Received 27 February 2023, accepted 14 March 2023, date of publication 21 March 2023, date of current version 28 March 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3260089

A Novel Hybrid Model to Predict Dissolved

Oxygen for Efficient Water Quality
in Intensive Aquaculture
WENJUN LIU1,2,3,4,5 , SHUANGYIN LIU 1,2,3,4 , SHAHBAZ GUL HASSAN1,2,3,4 ,
YINGYING CAO1,2,3,4,5 , LONGQIN XU1,2,3,4 , DACHUN FENG1,2,3,4 , LIANG CAO1,2,3,4 ,
WEIJUN CHEN6 , YAOCONG CHEN1,2,3,4 , JIANJUN GUO 1,2,3,4 ,
TONGLAI LIU 1,2,3,4 , AND HANG ZHANG5
1 Guangzhou Key Laboratory of Agricultural Product Quality, Safety Traceability Information Technology, Zhongkai University of Agriculture and Engineering,
Guangzhou 510225, China
2 Academy of Smart Agricultural Engineering Innovations, Zhongkai University of Agriculture and Engineering, Guangzhou 510225, China
3 Smart Agriculture Engineering Technology Research Center of Guangdong Higher Education Institues, Zhongkai University of Agriculture and Engineering,

Guangzhou 510225, China

4 Guangdong Provincial Agricultural Products Safety Big Data Engineering Technology Research Center, Zhongkai University of Agriculture and Engineering,

Guangzhou 510225, China

5 College of Computer and Information Engineering, Tianjin Agricultural University, Tianjin 30000, China
6 Hengxing Intelligent Agriculture (Guangzhou) Development Co., Ltd., Guangzhou 511453, China

Corresponding authors: Tonglai Liu ([email protected]) and Hang Zhang ([email protected])

This work was supported in part by the National Natural Science Foundation of China under Grant 61871475; in part by the Special Project
of Laboratory Construction of Guangzhou Innovation Platform Construction Plan under Grant 201905010006; in part by the Guangzhou
Key Research and Development Project under Grant 202103000033 and Grant 201903010043; in part by the Guangdong Science and
Technology Project under Grant 2020A1414050060 and Grant 2020B0202080002; in part by the Innovation Team Project of Universities
in Guangdong Province under Grant 2021KCXTD019; in part by the Characteristic Innovation Project of Universities in Guangdong
Province under Grant KA190578826; in part by the Guangdong Province Enterprise Science and Technology Commissioner Project under
Grant GDKTP2021004400; in part by the Guangdong Science and Technology Planning Project under Grant 2016A020210122 and Grant
2015A040405014; in part by the Meizhou City S&T Planned Projects under Grant 2021A0305010; in part by the Rural Science and
Technology Correspondent Project of Zengcheng, Guangzhou, under Grant 2021B42121631; in part by the Educational Science Planning
Project of Guangdong Province under Grant 2020GXJK102 and Grant 2018GXJK072; in part by the Guangdong Province Graduate
Education Innovation Program Project under Grant 2022XSLT056 and Grant 2022JGXM115; in part by the Guangdong Provincial
Graduate Education Innovation Plan Project under Grant 2022XSLT056; in part by the Basic Ability Enhancement Project for Young and
Middle-Aged Teachers of Universities in Guangxi under Grant 2021KY0799; in part by the Technical Service Project for Xingning Pigeon
Industrial Park of the Zhongkai University of Agriculture and Engineering (Construction and Promotion of Visual Information Platform);
in part by the National Innovation Industry Training Program for College Students in 2022 under Grant 202211347023; in part by the
National Innovation Industry Training Program for College Students in 2021 under Grant 202111347004; in part by the Guangdong
Innovation Industry Training Program for College Students in 2022 under Grant S202211347083 in part by the Key Construction
Discipline Research Ability Enhancement Project of Guangdong Province in 2022 under Grant 2022ZDJS022, in part by the Natural
Science Foundation of Guangdong Province in 2023 under Grant 2023A1515011230, in part by the Science and Technology Planning
Project in Yunfu under Grant 2022020303, and in part by the Science and Technology Program of Guangzhou under Grant 2023E04J0037.

ABSTRACT Dissolved oxygen content is a key indicator of water quality in aquaculture environment.
Because of its nonlinearity, dynamics, and complexity, which makes traditional methods face challenges
in the accuracy and speed of dissolved oxygen content prediction. As a solution to these issues, this
study introduces a hybrid model consisting of the Light Gradient Boosting Machine (LightGBM) and the
Bidirectional Simple Recurrent Unit (BiSRU). Firstly, Linear interpolation and smoothing were used to
identify significant parameters. LightGBM algorithm then determines the significance of dissolved oxygen
by eliminating irrelevant variables and predicting dissolved oxygen in intensive aquaculture. Finally, the
attention method was implemented to map the weighting and learning parameter matrices, so enabling the
BiSRU’s hidden states to be assigned different weights. The findings shown that the presented prediction
model can accurately anticipate the fluctuating trend of dissolved oxygen over a 10-day period in just
122 seconds, and the accuracy rate reached 96.28%. Comparing the model effects of LightGBM-BiSRU,

The associate editor coordinating the review of this manuscript and

approving it for publication was Yiqi Liu .

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

29162 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 11, 2023
W. Liu et al.: Novel Hybrid Model to Predict Dissolved Oxygen for Efficient Water Quality in Intensive Aquaculture

LightGBM - GRU, LightGBM-LSTM, and BiSRU - Attention takes the least time. Its higher prediction accuracy
can provide an essential reference for intensive aquaculture water quality regulation.

INDEX TERMS Non-linear, LightGBM, BiSRU, attention mechanism.

I. INTRODUCTION
In terms of aquaculture production, China dominates the models can predict dissolved oxygen content at future
global stage; the nation is responsible for over 70% of moments, they are still inadequate in terms of speed of com-
the world’s total. In 2021, China’s aquaculture output was putation and the capture of global contextual information,
53.88 million tons, accounting for 80.50% of the total aquatic make it challenging to fulfill the demand of accurate and
product [1]. Because aquatic organisms perform a range of fruitful aquaculture production.
physiological activities in water, such as respiration, feeding, The Bidirectional Simple Recurrent Unit (BiSRU) is a
excretion and reproduction, the quality and production of two-layer network structure with reverse stacking, which can
aquatic goods are directly influenced by the water quality acquire the past and the future information but also has a
of an aquatic environment [2]. Dissolved oxygen (DO) is a highly parallelized architecture. This network is capable of
key indication of water quality since it is essential to the sur- not only sequence modeling, but also improves the gradient
vival of aquatic animals and is used by their metabolism [3]. disappearance problem, making BiSRU widely used in many
Excessive or insufficient DO can affect the healthy growth fields. For example, Jie used a prediction method based on
of farmed fish, shrimp, and other organisms, easily resulting BiSRU to achieve high accuracy prediction for intrusion
in disease outbreaks and even mass mortality, which would detection of industrial control systems [14]. Ding et al. used
result in significant economic losses for business [4], [5]. For an intrusion detection model combining CNN and BiSRU to
this reason, predicting dissolved oxygen concentrations and achieve accurate prediction of network intrusion [15]. Ding
their trends in advance, regulating dissolved oxygen concen- proposed an effective model for network security protection
trations in a timely manner and ensuring healthy growth of using BiSRU in conjunction with feature reduction for iden-
aquatic products in a comfortable environment are important tifying anomalous traffic [16].
for preventing water quality deterioration, reducing the risk In Machine learning, data is processed using attention
of aquaculture and the healthy and sustainable development mechanism, which is used to determine contribution size
of intensive aquaculture [6]. between input and out data, making it applicable to a variety
Many findings have been made as a consequence of exten- of disciplines. For example, Jiang proposed a combined
sive study on dissolved oxygen prediction models utilizing LSTM, transformer and attention mechanism for indoor
machine learning conducted by academics both domestically temperature prediction model, which achieves accurate and
and overseas. For example, Liu introduced a forecasting efficient prediction of room temperature trends [17]. Zhang
model that integrates grey correlation degree, empirical integrated transformer model and multiple attention mech-
wavelet transformations, and the particle swarm optimization anism to develop an attention network framework based on
gravity search technique. Experiments show that this model Transformer Encoder, which effectively achieved accurate
can more accurately analyze the trend of dissolved oxygen prediction of stock trends [18]. Mei presented a hybrid model
[7]; Ren used Variational Mode Decomposition (VMD) to based on CNN, GRU and attention mechanisms, in which
segregate and denoise the original data before feeding the different neuron weights can be adjusted by the attention layer
decomposed data into a Deep Belief Network (DBN) for pre- to achieve accurate prediction of water quality [19]. Li pro-
diction [8]. Cao utilized principal component analysis (PCA) posed a dissolved oxygen prediction model combining stack
to filter the essential elements that would affect the dissolved structure, multi-attention mechanism and TCN, which can
oxygen, then K-means clustering and GRU were employed effectively improve the prediction accuracy of water quality
to create the dissolved oxygen prediction model [9]; Shi parameters in Marine pastures and bring positive influence
suggested a dissolved oxygen prediction model based on clus- to the development of Marine fisheries [20]. Duan achieved
tering and an enhanced extreme learning technique [10] and effective prediction of tool wear status using a hybrid atten-
Huang proposed hybrid model based on CEEMDAN-LZC tion based on parallel deep learning [21]. Therefore, the
and GOBLPSO to improve efficiency [11]. A novel model drawbacks of conventional approaches may be addressed by
based on feature extraction was proposed, which improved including an attention mechanism into the model to concen-
predictive performance and provide accurate predictions for trate on data that is more important to the present job among
dissolved oxygen levels [12]. Nong proposes a dissolved the numerous inputs by adaptively learning the proportion of
oxygen prediction model based on support vector regression weights.
combined with multi-feature engineering and optimisation In intensive aquaculture, there are complex biochemical
methods, and the implementation shows that the model can reactions in the water column and complex mechanisms
effectively improve the accuracy of the prediction model [13]. of action between the factors. The accuracy and computa-
Although the above-mentioned dissolved oxygen prediction tional complexity of dissolved oxygen prediction for water

VOLUME 11, 2023 29163

 W. Liu et al.: Novel Hybrid Model to Predict Dissolved Oxygen for Efficient Water Quality in Intensive Aquaculture

FIGURE 1. Raw data of water quality parameters of intensive aquaculture.

quality parameters in intensive aquaculture suffer by feeding which leads to a complex prediction model network structure
the all factors into the dissolved oxygen prediction model, and redundancy or overlapping information. LightGBM is

29164 VOLUME 11, 2023

W. Liu et al.: Novel Hybrid Model to Predict Dissolved Oxygen for Efficient Water Quality in Intensive Aquaculture

FIGURE 2. Raw data of water quality parameters of intensive aquaculture.

a tree-based boosting algorithm that can not only process significantly improve the speed of the algorithm in large sam-
data efficiently but also reduce memory consumption and ple applications. Wang et al. constructed LightGBM model

VOLUME 11, 2023 29165

 W. Liu et al.: Novel Hybrid Model to Predict Dissolved Oxygen for Efficient Water Quality in Intensive Aquaculture

FIGURE 3. Internet of Things based real time monitoring system.

FIGURE 4. Histogram algorithm.

FIGURE 5. Growth of leaf-wise tree-based decision tree learning process.

for corporate financing risk prediction, which can effectively prediction model, effectively providing decision support for
improve the accuracy of risk prediction [22]. Sun et al. investors to invest in cryptocurrency [23]. Ren constructed
adopted LightGBM to build a cryptocurrency price trend a CNN-LSTM-LightGBM based attention mechanism for

29166 VOLUME 11, 2023

W. Liu et al.: Novel Hybrid Model to Predict Dissolved Oxygen for Efficient Water Quality in Intensive Aquaculture

FIGURE 6. Detailed structure of SRU hidden layer.

short-term wind power forecasting model, which achieves the smoothing method as shown in eq 1
accurate prediction [24].
xm−1 + xm+1
This study proposed a hybrid model consist of a LightGBM xm = (1)
(Light Gradient Boosting Machine), Bidirectional Simple 2
Recurrent Unit (BiSRU), and Attention mechanism to over- where xm−1 and xm+1 are value at the m − 1 and m + 1 times,
come the limitations of traditional approaches for dissolved respectively, and xm is the abnormal data at the m-th time.
oxygen prediction. The LightGBM was used to identify the Figures 1 and 2 demonstrate noise reduction is decreased
significant parameters affecting the dissolved oxygen con- successfully monitoring data for intensive aquaculture.
centration in intensive aquaculture. A nonlinear hybrid model
was proposed by simplifying the network architecture using a B. EXPERIMENTAL SETUP
bidirectional simple loop unit and an attention mechanism to The experimental environment include: processor i7-11800H,
predict of dissolved oxygen. The experimental results showed CPU frequency 2.3GHz, memory 16.0GB, Windows 10
that the forecasting results of this model can provide technical (64-bit), python3.8 (64-bit), integrated development environ-
support for the accurate control of water quality parameters ment Anaconda3.
in intensive aquaculture. LightGBM initial parameters are: num_leaves are 1000,
learning_rate is 0.2, feature_fraction is 0.8, bagging_fraction
is 0.8, max-bin is 800, and boosting_type is gbdt.
II. MATERIALS AND METHODS
A. DATASET AND PRE-PROCESSING C. MODEL PERFORMANCE METRICS
From December 7, 2020 to January 18, 2021, water quality To evaluate the performance of the dissolved oxygen forecast
sample data were collected for this research, which comprised model for intensive aquaculture water quality parameters,
five water quality indicators and a total of 6050 observation Root Mean Square Error (RMSE), Mean Absolute Error
samples. Testing and training and testing dataset curve can be (MAE) and Mean Square Error (MSE) were selected as error
seen in Figure 1. evaluation indicators. The higher the value of these indicators
This section may be divided by subheadings. It should converges to zero, the higher the prediction accuracy. The
provide a concise and precise description of the experimental higher the value of the coefficient of determination R2 , the
results, their interpretation, as well as the experimental con- better the model fits the data. The specific formulae are shown
clusions that can be drawn. in equations (2)-(5).
As a consequence of environmental factors during the data r
collection process, such as sensor aging or surface contami- 1 XN  p 2
ERMSE = yi − ŷi (2)
nation, there is noise interference in the acquired data. Noise N i=1
reduction processing must be applied to the original signal of 1 XN
2
EMSE = yi − ŷi (3)
the monitored water quality data in order to reduce noise and N i=1
recover the actual signal as show in Figure 2. In the process 1 X N
EMAE = yi − ŷi (4)
of noise reduction, the abnormal data is filled with the mean N i=1

VOLUME 11, 2023 29167

 W. Liu et al.: Novel Hybrid Model to Predict Dissolved Oxygen for Efficient Water Quality in Intensive Aquaculture

FIGURE 7. Structure of LightGBM-BiSRU-Attention network model.

PN
2
yi − ŷi TABLE 1. Feature importance of water quality parameters.
R2 = 1 − Pi=1
N
(5)
i=1 (yi − ȳi )2

where: is yi the true value ȳ of sample i ; is the mean value of

the sequence of true values; is ŷi the model predicted value of
sample i ; N is the number of test sample sets.

D. DATA ACQUISITION
In this paper, a pond with a depth of 1.5 m and an area of
200 m2 was selected in an aquaculture Dongchong Town,
Nansha District, Guangzhou in China. The pond equipped
modern aquaculture IOT based system for real time moni- exclude them from selection, avoiding significant additional
toring and data collection which includes: temperature, dis-
memory consumption [26].
solved oxygen, ammonia nitrogen oxygen sensors, ultraviolet
filters, aerators, and biological towers as shown in Figure 3.
F. SIMPLE RECURRENT UNIT (SRU)
Data was collected after every 10 minutes and store it on the
server for real-time viewing on the terminal through the IoT Lei proposed Simple Recurrent Unit (SRU) to simplify
architecture. parameters, parallel processing and increase the speed of
model runs, and has been widely used in many fields [27].
When compared to conventional recurrent and acyclic
E. LIGHT GRADIENT BOOSTING MACHINE (LIGHTGBM) designs, the SRU performs better since its gate computations
LightGBM uses a histogram-based algorithm and a leaf-by- are not dependent on those of prior time steps. The SRU
leaf tree structure, which can effectively improve computa- architecture is shown in Figure 6.
tional efficiency and reduce memory consumption [25] as Both the ‘‘forget gate’’ and the ‘‘memory unit’’ are essen-
show in in Figure 4. In XGBoost, the tree is grown using tial components of the SRU. In order to fine-tune the memory
the level-wise growth strategy, and the level-wise algorithm device, forget gate was employed, which symbolizes for the
searches and splits the leaves at each level, which has the significance of the preceding step to the current state. The
problem of extra consumption and leads to inefficiency. For memory component unit performs the calculations necessary
this reason, LightGBM is optimized with a leaf-wise algo- to determine the final output state. In general, the parameter
rithm that performs a deep leaf-by-leaf search and split, formulation of a single-layer SRU model is as follows:
as shown in Figure 5. The depth-based leaf-wise algorithm
can efficiently find nodes with small information gain and x̃t = W xxt (6)

29168 VOLUME 11, 2023

W. Liu et al.: Novel Hybrid Model to Predict Dissolved Oxygen for Efficient Water Quality in Intensive Aquaculture

TABLE 2. Error comparison of models.

FIGURE 8. Proposed LightGBM-BiSRU-Attention model-based DO prediction.

FIGURE 9. Proposed LightGBM-BiSRU-Attention model comparison.

ft = σ (Wf xxt + bf ) (7) processes the actual sequence front to back rather than super-
rt = σ (Wr xxt + br ) (8) imposing a forward and reverse SRU on the input signal
and connecting the two SRUs toward an output unit. The
Ct = ft xCt−1 + (1 − ft ) xx̃t (9)
traditional time series model is prone to problems such as
ht = rt xg (Ct ) + (1 − rt ) xxt (10) gradient disappearance and poor training effectiveness [28].
xt is current moment input; x̃t is linear transformation at the Ling proposed a detection approach based on BiSRU to
current moment, ft is the forget gate, rt is the reset gate, ct and successfully address the aforementioned issues [29]. This
ht are memory unit and hidden layer state. wr and wf are reset model uses parallel computing to speed up model training,
and forget gate, respectively. makes each time step independent of the one before it,
employs skip connections to solve the gradient disappearance
G. BIDIRECTIONAL SIMPLE RECURRENT UNIT (BISRU) problem, and improves information capture about the time
The BiSRU model improves on the standard time series series’ characteristics with respect to its positive and negative
model. The main concept is that the traditional model bi-directional structure. BiSRU is an improved version of

VOLUME 11, 2023 29169

 W. Liu et al.: Novel Hybrid Model to Predict Dissolved Oxygen for Efficient Water Quality in Intensive Aquaculture

FIGURE 10. Comparison of BiSRU-Attention and LightGBM-BiSRU-Attention models.

FIGURE 11. Comparison of LightGBM-BiSRUand LightGBM-BiSRU-Attention models.

FIGURE 12. Comparison of LightGBM-BiSRU, LightGBM-LSTM and LightGBM-GRU models.

BiGRU that retains the modeling capabilities while using less recognition [34] using attention mechanism. By calculating
computation (and hyperparameters) [30]. the correlation between the data, the attention mechanism
gives different weights to different feature data. This makes
H. ATTENTION MECHANISM it easier to find useful information in the input data and the
Bahdanau et al. presented the Attention mechanism to target output than in the original data, brings out the most
overcome the issue that input data information cannot be important features related to the prediction, and improves the
completely collected, which simulates the resource allocation quality of the output.
mechanism of human brain attention [31]. Attention, ignor- In the attention mechanism structure, Xt (t ∈ [1, n]) is the
ing irrelevant information while amplifying desired infor- input to the BiSRU network, ht (t ∈ [1, n]) represents to the
mation [32]. Exceptional achievements have been obtained hidden layer output obtained from each input BiSRU, and αt
in the domains of machine translation [33] and voice (t ∈ [1, n]) is the attention probability distribution value of

29170 VOLUME 11, 2023

W. Liu et al.: Novel Hybrid Model to Predict Dissolved Oxygen for Efficient Water Quality in Intensive Aquaculture

FIGURE 13. Critical indexes-based model comparison.

the BiSRU hidden layer output by the Attention mechanism, III. RESULTS AND DISCUSSION
and y is the final output value of the Attention mechanism. Dissolved oxygen is influenced by various factors and has
a complex mechanism of action. Suppose all parameters
I. THE HYBRID FORECASTING MODEL BASED ON are directly input into the prediction model. In that case,
LIGHTGBM - BISRU - ATTENTION it will easily lead to a complex model network structure and
LightGBM-BiSRU-Attention are organically integrated to high computational complexity. Therefore, it is necessary to
construct a non-linear dissolved oxygen model for intensive eliminate the multiple covariances, extract influencing key
aquaculture, which improves the performance of the dis- factors, and optimize the prediction model network structure
solved oxygen prediction model. The basic ideas are: firstly, to improve the prediction performance. LightGBM has an
the dissolved oxygen data of intensive aquaculture are pre- ability to screening key influencing factors. Therefore, this
processed, secondly, the structure of the BiSRU prediction paper selected LightGBM to screen the key influencing fac-
model was simplified by using LightGBM to screen for key tors of dissolved oxygen and obtained the contribution of
influences on water quality parameters dissolved oxygen. each feature, as shown in Table 1. Among them, the first
Subsequently, the prediction model of BiSRU-Attention was three influencing factors have a greater impact on dissolved
trained and predicted to obtain the final prediction results, oxygen, so the study screened the key influencing factors are
which effectively improved the prediction accuracy of the acidity, conductivity, and turbidity, which were also unan-
combined model. Proposed non-linear hybrid prediction imously approved by experts in the field of aquaculture,
model is shown in Figure 7, and the specific steps are as to construct the training and testing data set for dissolved
follows. oxygen prediction model.
Step 1: Collect water quality parameter dissolved oxygen The results of Proposed LightGBM-BiSRU-Attention
time series data online through the Intensive Aquaculture model are shown in figure 8. The proposed LightGBM-
Internet of Things (IoT) cloud platform and pre-process them. BiSRU-Attention hybrid model is an innovative and useful
Step 2: Screen of key influences on dissolved oxygen technique for dissolved oxygen prediction, since it has a
for water quality parameters in intensive aquaculture using robust learning capacity and achieves high generalization
LightGBM, eliminating multivariate redundant information performance.
and streamlining the prediction model structure. The conventional LightGBM-GRU, LightGBM-LSTM,
Step 3: The split samples are the training and test datasets, BiSRU-Attention, and LightGBM-BiSRU were chosen for
and the best BiSRU-Attention prediction model is obtained comparison analysis based on the identical original data to
by feeding the training dataset. analyze the performance of the LightGBM-BiSRU-Attention.
Step 4: Testing BiSRU-Attention and proposed LighGM- Figure 9 and Table 2 respectively provide the fitting curve
BISRU-Attention on data set by comparing the result to verify comparison graphs and the prediction performance compari-
the result for proposed model prediction. son data of their five models.

VOLUME 11, 2023 29171

 W. Liu et al.: Novel Hybrid Model to Predict Dissolved Oxygen for Efficient Water Quality in Intensive Aquaculture

The non-linear hybrid LightGBM-BiSRU-Attention model the LightGBM-BiSRU-Attention model, which incorporates
proposed in this paper was compared with the standard all of the components of LightGBM, BiSRU, and Attention,
LightGBM-GRU model, LightGBM-LSTM model, BiSRU- is superior for serial prediction of the water quality parameter
Attentionmodel and LightGBM-BiSRU model, which shows dissolved oxygen in high-density aquaculture.
the better prediction effect.
Figure 10 illustrates the BiSRU-Attention model’s predic- IV. CONCLUSION AND FUTURE WORK
tion findings, which were obtained without applying the key This study contributes as follows:
factor screening method and trained using data on intensive
1. The data is preprocessed and filled using linear inter-
aquaculture water quality, had anticipated values that dif-
polation, and smoothing corrects anomalous data to
fered considerably from the actual measured values. Table 2
improve prediction accuracy.
shows that when comparing LightGBM-BiSRU-Attention
2. LightGBM was used to assess the degree of other water
with BiSRU-Attention under the identical circumstances,
quality parameters on dissolved oxygen, taking into
the evaluation metrics MSE, MAE, and RMSE dropped by
consideration their correlation strength.
0.02%, 0.39%, and 0.38%. While R2 rose 2.69%. It is con-
3. A hybrid dissolved oxygen prediction model (Light-
firmed that the underlying model’s performance and predic-
GBM - BiSRU - Attention ) is proposed. The bidirec-
tion accuracy can be significantly enhanced by the LightGBM
tional structure (BiSRU bidirectional simple recurrent
approach employed in this article.
unit) is used to transform the upcoming information
Figure 11 shows that the LightGBM-BiSRU-Attention
into current time point prediction, select the key points
model provides a better match than the LightGBM-BiSRU
via the attention mechanism which enables accurate
model when comparing anticipated outcomes. Table 2
prediction of dissolved oxygen.
shows that when comparing LightGBM-BiSRU-Attention to
LightGBM-BiSRU, the evaluation metrics MSE,MAE, and A new hybrid model (LightGBM-BiSRU-Attention) is
RMSE all improve by 0.03%, 0.45%, and 0.48%, respec- proposed in this paper to address the issues of slow
tively, while R2 improves by 2.76%. This exemplifies how operation speed, complex network structure, and insuf-
incorporating an attention mechanism into a model may ficient capture of global contextual information for dis-
improve its output quality and boost its overall performance, solved oxygen prediction of non-linear and non-smooth
bringing with it time and cost savings. The model is slower intensive aquaculture water quality parameters directly
than the LightGBM-BiSRU model but quicker than the by traditional dissolved oxygen prediction methods. The
comparative models because of the attention mechanism’s LightGBM-BiSRU-Attention model was constructed by
inclusion in the whole. integrating the three LightGBM, BiSRU, and Attention
In Figure 12, we can see that the LightGBM-BiSRU-based approaches. The results clearly show that when RMSE, MAE,
model provides the best fit, while the LightGBM-LSTM- MSE, and R2 are used, the proposed method outperforms
based model deviates the most from the actual value. Table 2 LightGBM-GRU, LightGBM-LSTM, BiSRU-Attention, and
shows that both LightGBM-GRU and LightGBM-LSTM LightGBM-BiSRU. In current intensive aquaculture, the
improved upon the baseline in terms of MSE (0.19%), MAE hybrid technique of LightGBM-BiSRU-Attention has higher
(1.35%), RMSE (1.44%), and T (465.0 s), whereas R2 predictive performance and is an excellent predictive method
(13.78%) was higher in both cases. Based on these results, for predicting dissolved oxygen time series.
it seems that GRU is superior than LSTM in terms of pre- This study has certain limitations and requires more inves-
diction accuracy even when using a smaller sample size. tigation. In the future, it is planned to investigate advanced
Due to the fact that LSTM has three distinct gate struc- algorithms such as: the bat algorithm, particle swarm opti-
tures whereas GRU only contains two, LightGBM-LSTM mization algorithm, and swarm spider optimization, can
has extra gate than the LightGBM-GRU, along with a more be combined with BiSRU for more accurate and efficient
complicated structure and additional parameters. Therefore, prediction of dissolved oxygen levels and further improve
the LightGBM-GRU-based prediction model outperforms the prediction capability. All figures and tables should be cited
LSTM-based in terms of accuracy and processing speed. The in the main text as Figure 3, Table 1, etc.
error indicators as: MSE, MAE, RMSE, and time to com-
plete a task are all improved by 0.22%, 2.54%, 2.45%, and ACKNOWLEDGMENT
257 seconds, respectively, using lightGBM-BiSRU, while R2 (Wenjun Liu, Shuangyin Liu, and Shahbaz Gul Hassan con-
is increased by 16.26%. BiSRU’s improved performance in tributed equally to this work.)
dissolved oxygen prediction may be attributed to its bidirec-
tional structure, which collects all relevant data, as well as its REFERENCES
greater memory capacity and enhanced time series prediction [1] F. Hu, ‘‘Development of fisheries in China,’’ Reprod. Breeding, vol. 1,
capabilities when compared to GRU. no. 1, pp. 64–79, 2021.
The proposed model has the minimum error, as seen more [2] S. Ayesha Jasmin, P. Ramesh, and M. Tanveer, ‘‘An intelligent frame-
work for prediction and forecasting of dissolved oxygen level and biofloc
clearly in Figure 13’s three-dimensional representation of the amount in a shrimp culture system using machine learning techniques,’’
assessment metrics for the prediction models. As a whole, Expert Syst. Appl., vol. 199, Aug. 2022, Art. no. 117160.

29172 VOLUME 11, 2023

W. Liu et al.: Novel Hybrid Model to Predict Dissolved Oxygen for Efficient Water Quality in Intensive Aquaculture

[3] M. H. Ahmed and L.-S. Lin, ‘‘Dissolved oxygen concentration predic- [24] J. Ren, Z. Yu, G. Gao, G. Yu, and J. Yu, ‘‘A CNN-LSTM-LightGBM based
tions for running waters with different land use land cover using a quan- short-term wind power prediction method based on attention mechanism,’’
tile regression forest machine learning technique,’’ J. Hydrol., vol. 597, Energy Rep., vol. 8, pp. 437–443, Aug. 2022.
Jun. 2021, Art. no. 126213. [25] W. Liang, S. Luo, G. Zhao, and H. Wu, ‘‘Predicting hard rock pillar stability
[4] R. Dehghani, H. Torabi Poudeh, and Z. Izadi, ‘‘Dissolved oxygen concen- using GBDT, XGBoost, and LightGBM algorithms,’’ Mathematics, vol. 8,
tration predictions for running waters with using hybrid machine learning no. 5, p. 765, May 2020.
techniques,’’ Model. Earth Syst. Environ., vol. 8, no. 2, pp. 2599–2613, [26] Y. Li and W. Chen, ‘‘A comparative performance assessment of ensem-
Jun. 2022. ble learning for credit scoring,’’ Mathematics, vol. 8, no. 10, p. 1756,
[5] X. Cao, N. Ren, G. Tian, Y. Fan, and Q. Duan, ‘‘A three-dimensional Oct. 2020.
prediction method of dissolved oxygen in pond culture based on [27] T. Lei, Y. Zhang, S. I. Wang, H. Dai, and Y. Artzi, ‘‘Simple recurrent units
attention-GRU-GBRT,’’ Comput. Electron. Agricult., vol. 181, Feb. 2021, for highly parallelizable recurrence,’’ 2017, arXiv:1709.02755.
Art. no. 105955. [28] Y. Li, Y. Yang, K. Zhu, and J. Zhang, ‘‘Clothing sale forecasting by
[6] W. Li, H. Wu, N. Zhu, Y. Jiang, J. Tan, and Y. Guo, ‘‘Prediction of a composite GRU–Prophet model with an attention mechanism,’’ IEEE
dissolved oxygen in a fishery pond based on gated recurrent unit (GRU),’’ Trans. Ind. Informat., vol. 17, no. 12, pp. 8335–8344, Dec. 2021.
Inf. Process. Agricult., vol. 8, no. 1, pp. 185–193, Mar. 2021. [29] J. Liu, C. Yu, Z. Hu, Y. Zhao, Y. Bai, M. Xie, and J. Luo, ‘‘Accurate
[7] H. Liu, R. Yang, Z. Duan, and H. Wu, ‘‘A hybrid neural network model for prediction scheme of water quality in smart mariculture with deep Bi-S-
marine dissolved oxygen concentrations time-series forecasting based on SRU learning network,’’ IEEE Access, vol. 8, pp. 24784–24798, 2020.
multi-factor analysis and a multi-model ensemble,’’ Engineering, vol. 7, [30] Z. Chen, Z. Hu, L. Xu, Y. Zhao, and X. Zhou, ‘‘DA-Bi-SRU for water
no. 12, pp. 1751–1765, Dec. 2021. quality prediction in smart mariculture,’’ Comput. Electron. Agricult.,
[8] Q. Ren, X. Wang, W. Li, Y. Wei, and D. An, ‘‘Research of dissolved vol. 200, Sep. 2022, Art. no. 107219.
oxygen prediction in recirculating aquaculture systems based on deep [31] D. Bahdanau, K. Cho, and Y. Bengio, ‘‘Neural machine translation by
belief network,’’ Aquacultural Eng., vol. 90, Aug. 2020, Art. no. 102085. jointly learning to align and translate,’’ 2014, arXiv:1409.0473.
[32] Y. Zhen, J. Fang, X. Zhao, J. Ge, and Y. Xiao, ‘‘Temporal convolution
[9] X. Cao, Y. Liu, J. Wang, C. Liu, and Q. Duan, ‘‘Prediction of dissolved
network based on attention mechanism for well production prediction,’’
oxygen in pond culture water based on K-means clustering and gated
J. Petroleum Sci. Eng., vol. 218, Nov. 2022, Art. no. 111043.
recurrent unit neural network,’’ Aquacultural Eng., vol. 91, Nov. 2020,
[33] F. Wang, W. Chen, Z. Yang, S. Xu, and B. Xu, ‘‘Hybrid attention for
Art. no. 102122.
Chinese character-level neural machine translation,’’ Neurocomputing,
[10] P. Shi, G. Li, Y. Yuan, G. Huang, and L. Kuang, ‘‘Prediction of dissolved
vol. 358, pp. 44–52, Sep. 2019.
oxygen content in aquaculture using clustering-based softplus extreme
[34] Q. Chen and G. Huang, ‘‘A novel dual attention-based BLSTM with hybrid
learning machine,’’ Comput. Electron. Agricult., vol. 157, pp. 329–338,
features in speech emotion recognition,’’ Eng. Appl. Artif. Intell., vol. 102,
Feb. 2019.
Jun. 2021, Art. no. 104277.
[11] J. Huang, S. Liu, S. G. Hassan, L. Xu, and C. Huang, ‘‘A hybrid model
for short-term dissolved oxygen content prediction,’’ Comput. Electron.
Agricult., vol. 186, Jul. 2021, Art. no. 106216.
[12] W. Cao, J. Huan, C. Liu, Y. Qin, and F. Wu, ‘‘A combined model of
dissolved oxygen prediction in the pond based on multiple-factor analysis
and multi-scale feature extraction,’’ Aquacultural Eng., vol. 84, pp. 50–59, WENJUN LIU received the bachelor’s degree
Feb. 2019. in computer and information engineering from
[13] X. Nong, C. Lai, L. Chen, D. Shao, C. Zhang, and J. Liang, ‘‘Prediction Tianjin Agricultural College, in 2021, where she
modelling framework comparative analysis of dissolved oxygen concen- is currently pursuing the master’s degree in elec-
tration variations using support vector regression coupled with multiple tronic information. Her research interests include
feature engineering and optimization methods: A case study in China,’’ machine learning and time-series prediction.
Ecol. Indicators, vol. 146, Feb. 2023, Art. no. 109845.
[14] J. Ling, Z. Zhu, Y. Luo, and H. Wang, ‘‘An intrusion detection method for
industrial control systems based on bidirectional simple recurrent unit,’’
Comput. Electr. Eng., vol. 91, May 2021, Art. no. 107049.
[15] S. Ding, Y. Wang, and L. Kou, ‘‘Network intrusion detection based on
BiSRU and CNN,’’ in Proc. IEEE 18th Int. Conf. Mobile Ad Hoc Smart
Syst. (MASS), Oct. 2021, pp. 145–147.
[16] P. Ding, J. Li, M. Wen, L. Wang, and H. Li, ‘‘Efficient BiSRU combined SHUANGYIN LIU graduated from China
with feature dimensionality reduction for abnormal traffic detection,’’ Agricultural University, in 2014. He worked with
IEEE Access, vol. 8, pp. 164414–164427, 2020.
the Zhongkai University of Agriculture and Engi-
[17] B. Jiang, H. Gong, H. Qin, and M. Zhu, ‘‘Attention-LSTM architecture neering. He is currently a Professor. His research
combined with Bayesian hyperparameter optimization for indoor temper-
interests include machine learning and big data
ature prediction,’’ Building Environ., vol. 224, Oct. 2022, Art. no. 109536.
processing.
[18] Q. Zhang, C. Qin, Y. Zhang, F. Bao, C. Zhang, and P. Liu, ‘‘Transformer-
based attention network for stock movement prediction,’’ Expert Syst.
Appl., vol. 202, Sep. 2022, Art. no. 117239.
[19] P. Mei, M. Li, Q. Zhang, G. Li, and L. Song, ‘‘Prediction model of drinking
water source quality with potential industrial-agricultural pollution based
on CNN-GRU-attention,’’ J. Hydrol., vol. 610, Jul. 2022, Art. no. 127934.
[20] D. Li, X. Zhang, Y. Yang, H. Yang, and S. Liu, ‘‘An interpretable hierar-
chical neural network insight for long-term water quality forecast: A study
in marine ranches of eastern China,’’ Ecol. Indicators, vol. 146, Feb. 2023, SHAHBAZ GUL HASSAN received the Ph.D.
Art. no. 109771. degree from the College of Information and Elec-
[21] J. Duan, X. Zhang, and T. Shi, ‘‘A hybrid attention-based paralleled deep trical Engineering, China Agricultural Univer-
learning model for tool wear prediction,’’ Expert Syst. Appl., vol. 211, sity, in 2017. He is currently a Lecturer with
Jan. 2023, Art. no. 118548. the College of Information Science and Tech-
[22] D.-N. Wang, L. Li, and D. Zhao, ‘‘Corporate finance risk prediction based nology, Zhongkai University of Agriculture and
on LightGBM,’’ Inf. Sci., vol. 602, pp. 259–268, Jul. 2022. Engineering. His current research interests include
[23] X. Sun, M. Liu, and Z. Sima, ‘‘A novel cryptocurrency price trend fore- intelligent information systems of agriculture and
casting model based on LightGBM,’’ Finance Res. Lett., vol. 32, Jan. 2020, artificial intelligence.
Art. no. 101084.

VOLUME 11, 2023 29173

 W. Liu et al.: Novel Hybrid Model to Predict Dissolved Oxygen for Efficient Water Quality in Intensive Aquaculture

YINGYING CAO received the bachelor’s degree YAOCONG CHEN received the bachelor’s degree
in computer and information engineering from in information science and technology from the
Tianjin Agricultural College, in 2021, where she Zhongkai University of Agriculture and Engi-
is currently pursuing the master’s degree in elec- neering, in 2020, where he is currently pursu-
tronic information. Her research interests include ing the master’s degree in agricultural engineering
computer vision. and information technology. His research interest
includes time-series forecasting.

LONGQIN XU received the M.S. degree from the

Faculty of Computers, Guangdong University of
Technology, in 2006. She is currently a Professor
with the College of Information Science and Tech- JIANJUN GUO was born in Handan, China,
nology, Zhongkai University of Agriculture and in 1982. He received the B.E. degree in optical
Engineering. Her main research interests include engineering from the South China Normal Univer-
intelligent information systems for agriculture, sity, Guangzhou, China, in 2012, the M.S. degree
artificial intelligence, and machine learning. in measurement and control technology and instru-
mentation from the Hebei University of Science
and Technology, Shijiazhuang, China, in 2006, and
the Ph.D. degree in optics from the South China
Normal University, Guangzhou, in 2016. He is
currently an Associate Professor with the College
of Information Science and Technology, Zhongkai University of Agriculture
and Engineering, Guangzhou. His current main research interests include
DACHUN FENG was born in Nanchong, China, agricultural automation and information, artificial intelligence, big data ana-
in 1973. He received the Ph.D. degree from the lytics, and mining.
South China University of Technology, China,
in 2009. He is currently a Professor with the
College of Information Science and Technology,
Zhongkai University of Agriculture and Engi-
neering. His current research interests include
intelligent information systems for agriculture, the
Internet of Things, artificial intelligence, and big
data.

TONGLAI LIU received the B.E. and M.E. degrees

from the Guilin University of Electronic Technol-
ogy, China, in 2007 and 2010, respectively, and
the Ph.D. degree from the Guangdong University
of Technology, in 2021. After the M.E. degree,
LIANG CAO received the M.S. degree in com- he joined the Guilin University of Electronic Tech-
puter technology engineering from Sun Yat-sen nology. He is currently an Associate Professor with
University, Guangzhou, China, in 2008. He is the Zhongkai University of Agriculture and Engi-
currently an Engineer with the Zhongkai Univer- neering. His current research interests include data
sity of Agriculture and Engineering, Guangzhou. mining, blockchain technology, edge computing,
His recent research interests include computer and database.
technology, the Internet of Things technology, and
applications.

WEIJUN CHEN is currently pursuing the Ph.D. HANG ZHANG received the Ph.D. degree in
degree with Jimei University. He is currently agricultural information technology from China
the Deputy General Manager of the Hengxing Agricultural University, Beijing, China, in 2016.
Intelligent Agriculture Development (Guangzhou) He is currently a Lecturer with the College of
Company. He is responsible for the formulation of Computer and Information, Tianjin Agricultural
special fish and shrimp feeds. He has published University, Tianjin, China. His research interests
many articles in core journals at home and abroad, include image processing and pattern recognition.
including three SCI articles in the past three years.
He is a member of the fifth batch of the Fujian
Province Hundred People Introduction Plan.

29174 VOLUME 11, 2023