernels used. Historical data is used to obtain the model parameters based on some optimization method. Another important categorization among prediction approaches exist based on how the multiple step ahead prediction levels are produced. Our contribution in this paper is a new approach to produce multiple step ahead prediction. Such approaches can be broadly classified into three groups as presented next. Any of the above machine learning approaches can be tied with the following approaches. A summary of these approaches is presented in Table 1.
Fixed Window Approach: In the first set of approaches, a fixed window of past observations acts as an input to the model. The output of the model is a number indicating the predicted value a certain time stamps ahead. A set of p models are developed for predicting p time stamps ahead prediction. Note that same input variables are used in all the models (Fig. 1). We use the term ‘Fixed Window’ approach to refer to this first group. Variable Window Approach: The second group of approaches are based on manipulating the input vector to the model (Fig. 2). The input to the model is a vector of past observa-
Please cite this article as: A. Rahman, J. Dabrowski and J. McCulloch, Dissolved oxygen prediction in prawn ponds from a group of one step predictors, Information Processing in Agriculture, https://doi.org/10.1016/j.inpa.2019.08.002
Information Processing in Agriculture
xxx (xxxx) xxx
3
Table 1 – Terms used for different multiple time stamp ahead prediction approaches. Terms
Approach
One Step to n Stamps
Proposed Method of group of predictors where each predictor predicts a step ahead but combined predicts multiple time stamps ahead. Conventional approach where a predictor predicts multiple time stamps ahead. A single predictor is used to predict a specific time stamp ahead sample. Same lag window is used by all the predictors. A recurrent approach where one predictor is designed for one time stamp ahead prediction. The prediction from step p is fed back as the last observation for the prediction of step p + 1 A recurrent neural network structure where only one predictor is designed to predict multiple time stamps ahead in one go.
Fixed Window
Variable Window
Recurrent Neural Network
Fig. 1 – Fixed Window approach to predict multiple time steps ahead using separate models.
Fig. 2 – Variable Window approach to predict multiple time stamps ahead using one model.
Fig. 3 – RNN approach to predict multiple time stamps ahead using one model.
tions. The output is set to predict only one time stamp ahead. For predicting 2 future time stamps ahead, the input vector is modified with the prediction made from time stamp one (Fig. 2) and this process is repeated for future time stamps. In general, for predicting p future time stamps ahead, earlier predictions from time stamps 1; 2; ; p 1 becomes part of the input vector. We use the term ‘Variable Window’ approach to refer to the second group.
the model is set to a vector of past observations (Fig. 3). The output of the model is the vector of p future time stamps. The model thus predicts p future time stamps in one go. This being a recurrent neural network, the internal states of the model are fed back and acts as input to the model. The feedback loop keeps track of the history of prediction and helps making better predictions. We use the RNN approach to refer to this group.
Recurrent Neural Network Approach: The third approach is based on Recurrent Neural Network (RNN). The input to
From experiments we have observed that prediction error increases as we predict further steps ahead. Step
Please cite this article as: A. Rahman, J. Dabrowski and J. McCulloch, Dissolved oxygen prediction in prawn ponds from a group of one step predictors, Information Processing in Agriculture, https://doi.org/10.1016/j.inpa.2019.08.002
4
Information Processing in Agriculture
ahead prediction normally ends up with smallest error. We are thus motivated to develop a classifier that imitates on step ahead prediction but predicts multiple time stamp ahead in disguise. The following section details the proposed approach.
3.
Proposed approach
In this section we present the philosophy and framework for the proposed ensemble prediction method. Let x1 ¼ x1 ð1Þ; x1 ð2Þ; ; x1 ðnÞ be a time series and we aim to make multiple time stamp ahead prediction on x1 . We train a step ahead predictor h1 on x1 . One step in x1 is equivalent to one time stamp in x1 . Hence one step ahead prediction from h1 will generate one time stamp ahead prediction for x1. Now
xxx (xxxx) xxx
let’s sample x1 at second time stamp such that x2 ¼ x1 ð1Þ; x1 ð1 þ 2Þ; x1 ð1 þ 2 2Þ; . We now train a step ahead predictor h2 on x2 . Note that one step in x2 is equivalent to two time stamps in x1 . Hence one step ahead prediction from h2 will generate two time stamps ahead prediction for x1 . Following this analogy if we compose a time series data xp by sampling x1 at p–th time stamp such that xp ¼ x1 ð1Þ; x1 ð1 þ pÞ; x1 ð1 þ 2 pÞ; and train a predictor hp on xp , a step ahead prediction from hp will generate p time stamps ahead prediction for x1. We followed this philosophy to generate p time stamps ahead prediction be concatenating the predictions from the predictors h1 ; h2 ; ; hp . The framework is presented in Fig. 4. We use the term ‘One Step to n Stamps’ to refer to the proposed approach. All the abbreviations of current and proposed approaches are presented in Table 1.
Fig. 4 – A framework of the proposed multiple time stamp ahead prediction method. To make a prediction of p time stamps ahead, a modified time series data xp is composed by concatenating samples that are p time stamps apart in the original time seriesx1 . A step ahead predictor is trained on xp . A step ahead prediction from xp constitutes p time stamps ahead prediction for x1. li past observations are considered for making step ahead prediction in time series xi where 1 i p. For ease of demonstration li ¼ 3 is assumed in this figure.
Fig. 5 – Satellite image of the two ponds used in the experiments. Please cite this article as: A. Rahman, J. Dabrowski and J. McCulloch, Dissolved oxygen prediction in prawn ponds from a group of one step predictors, Information Processing in Agriculture, https://doi.org/10.1016/j.inpa.2019.08.002
Information Processing in Agriculture
xxx (xxxx) xxx
5
Fig. 6 – Time series data over the whole data collection period from the two ponds studied in this paper.
4. Table 2 – Error comparison between different learning algorithms as part of the proposed approach.
Pond 1 Pond 2
MAE RMSE MAE RMSE
SVM
NN
LR
0.3703 0.5887 0.1155 0.1673
0.4664 0.6683 0.1333 0.1869
0.5437 0.7436 0.1258 0.1794
Experimental setup
The dataset used in this study comprises a set of DO readings from two aquaculture prawn ponds in Queensland, Australia. Pond 1 was large 0.18 ha grow-out pond and Pond 2 was a small 0.022 ha nursery pond (Fig. 5). The large pond is an outdoor pond and the nursery pond is a covered pond (not a tank). Both ponds were dug ponds, about 1.8 m deep, and have a plastic lining. The ponds are filled to about 30–40 cm from the pond lip. Tiger prawn was cultivated in both ponds.
Fig. 7 – Average prediction error at different lags for Pond 1. The average error across the 50 time-stamp-ahead predictions were computed for each lag.
Please cite this article as: A. Rahman, J. Dabrowski and J. McCulloch, Dissolved oxygen prediction in prawn ponds from a group of one step predictors, Information Processing in Agriculture, https://doi.org/10.1016/j.inpa.2019.08.002
6
Information Processing in Agriculture
A YSI EXO2 Multiparameter Sonde sensor [27] was placed in the two ponds to acquire DO readings at 15-minute intervals. The sensor was placed at approximately 1 m depth from the water surface. Readings are automatically sampled and transmitted to CSIRO’s Senaps data platform [28]. The sensor was placed in Pond 1 from 2nd Dec 2016 to 28th Feb 2017 and in Pond 2 from 2nd March 2017 to 3rd June 2017. Occasionally,
xxx (xxxx) xxx
the sensor was removed from the ponds for cleaning. Cleaning events introduce anomalous readings and missing data. Owing to the higher volume of water, algae, and animals in the larger grow-out pond (Pond 1), the DO varies over a wider range than that in the smaller nursery pond (Pond 2). Furthermore, the grow-out pond has a more irregular stochastic component than that of the nursery pond.
Fig. 8 – Average prediction error at different lags for Pond 2. The average prediction error across the 50 time-stamp-ahead predictions were computed for each lag.
Fig. 9 – Prediction error for Pond 1.
Fig. 10 – Prediction error for Pond 2. Please cite this article as: A. Rahman, J. Dabrowski and J. McCulloch, Dissolved oxygen prediction in prawn ponds from a group of one step predictors, Information Processing in Agriculture, https://doi.org/10.1016/j.inpa.2019.08.002
Information Processing in Agriculture
We have used around 70% of the data for training and remaining 30% for testing. The RNN experiments were conducted in Python. Experiments on all the remaining existing approaches and the proposed approach were conducted in MATLAB. SVM was used as the base predictor in MATLAB and we used the libSVM implementation [29]. As data was collected at 15 min intervals, each time stamp in the graphs presented in the remaining figures in the paper correspond to 15 min. We presented prediction results up to 50 time stamps ahead and that is equivalent to bit more than 12 h ahead prediction.
xxx (xxxx) xxx
7
We have used Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) to evaluate the accuracy of the predictions. Given a set of targets y1 ; y2 ; ; yN and corresponding P predictions p1 ; p2 ; ; pN , MAE is computed as N1 Ni¼1 yi pi h P 2 i12 . and RMSE is computed as N1 Ni¼1 yi pi
5.
Results and discussion
Fig. 6 presents the time series dissolved oxygen data from the two ponds we studied in this research. It can be observed the
Fig. 11 – Time series rediction results on the test data of Pond 1 using the proposed method. The red line indicates the minimum level of 5 mg/L DO to be maintained in the ponds. The proposed method was able to predict DO levels bellow 5 mg/ L even at 50 time stamps ahead. Please cite this article as: A. Rahman, J. Dabrowski and J. McCulloch, Dissolved oxygen prediction in prawn ponds from a group of one step predictors, Information Processing in Agriculture, https://doi.org/10.1016/j.inpa.2019.08.002
8
Information Processing in Agriculture
diurnal pattern of DO i.e. the higher peaks represent maximum DO during daytime and lower peaks represent minimum DO during night time. It can be observed that the lowest DO levels dropped below 5 in Pond 1 on several occasions whereas the lowest DO levels in Pond 2 never crossed that lowest threshold. We will observe how these different patterns are predicted by the proposed and existing methods in following sections. Note that, during the data collection process we never reached a situation where DO levels were critically low (bellow 2 mg/L). Hence, we are unable to assess the effectiveness of the approaches to predict critically low DO values in this paper. We have investigated the performance of different linear and nonlinear learning algorithms as part of the proposed
xxx (xxxx) xxx
prediction framework. Table 2 presents the performance of SVM, NN, and LR. Average MAE and RMSE error across 50 time stamps ahead prediction is presented here. It can be observed that SVM obtains lowest fitting error (both MAE and RMSE) for both ponds. We thus used SVM for analysis in the following sections. We also conducted some experiments to obtain the best lag window size for the proposed approach. As the proposed approach samples from past observations, a large window size w will lead to the requirement of w d past observations if we are forecasting d time stamps ahead. A small window size is thus preferable to reduce the requirement of high volume of past observations. Fig. 7 and Fig. 8 presents the prediction errors as a function of lag window size for Pond 1 and
Fig. 12 – Time series prediction results on the test data of Pond 2 using the proposed method. We never reached a situation in this pond where the DO level went bellow 5 mg/L – the minimum level to be maintained. Please cite this article as: A. Rahman, J. Dabrowski and J. McCulloch, Dissolved oxygen prediction in prawn ponds from a group of one step predictors, Information Processing in Agriculture, https://doi.org/10.1016/j.inpa.2019.08.002
Information Processing in Agriculture
xxx (xxxx) xxx
9
Fig. 13 – Prediction error comparison for Pond 1.
Fig. 14 – Prediction error comparison for Pond 2. Please cite this article as: A. Rahman, J. Dabrowski and J. McCulloch, Dissolved oxygen prediction in prawn ponds from a group of one step predictors, Information Processing in Agriculture, https://doi.org/10.1016/j.inpa.2019.08.002
10
Information Processing in Agriculture
Pond 2 respectively. Each bar in the graphs represents the average error computed from 50 days ahead prediction. Looking at the trend, we consider a lag window size of 6 to be optimum for Pond 1 and same of size 10 for Pond 2. Fig. 9 and Fig. 10 present the average prediction error on the two ponds at multiple days ahead. Note that the prediction error is relatively higher when predicting at a higher day ahead. This is consistent with any prediction method. Note that prediction error is relatively lower at Pond 1 compared to Pond 2. Fig. 11 and Fig. 12 presents the prediction results on the test data of Pond 1 and Pond 2 at multiple time steps ahead. Note that prediction aligns very well when predicting one time step ahead for both ponds. However, the predictions deviate from actual values when predictions are made 25 and 50 time stamps ahead leading to higher prediction error. This is consistent with the error graphs presented in Fig. 9 and Fig. 10. It is important to note that the low DO levels in Pond 1 (below 5 mg/L) was successfully predicted by the proposed method even at 50 times ahead. Fig. 13 and Fig. 14 presents the prediction error for Pond 1 and Pond 2 when predictions are made multiple time steps ahead using proposed and some of the existing methods (presented in Section 2). Note that prediction error is higher for all methods when predictions are made at higher time steps ahead. At smaller days ahead predictions, the error from proposed and existing method are very similar. It can be observed from Table 3 and Table 4 that the 1 time stamp ahead error (both MAE and RMSE) are very similar for all approaches. However, at higher time stamps ahead predictions, the error from proposed method is significantly lower than the other methods. As can be observed from Table 3 and Table 4, the difference of error between 50 time stamps ahead prediction and 1 time stamp ahead prediction for the proposed method is much lower compared to the other exist-
xxx (xxxx) xxx
ing approaches. As the proposed method makes one step ahead prediction for any time-stamp-ahead prediction, the error is low. Also, prediction error made on Pond 2 was lower than that in Pond 1. This can be attributed to the higher variability of DO levels in Pond 1 compared to Pond 2. There are a number of observations on the proposed ‘One Step to n Stamps’ approach. (i) In order to predict up to p time stamps ahead, we need to train a total of p predictors. This can be an issue if data is of high resolution and that means p can be a very large number if far ahead predictions are needed. That will require a large number of models; (ii) Another issue of the proposed approach is the requirement of past observations proportional to the size of the prediction length. If we are using a lag window of size l and making p time stamp ahead prediction, the lag sample size will be l p. This can be a large number and will require a large number of past observations to make the prediction.
6.
Conclusion
The paper presents an approach that can predict dissolved oxygen multiple time stamps ahead with relatively low error compared to other contemporary approaches. The key trick is to design a step ahead predictor where a step is equivalent to multiple time stamps. As step ahead prediction errors are normally lower than multiple step ahead predictions, the proposed approach obtains lower error compared to other methods even at multiple time steps ahead. An issue is that the proposed approach requires a higher window of past observations and a large number of models compared to some other existing approaches. This may limit the number of time steps ahead prediction it can make. In future we aim to work on reducing the lag window size and number of models in the proposed approach. Also, inclusion of past observations of other variables like temperature, salinity, wind speed, solar
Table 3 – Error comparison between alternative approaches for Pond 1. MAE
One Step to n Stamps Fixed Window RNN Variable Window
RMSE
1 Time Stamp Ahead
50 Times Stamps Ahead
Difference
1 Time Stamp Ahead
50 Times Stamps Ahead
Difference
0.074 0.082 0.068 0.086
0.521 1.06 1.784 1.156
0.447 0.978 1.716 1.070
0.053 0.053 0.168 0.052
0.217 0.785 0.703 0.509
0.164 0.732 0.535 0.457
Table 4 – Error comparison between alternative approaches for Pond 2. MAE
One Step to n Stamps Fixed Window RNN Variable Window
RMSE
1 Time Stamp Ahead
50 Times Stamps Ahead
Difference
1 Time Stamp Ahead
50 Times Stamps Ahead
Difference
0.035 0.035 0.028 0.035
0.147 0.661 0.493 0.363
0.112 0.626 0.465 0.328
0.053 0.053 0.168 0.052
0.217 0.785 0.703 0.509
0.164 0.732 0.535 0.457
Please cite this article as: A. Rahman, J. Dabrowski and J. McCulloch, Dissolved oxygen prediction in prawn ponds from a group of one step predictors, Information Processing in Agriculture, https://doi.org/10.1016/j.inpa.2019.08.002
Information Processing in Agriculture
radiation etc. are likely to improve the prediction accuracy of DO levels. We aim to consider them in the future while this paper focussed on autoregressive approaches only.
Declaration of Competing Interest All authors are from CSIRO Australia. Reviewers from CSIRO should be avoided.
R E F E R E N C E S
[1] Felix SS. Advances in shrimp aquaculture management. India: Daya Publishing House; 2009. [2] Robertson CE. Australian prawn farming manual: Health management for profit; 2006. link: http://era.daf.qld.gov.au/ id/eprint/2055/. [3] Boyd CE. Guidelines for aquaculture effluent management at the farm-level. Aquaculture 2003;226(1–4):101–12. [4] Ferreira NC, Bonetti V, Seiffert WQ. Hydrological and Water Quality Indices as management tools in marine shrimp culture. Aquaculture 2011;318(3–4):425–33. [5] Boyd CE, Tucker CS. Pond aquaculture water quality management. Boston: Kluwer Academic Publishers; 1998. [6] FAO. GROW-OUT PHASE; 2019. link: http://www.fao.org/3/ y4100e/y4100e08.htm. [7] Han JH. Comparing Models for Time Series Analysis; 2018. link: https://repository.upenn.edu/wharton_research_ scholars/162. [8] Khan UT, Valeo C. A new fuzzy linear regression approach for dissolved oxygen prediction. Hydrol Sci J 2005;60(6):1096–119. [9] Abba SI, Hadi SJ, Abdullahi J. River water modelling prediction using multi-linear regression, artificial neural network, and adaptive neuro-fuzzy inference system techniques. Procedia Comput Sci 2017;120:75–82. [10] Miao X, Deng C, Li X, Gao Y, He D. A hybrid neural network and genetic algorithm model for predicting dissolved oxygen in an aquaculture pond. In: Proc. International Conference on Web Information Systems and Mining (WISM), Sanya. p. 415–9. [11] Khan UT, Valeo C. Optimising fuzzy neural network architecture for dissolved oxygen prediction and risk analysis. Water 2017;9(6):381. [12] Chen Y, Yu H, Cheng Y, Cheng Q, Li D. A hybrid intelligent method for three-dimensional short-term prediction of dissolved oxygen content in aquaculture. PLoS One 2018;13 (2). [13] Huanand J, Liu X. Dissolved oxygen prediction in water based on K-means clustering and ELM neural network for aquaculture. Trans Chin Soc Agric Eng 2016;32(17):174–81. [14] Xiao Z, Peng L, Chen Y, Liu H, Wang J, Nie Y. The dissolved oxygen prediction method based on neural network. Complexity 2017.
xxx (xxxx) xxx
11
[15] Zhang Y, Fitch P, Vilas MP, Thorburn PJ. Applying multi-layer artificial neural network and mutual information to the prediction of trends in dissolved oxygen. Front Environ Sci 2019;7:46. [16] Ta X, Wei Y. Research on a dissolved oxygen prediction method for recirculating aquaculture systems based on a convolution neural network. Comput Electron Agric 2018;145:302–10. [17] Ren Q, Zhang L, Wei Y, Li D. A method for predicting dissolved oxygen in aquaculture water in an aquaponics system. Comput Electron Agric 2018;151:384–91. [18] Ahmed AM. Prediction of dissolved oxygen in surma river by biochemical oxygen demand and chemical oxygen demand using the artificial neural networks (anns). J King Saud Univ – Eng Sci 2017;29:151–8. [19] Malek S, Mosleh M, Syed SM. Dissolved oxygen prediction using support vector machine. Int Scholar Sci Res Innovat 2018;8(1):46–50. [20] Tarmizi A, Ahmed A, El-Shafie A. Dissolved oxygen prediction using support vector machine in terengganu river. Middle-East J Sci Res 2014;21(11):2182–8. [21] Ji X, Shang X, Dahlgren RA. Prediction of dissolved oxygen concentration in hypoxic river systems using support vector machine: a case study of Wen-Rui Tang River, China. Environ Sci Pollut Res 2017;24(19):16062–76. [22] Shi P, Li G, Yuan Y, Huang G, Kuang L. Prediction of dissolved oxygen content in aquaculture using clustering-based softplus extreme learning machine. Comput Electron Agric 2019;157:329–38. [23] Xu L, Liu S, Li D. Prediction of water temperature in prawn cultures based on a mechanism model optimized by an improved artificial bee colony. Comput Electron Agric 2017;140:397–408. [24] Olyaie E, Abyaneh HZ, Mehr AD. A comparative analysis among computational intelligence techniques for dissolved oxygen prediction in delaware river. Geosci Front 2017;8:517–27. [25] Dabrowski JJ, Rahman A, George A, Arnold S, McCulloch J. State space models for forecasting water quality variables: an application in aquaculture prawn farming. In: Proc. 24th ACM SIGKDD international conference on knowledge discovery data mining (KDD) London, UK. p. 177–85. [26] Dabrowski JJ, Rahman A, George A. Prediction of dissolved oxygen from ph and water temperature in aquaculture prawn ponds. In: Proc. of the Australasian joint conference on artificial intelligence – Workshop MLSDA, Wellington, New Zealand. p. 2–6. [27] YSI, EXO2 Multiparameter Sonde; 2018. link: https://www. ysi.com/EXO2. [28] Biggins D, Coombe M, Hugo D, McCulloch J, Neumeyer P, Pasanen J, et al. A platform for integrating time-series with modelling systems. Proc MODSIM 2017, Hobart, Tasmania, 2017. [29] Chang CC, Lin CJ. LIBSVM: a library for support vector machines. ACM Trans Intell Syst Technol 2011.
Please cite this article as: A. Rahman, J. Dabrowski and J. McCulloch, Dissolved oxygen prediction in prawn ponds from a group of one step predictors, Information Processing in Agriculture, https://doi.org/10.1016/j.inpa.2019.08.002