Available online at www.sciencedirect.

com

Procedia Engineering 31 (2012) 1194 – 1199

International Conference on Advances in Computational Modeling and Simulation

Prediction of water quality time series data based on least

squares support vector machine
Guohua Tan*, Jianzhuo Yan, Chen Gao, Suhua Yang
College of Electronic Information and Control Engineering, Beijing University of Technology, Beijing ,China

Abstract

Actual water quality monitoring sites due to inadequate and abnormal by the impact to water quality
warning has brought new challenges. How do based on limited monitoring data to predict the water
quality to address the shortage of river water quality monitoring sites and data coverage of false alarms
caused by abnormal, early warning of water pollution are of great significance. According to river
water monitoring data, the small sample properties, is proposed based on least squares support vector
machine prediction of water quality parameters, this method has strong ability to predict the true
value, and the global optimization and good generalization. This method is applied in the river water
quality measurement data, after training the LS-SVM model for water quality parameters of water quality
monitoring system to predict, in the same sample under the BP network and RBF network prediction.
Experimental results show that the small sample case with noise, least squares support vector machine
method is better than multi-layer BP and RBF neural network, to better meets the requirements of water
quality prediction. Comparing data of the two experiments shows that new module makes the interest
mining more effective.

© 2011 Published by Elsevier Ltd. Selection and/or peer-review under responsibility of Kunming
University of Science and Technology Open access under CC BY-NC-ND license.

Keywords: Support Vector Machines, Time series data forecast, Small sample, Water Quality Monitoring,

*Corresponding author. Tel:+86-15210835051

Email address: [email protected]

1877-7058 © 2011 Published by Elsevier Ltd. Open access under CC BY-NC-ND license.
doi:10.1016/j.proeng.2012.01.1162
 Guohua Tan et al. / Procedia Engineering 31 (2012) 1194 – 1199 1195

1. Introduction

Forecasting changes in water quality of water environment, planning, evaluation and management of
the foundation is to maintain and manage the current basis for water quality, water quality by predicting
the evolution of trends can be understood in order to find the reasons for deterioration of water quality,
and can guide the development of control measures.
But in the actual water quality control, due to the limited number of water quality monitoring sites,
resulting in lack of monitoring data, monitoring data also lead to abnormal environmental impact, which
bring more to the prediction of water quality challenges.
The current domestic small samples there are two categories of prediction methods:
(1)Prediction model based on neural networks[1~2], BP networks using nonlinear function
approximation theory to predict the sample.
(2)Using principal component analysis (PCA) to model the main features of extracted information,
high dimensional complex data dimension reduction, data reconstruction through the correct estimate of
the data.
In recent years, researchers at home and abroad to support vector machine classification and regression
were a large number of studies, such as literature [3~4] to support vector machines, artificial neural
networks have been studied, indicating that the support vector machine classification and regression have
a greater advantages, can solve the small sample, nonlinear, high dimension and local minimum points
and other practical issues, is considered a better alternative to artificial neural network algorithm.
In this paper, least squares support vector machine LS-SVM algorithm to construct non-linear time
series forecasting model to predict the water quality.

2. Forecasting methods and implementation process

2.1 Phase space reconstruction theory

Phase space reconstruction theory [5] is the basis for time series prediction, Packard and others and
Takens[6] delay coordinates proposed method using time series phase space reconstruction. Phase space
reconstruction theory is that the system is the evolution of any component interact with other components
of the decision, therefore, the evolution of each component in the system are implicit in all of the
information. When a state space reconstruction, only consider a component, and it is a fixed time delay in
some point as a new dimension processing.
In this paper, in order to reduce modeling errors, the first for the raw data processing and data zero-
mean normalization, and then according to Takens theory of phase space reconstruction, about one-
dimensional time series into a matrix form to obtain the relationship between data relations, and thus tap
into as much as possible the amount of information.

2.2 LS-SVM time series prediction algorithm

Suyken et proposed least squares support vector machines[7~9] in recent years, important results of
statistical learning theory is one, least squares support vector machine training process follows the
structural risk minimization principle, and the general vector machine has the computational complexity
compared to low, the advantages of high operation speed.
Least squares support vector machines time series prediction algorithm described as follows:
1196 Guohua Tan et al. / Procedia Engineering 31 (2012) 1194 – 1199

d
Training samples for linear regression problems, training data set ( xi , yi ), i 1, 2 « n, xi  R is the
i sample of the input mode, yi  R corresponding to the desired output samples.
Linear regression function:
y ( x) wT x  b  [1]
According to SRM criteria will be returned to the problem into the following constrained quadratic
optimization problem, it is the only optimal solution exists:
1 t J n 2 [2]
min J ( w, [ ) w w [
2 2
¦
i 1
i

Constraints:
 yi wT xi  b  [i , i 1, « , n  [3]
LS-SVM regression, with ad low-dimensional dh
space to high dimensional space (Hilbert space) [10]of
the nonlinear mapping M (˜) : R o R , the low-dimensional nonlinear regression into high-dimensional
space, linear regression, the definition T
of high-dimensional space for the inner product
K ( xi , x j )
operation K ( xi , x j ) (M ( xi )M ( x j )) , where ( xi , x j ) replacing the inner product operation , the
introduction of Lagrange function, the constrained optimization problem into the linear equation obtained
nonlinear LS-SVM regression model:
n
f ( x) ¦ a K ( x , x)  b
i i
i 1 [4]
Type in the Mercer kernel function is to satisfy the conditions can be any symmetric function, we use
the radial basis function RBF:
2
­ x ,x ½
° i j °
K ( xi , x j ) exp ® ¾
2V 2
¯° ¿°  [5]

2.3 Implementation process

Based on the above algorithm, this paper proposes using least squares support vector machine LS-
SVM (Least Squares Support Vector Machine) algorithm to construct non-linear time series forecasting
model to predict the water quality. The method mainly consists of the following modules completed, the
overall implementation framework shown in Figure 1.

9GSVRK 6XKZXKGZSKTZ 9KZ'RMUXOZNS 3UJKR *GZGLUXKIGYZ ,UXKIGYZ

*GZG *GZG 6GXGSKZKXY :XGOTOTM UVKXGZOUTY *GZG

Fig. 1 Implementation framework of predict the water quality

The above frame mainly has four steps, Pretreatment Data, Set Algorithm Parameters, Model
Training and Data forecast operations. The details are below.

2.3.1 Pretreatment Data

Construction of training samples under the conditions of small samples, to validate the model and
the real value of the predicted anti-noise performance, the training samples and test samples to make the
 Guohua Tan et al. / Procedia Engineering 31 (2012) 1194 – 1199 1197

following treatment:

(1)The experimental data were first pre-entered; select the data and the normalized, after
transformation, the data values in between 0 and 1.
(2)In the measurement data, the variance of 0.01 were added random white noise, the direct use of
these data with the noise form the training samples

2.3.2 Set Algorithm Parameters

LS-SVM model with the measured data to predict water quality, the use of radial basis function as
the RBF kernel function, the use of LSSVMLab matlab toolbox algorithm framework as the
basic package, parameter Gam=10, sig2=0.5, Type='function estimation'; with BP Neural network
model forecasts of water quality measurements, using a standard3-layer BP network, an input
node, an output node, 20 hidden nodes; with RBF neural network model forecasts of water
quality measurement data, using generalized regression neural Network GRNN (Generalized Regression
NN).

2.3.3 Model Training

Using trainlssvm function training the network, the regression coefficient and deviation b, so as to
arrive as in [4] shows the prediction model.

2.3.4 Data forecast operations

Trained by the previous LS-SVM, BP and RBF models predict water quality parameters of the
experimental data.

3. Experimental Results and Analysis

The LS-SVM algorithm is applied to a section of the river water quality monitoring data to predict the
total phosphorus, and with the BP (Back Propagation), RBF (Radial Basis Function) network method, a
comparative study of prediction.
3.1 Experimental Results
LS-SVM, BP network and RBF network training for the three models and error learning curve shown
in Figure 2, "Solid line" represents the total phosphorus in water quality monitoring parameters of the
actual value, "Dotted line" represents the predicted value, its deviation from the reflection of the
phosphorus measured deviation from the actual value of the degree of the graph shows the deviation from
the actual deviation value has no strict counterpart, because after the data preprocessing and phase space
reconstruction after the great changes of their range; eight output of the test sample and the relative error
of the forecast shown in Table 2 are 0.064 percent, 0.529 percent and 0.552 percent,experimental results
Table 2 below.
The total phosphorus in water quality monitoring parameters of the actual value, see Table 1.
Forecasting value comparison of LS-SVM,RBF and BP network, see Table 2.

Table 1 The total phosphorus in water quality monitoring parameters of the actual value Model.

Month 1 2 3 4 5 6 7 8 9 10
1198 Guohua Tan et al. / Procedia Engineering 31 (2012) 1194 – 1199

TP 0.0763 0.0312 0.0607 0.0380 0.1260 0.1152 0.1280 0.1172 0.1273 0.1233

Input of the three model simulation compared with the predicted value of the map shown in Figure 2.

The first chart is LS-SVM model, the second chart is BP network, the third chart is RBF network

Fig. 2 models the input and output value of the comparison chart

The output of the model are compared with the input value in three models obtained the best predictive
model output compared in Table 2.

Table 2 Forecasting value comparison of LS-SVM,RBF and BP network

0RGHO LS-SVM BPQHWZRUN RBF QHWZRUN

Month True Value/ Predictive Relative Predictive Relative Predictive Relative

(mg/L) Value/ Value/% value/ Value/% value/ Value/%
(mg/L) (mg/L) (mg/L)
1 0.0763 0.0686 0.10 0.0320 1.42 0.0297 1.39
2 0.0312 0.0268 0.14 0.0184 0.41 0.0180 0.42
3 0.0607 0.0570 0.06 0.0267 0.56 0.0358 0.41
4 0.0380 0.0372 0.02 0.0144 0.62 0.0307 0.19
5 0.1260 0.1121 0.11 0.0365 0.71 0.0705 0.44
6 0.1152 0.1082 0.06 0.0552 0.52 0.0023 1.02
7 0.1280 0.1280 0.00 0.0627 0.51 0.0601 0.53
8 0.1172 0.1136 0.03 0.0761 0.35 0.0726 0.38
9 0.1273 0.1234 0.03 0.1132 0.11 0.0865 0.32
10 0.1233 0.1085 0.12 0.1134 0.08 0.0715 0.42
Average ü ü 0.064 ü 0.529 ü 0.552
Value
3.2. Comparative Analysis
Comparison of experiment with LS-SVM can be seen to predict the time series based on the true
value of water quality measurement data, the data fit a good degree of good, better than the BP network
method and the RBF Network. This is mainly because the LS-SVM model can take full advantage of the
distribution of training samples, based on discriminate function constituted part of the training samples,
and the LS-SVM model processing ability of small samples, in the case of small samples can be well
Conduct statistical learning.

4. Conclusions

Based on the timing of water quality monitoring data with small samples and noise characteristics of
 Guohua Tan et al. / Procedia Engineering 31 (2012) 1194 – 1199 1199

the data presented method based on LS-SVM time series prediction model for water quality monitoring.
It will be by means of phase space reconstruction Reconstruction of time series data for the vector data,
obtained through the normal LS-SVM model to learn on small samples, and can produce good
performance of the prediction model, simulation results show that the LS-SVM based Water quality
prediction model, root mean square error and mean relative error than the BP network method, RBF
network method is much smaller, indicating that the LS-SVM method has a high prediction accuracy,
more applicable to real-time water quality data with small sample forecast.

Acknowledgements

The data of Beijing Water Authority, and the matlab experiment framework used LSSVMLab
matlab toolbox download from http://www.esat.kuleuven.be/sista/lssvmlab/,a special thanks here.

References

[1] Peng T M ˈ Hubele N F,Karady G G.Advancement in the application of neural networks for short-term load
forecasting.IEEE Trans on Power Systems,1992,7(1):250̚257.
[2] Park D C,El-Sharkawi M A et al.Electric load forecasting using an artificial neural networks.IEEE Trans on Power
Systems,1991,6(2)˖442̚448.
[3] J Tian Hanˈ13oüSuk YangˈJong Moon LeeˊA new condition monitoring and fault diagnosis system of induction
motors using artificial intelligence algorithms[J]ˊElectric Machines and Drivesˈ2005˖1967ü1974
[4] Ze DongˈPu ElanˈXiüChao YinˊSimulation study on sensor fault diagnoses of the temperature of the boiler highü
temperature part metal wall[J]ˊMachine Learning and Cyberneticsˈ2004ˈ5 (26ü 29)˖3003ü3008.
[5] Takens F. Detecting Strange Attractors in Turbulence [C]// Dynamical Systems and Turbulence, Lecture Notes in
Mathematics. Berlin: Springer-Verlag, 1981, 898: 366-381.
[6] Packard N H, Crutchfield J P, Farmer J D, et al. Geometry From a Time Series [J]. Physical Review Letters (S0031-9007),
1980, 45(9): 712-716.
[7] Vapnik V NˊAn overview of statistical learning theory[J].IEEE Trans. on Neural Networks,1999,10(5):988-999.
[8] Platt J.Fast training of support vector machine using sequential minimum optimization[C]ˊAdvance in Kernel Methods-
support Vector Learning, Cambridge, 1999:185-208.
[9] Suykens J A K,Lukas L,Vandewalle J.Sparse approximation using least squares support vector machine[C].IEEE Int.
Symposium on Circuit and Systems(ISCAS 2000),Geneva,Switzerland,2000.
[10] R. J. Duffin and A. C. Schaeffer, A class of no harmonic Fourier series, Trans. Amer. Math. Soc. 72 (1952), 341 - 366.