Integration Method of Local-global SVR and Parallel Time Variant PSO in Water Level Forecasting for Flood Early Warning System

Flood is one type of natural disaster that can’t be predicted, one of the main causes of flooding is the continuous rain (natural events). In terms of meteorology, the cause of flood is come from high rainfall and the high tide of the sea, resulting in increased the water level. Rainfall and water level analysis in each period, still not ab le to solve the existing problems. Therefore in this study, the proposed integration method of Parallel Time Variant PSO (PTVPSO) and Local-Global Support Vector Regression (SVR) is used to forecast water level. Implementation in this study combine SVR as regression method for forecast the water level, Local-Global concept take the role for the minimization for the computing time, while PTVPSO used in the SVR to obtain maximum performance and higher accurate result by optimize the parameters of SVR. Hopefully this system will be able to solve the existing problems for flood early warning system due to erratic weather.

flood detection by forecast the water level will be needed by the various parties to obtain a lot of weather conditions more quickly and accurately.
Forecasting is the estimation process based on past historical data to estimate in the future [8]. Various forecasting methods have been developed nowadays. The approach developed include stochastic and deterministic approach. Forecasting models mostly using a deterministic approach, which is a condition when t or t-1 and t+1, with a daily period [9]. The usual method for predicting linear data such as exponential smoothing models, moving average (MA) and autoregressive integrated moving average (ARIMA). While the methods used to forecast the data is non-linear threshold autoregressive (TAR), autoregressive conditional heteroskedasticity (ARCH), artificial neural networks (ANN) and Support Vector Regression (SVR) [8].
In the previous studies, using comparison of two methods: multivariate adaptive regression splines (MARS) with support vector regression (SVR) to predict the sales of IT products, SVR gives the root mean squared error (RMSE) which is smaller in the amount of 9:36 in sales Notebook, compared with MARS method that gives RMSE value of 293.3 [10]. In addition, according to research [11], a comparative study of the K -Means method and SVR prove that on 6 opinions sector use of the internet hotspot, SVR gives much better results in 5 sectors than K-Means.
In order to obtain optimal performance of SVR, many methods used to perform the optimization of SVR. Several methods of optimization variables are algorithms cross entropy (CE), bee colony, ant algorithm and particle swarm optimization (PSO). A study by the method of adaptive fuzzy k-nearest neighbor optimized by using parallel time variant particle swarm optimization for prediction of bankruptcy automatically shown to have higher accuracy, which amounted to 81.67% compared to the method FKNN in the amount of 78.75% [12], where PTVPSO approach aimed at search optimization coefficients in the calculation of SVR.
Mentioned in his research that the hybrid SVR-PSO can choose the appropriate input discriminatory features, reducing the average time execution and improve the accuracy of forecasting [13]. Therefore, in this study the proposed integration method of parallel time variant PSO (PTVPSO) and support vector regression (SVR) is used to forecast water level. Implementation in this study combine SVR as regression method for forecast the water level, while PTVPSO used in the SVR to obtain maximum performance and higher acc urate result. Hopefully this system will be able to solve the existing problems for flood detection due to erratic weather. The structure of this paper is as follows. In section 2, some important definitions are given. Section 3 describes the details of our improvement on the SVR-PTVPSO algorithm for forecasting water level. Section 4 presents some simulation results and compares the improved method. The final section provides conclusion.

Definitions and Preliminaries
Kernel method is a class of algorithms for analysis or pattern recognition, which elements are known best are the SVM. General assignment pattern analysis is to find and study the common types of relationships such as clusters, classification, and ranking. Kernel methods mapped data into a higher dimensional space with the hope that in the higher space, the dimensions of the data can be structured. The most frequent type of kernel used is Gaussian's Kernel, which is one implementation of Radial Basis Function kernel [14].
SVM has been efficiently applied for classification and regression [15]. In this paper we used SVM for regression (SVR) [16]. One of the simplest methods used to obtain the line or field of optimal hyperplane SVM algorithm is very complex sequential method developed by [17]. The following steps: and and (6) 4. Go back to the third step, until the condition IterMax or ( (| |) and (| |) ) Support vector, if the data has value ( ) 5. The forecast function is: 6. End PSO (Particle Swarm Optimization) is one of optimization techniques based on the metaphor of social interaction and communication as groups of birds or fish. This model will be simulated in space with specific dimensions with a number of iterations so that in each iteration, the particle's position will increasingly lead to the intended target (minimizing or maximizing functions). This is done until the maximum iteration or other termination criteria [18]. PSO is added by inertia weight and acceleration coefficient to improves performance. Each particle has position and velocity, and updates that in every iterating [19]. Below is a translation of PSO approach [12]: Where in the vector Pi represents the best positions of the particles prior to i which gives the best value, which is known as pbest. Vector is the vector particles Pg best among all the particles in the population, which is known as gbest. r 1 and r 2 are random numbers generated between the values [0,1]. Variables that exist in SVR methods used for forecasting water level, at -where to get this purpose, PTVPSO (Parallel Time Variant PSO) continuously coupled to dynamically conducts parameter optimization and feature selection simultaneously. Features obtained from optimizing the use as input into the SVR models applied in parallel so that a method that is optimized SVR. For the inertia weight, which is used to balance the local exploitation and global exploration, a large inertia weight used for global search, while small inertia weight is used for local search [20], the inertia weight is update according to following equation: Where w min and w max predefined as maximum weight and minimum weight which initialize before training models, in the other side t max is maximum iterations of PSO, and t is current iterations. Usually the value of w min is 0.4 and the value of w max is 0.9 because the value provide a good results [12]. This equation is also-known-as time varying inertia weight (TVIW). In the other hand, c 1 and c 2 are coefficients acceleration that has function for balancing the global exploration and local exploitation for better performance of PSO, time varying acceleration coefficient (TVAC) is update according to following equation: Where c 1f , c 1i , c 2f , and c 2i are constants, t max is maximum iterations of PSO, and t is current iterations. The value which is used for c 1i is 2.5, c 1f is 0.5, c 2i is 0.5 and c 2f is 2.5 because the value provide with a good results [12]. To check the errors from the forecasting method, it can be determined by calculating the difference between the original value of the water level data with the value of forecasted data. Common methods used in the measurement accuracy of forecasting, there are several kinds of method. Many researcher prefer to use MAPE (mean absolute percentage error) because MAPE is easy to understand, MAPE use percentage value to show the error value [21]: Besides MAPE we also use MAE (mean absolute error) to show the error value [22], the formula of MAE is according to following equation: Where ̂ is the forecasted data of water level and is the actual data of water level. n is the amount of predicted data.

LGSVR-PTVPSO Algorithm
In this section, we discuss about SVR and PTVPSO, how the performance of these two methods and how these two method are integrated each other, where the SVR work as a basic method for forecasting and PTVPSO work as optimization methods. First things to do is, initialize and calculate PTVPSO parameters such as c 1 , c 2 , r 1 , r 2 and w, and determine the range for SVR parameter. Train those parameters using SVR model, so we can calculate the fitness using training data in SVR method. We can determine pbest and gbest, to change the reference for the next iterations. The parameter to be optimized in PSO is λ, Ɛ, cLR and C. RBF   It can be said that for the training phase, the SVR methods are inside PTVPSO methods, but after we get the optimal parameters with the biggest fitness value for training data, we can use the parameters for testing phase. To do testing phase, we only use SVR method itself to forecast the water level. Kernel method has been used in this study is RBF Gaussian. Figure 1 shows the proposed method we used LGSVR-PTVPSO (Local Global SVR-PTVPSO), Global SVR can be interpreted in testing phase we use a large iterations to find the optimal parameters for SVR, in the other hand Local SVR can be interpreted in training phase we use a small amount of iterations to find the optimal parameters.

Experiment And Analysis
This section discussed the experiments have been performed as well as analysis of the experiment. Several trial test has been conducted for SVR features that used to optimize water level forecasting, where the variations of SVR feature are 3, 4 and 5 days feature. The data in Table 1 expressed in meters (m), we can clearly see to forecast the 1st day, it takes the value of F1 to F5 which is the actual value of the 5 days prior to day 1 which is amount of the Y, for forecasting the 2nd day. Actual value on the 1st day was used as the last feature or F5 to make the forecast, as well as for the next day. The visualization prediction experiment using 5 features or using 5 days in advance wi ll be presented in Table 1:  The second experiment of SVR feature is using four features or by using the value of 4 days prior to actual days to forecast the water level. Figure 2 is the best result of training and testing using 4 days features of SVR, the best trials that have been performed, obtained MAPE value which is amount to 0.0459 and MAE value which is amount to 0.177882 for training phase and MAPE value which is amount to 0.03708 and MAE value which is amount to 0.14336 for testing phase. The third experiment of SVR feature is using five features or by using the value of 5 days prior to actual days to forecast the water level. Figure 3 is the best result of training and testing using 5 days features of SVR, the best trials that have been performed, obtained MAPE value which is amount to 0.04706 and MAE value which is amount to 0.18088 for training phase and MAPE value which is amount to 0.039335 and MAE value which is amount to 0.15208 for testing phase. Table 2 describes the comparison of execution time and error rate some variation of features SVR used in this experiment, for training result the best value obtained from using 4 days features of SVR or using the value of 4 days prior to actual days to forecast, we can clearly see the data on the result of the prediction is very precise with the actual data on the training phase, while for the testing phase the best value obtained from using 3 days features of SVR or using the value of 3 day prior to actual days to forecast, it appears that the predicted result and the actual data of water level is very precise using 3 days features of SVR, while the fastest execution time of the experiment using 3 days features of SVR for training phase and 5 days features of SVR for testing phase, the factors that affect the speed of computation time/running time (RT) is the tabulated data used in this experiment, due to the data used in training phase more numerous compared with the data in the testing phase.  It appears when the testing phase and training phase on the forecasting process which obtain the best value when using 3 days until 4 days in advance, it does give an indication that the water level ranging from 3 to 4 days give the significant influence on the actual day, either the water level are going to increase or decrease, but when using 5 days in advance the data may be too far away to predicted the actual days.

Conclusion
In this paper we propose the algorithm Local-Global SVR with Parallel Time Variant PSO which is successfully used for forecast the water level for flood early waring system. For SVR, input feature are crucial problems, we get 3 days feature optimum for training phase and 4 days feature optimum for testing phase. The results indicate that the more we used days for feature didn't made the forecasting more accurate, the water level was highly influenced by 3 to 4 days before the actual days. In this algorithm PTVPSO was us ed to optimize the parameters of SVR such as λ, ε, cLR and C and this method can obtain suitable parameters of SVR to improve the forecasting performance of the proposed model. In the future we will study other water level forecasting using some attributes that influenced this object.