Dependable Flow Modeling In Upper Basin Citarum Using Multilayer Perceptron Backpropagation

Dependable Flow is the amount of discharge available to meet water needs with calculated risk of failure. In planning the provider of water projects must first look for the dependable flow, the purpose of which is to determine the planning discharge which is expected to always be available in the river [1]. The hydrological process can be simply described by the relationship between elements of input namely rain, process and output, namely in the form of flow [2]. The relationship of rainfall-runoff is an important problem in hydrology and is the most fundamental component in the process of evaluating water resources [3].

(physical process based), continuous models (continuous events), lumped models, distribution models and single models [7]. These models are formed by a set of mathematical equations that reflect the behavior of parameters in hydrology, so that the parameters contained in the equation have physical meaning [8]. ANN is one form of artificial intelligence that has the ability to learn from data and does not require a long time in making models [7]. This model uses a set of linear and non-linear mathematical equations that do not take into account the physical process at all, the most important in this model is that the output produced approaches the real one [8]. In addition, ANN is also able to identify the structure of the model and is also effective in connecting simulation inputs and outputs and forecasting models [7]. Benefits of using ANN is his ability to study relationships preexisting unknown between the input and output data of each system. Besides the modeling with ANN has that attribute desirable and learning abilities from examples without explicit physical data.
The application of artificial neural networks has been widely used by previous researchers, such as, [9] using the multilayer perceptron backpropagation method as an optimization of runoff rainfall on the island of Java, namely the Ciujung River, Banten, [10] who used artificial nerves to predict discharge at The Jinsha River Basin, China. [11] made a prediction model of river water salinity using artificial neural networks, and [12], using artificial neural networks to predict rainfall and climatology. Based on the above studies, according to researchers the method of artificial neural networks needs to be applied to predict an increase in the number of new students, because by using this method to obtain good accuracy and fast process. [13] used an ANN to model the relationship between rainfall and daily runoff in the Leaf River, Collins, Mississippi, USA. In order to evaluate the ANN performance, they compared the results obtained to the results of the SACRAMENTO model and autoregressive models of the ARMAX type. In this situation, ANNs presented the smallest errors and the smallest differences in volume. Good results were also obtained by applied ANN to flow forecasting for flood control in the Red River, Canada, and compared the results to linear and nonlinear multiple regression models; [14] applied ANN to three basins in Kansas, USA, to model the mean monthly flow, and also compared the results with empirical models; and [15] applied the ANN to three different USA basins: the Fraser River, Colorado, Raccoon Creek River, Iowa and the Little Patuxent River in Maryland, for monthly and daily forecasting horizons. Modarres [16] tested multilayer perceptron models in the Zayandehrud watershed in Iran. In all cases, the ANNs presented very good results, superior to the models to which they were compared.
Technological developments and the complexity of getting the parameters needed by using empirical formulas, moreover it is quite difficult to get results from quite a lot of data. To simplify this, the authors conducted a reliable discharge modeling that would be based on artificial neuron networks with the help of MATLAB software. This artificial nervous system method is useful only for real time not for recalculation.

II. Artificial Neural Network
Artificial neural networks (ANN) are information processing systems that have characteristics similar to biological neural networks, which are artificial representations of the human brain, which contain millions of nerve cells (neurons) and function to process information. Neurons have the same characteristics in ANN, consisting in groups called layers. Neurons in one layer are connected in the other adjacent layers. The strength of the relationship between adjacent neurons is represented in the strength of the relationship or weight.
An ANN generally consists of three layers, namely input layer, hidden layer and output layer. The input layer consists of neurons that receive an input from the outside environment. The input entered is a description of a problem. The hidden layer consists of neurons that receive input from the input layer, and then bring the output to the next layer. The output layer is called the output units, consisting of neurons that receive output from the hidden layer and send it to the user. Rosenblatt (1957) in Hung et al., (2009) introduced the first simplest form of artificial neural network called a perceptron, which consists of only one layer [17]. Inputs are given directly to the output unit via a weighting connection. In 1960 a multi layer perceptron (MLP) was developed and gradually became a neural network topology that was widely used in various fields, and included in the feedforward network category. In this network, besides input and output units, there are other units (often called hidden layer).
For system modeling, there are several advantages to using artificial neural networks [18], namely: a. Nonlinear system. Artificial neural networks have the ability to map relationships that are not linear, so it is very promising to overcome the problem of nonlinear control. b. Learning and adapting. Artificial neural networks that have been trained to use a set of data will be able to issue output equal to or close to the input pair output given, even able to issue output with inputs that have never been trained. Artificial neural networks can also be trained online. c. Multivariable system. Applicable artificial neural network for multi-variable processes. The common artificial neural network structure is a multi layer perceptron (MLP). Figure 1 illustrates MLP structure, which consists of input, hidden and output layers. The output produced by ANN for a particular input pattern depends on the value (weight) of the relationship between neurons in the ANN. In ANN, for pattern recognition applications an initial training is needed so that it can be used in receiving input from outside. An ANN can solve a complicated problem if the appropriate weight values are used between neurons in different layers. The right weight value is obtained through the training process. The training process is a process to change the weight between neurons so that a network can solve a problem. An important step in developing an ANN model is determining the weight of the matrix through training. There are two types of training mechanisms, namely supervised training and unsupervised training. Supervised training requires supervision from outside to guide the training process. This algorithm uses a number of pairs of input-output data that are used for example, where data used as an example should use data that is already known to be correct. The output from the network is then compared to the expected output data (sample output) to get the difference between the estimated output and the actual output. This difference is used to change the weight of the network to obtain the same output or close to the target. The mechanism of a supervised training called the backpropagation training algorithm [19] is generally widely used in engineering applications. Because ANN does not consider the physical problem, ANN is a blackbox model but can detect physical processes in ANN models that have been trained [20], [21], [22].

III. Material and Method
A. Location In this study, Citarum Upper Basin is the location of the study. Increasing the number of population which is increasing every year in the Citarum Upper Basin has an impact on increasing land needs. Changes in forest areas and agricultural land into residential land and various other uses have caused many negative impacts on land and water resources that occur in watershed areas.
In this study, the location of the observation was the Cikapundung sub-watershed and the

B. Data and Software
The data used are daily rainfall data, discharge data and water level data for 10 years from 2008-2017. Rainfall data for Cikapundung sub-watershed is obtained from 3 rain observation stations (Dagopakar, Margahayu, Mateo Lembang) and used as regional rainfall data for Citarik subwatershed is obtained from 3 rain observation stations too (Jatiroke, Tanjungsari, Cibiru). Data sources were obtained from BBWS Citarum and PUSAIR. Topographic data is obtained from the Geospatial Information Agency (BIG). The software used to process numerical data is used by Microsoft Excel 2010 and to create a model used MATLAB R2015a which has several features grouped based on certain applications (Toolbox) related to Artificial Neural Networks namely Neural Network Tools.

C. Modeling Procedure
The modeling process with ANN is carried out in several stages, the first providing data to be used as input and data that will be targeted. The data is further divided into training data and testing data. Training data is used to train models that are made to approach the targeted results. While testing data is used to test the model whether the model functions to predict well.
In this case, a model with the MLP-BP (Multilayer Perceptron Back-Propagation) method was used. The correlation formed from the model will be presented using the Nash-Sutcliffe Efficiency (1) and R2 (2) as indicated by formulas, to ensure the model has a small error value, the approach with mean square error (MSE) is formulated (3).

(1)
Where is modeled data, is the observation data and is the average of the observation data. (2) Where Y is the data modeled and X is the observation data. (3) Where is the data modeled and is the observation data.

D. Aplication
Training data and testing data use rainfall, water level and discharge data from 2008-2017. For training used data from 2008-2012 and testing used data from 2013-2017. The transformation process is carried out to change the input and output data values to a scale of 0 to 1 using the following equation: Three artificial neural network structures were formed to be tested, each of which was as follows: -Model 1 Model 1 has 2 input nodes, rainfall and discharge data, 8 hidden layer nodes and 1 layer output node as in Fig. 5 and Fig. 8 -Model 2 Model 2 has 3 input nodes, rainfall data, discharge and water level, 12 hidden layer nodes and 1 layer output node as in Fig. 6 and Fig. 9 Model 3 has 3 input nodes, rainfall data, discharge and water level, 12 hidden node layers and 1 output layer node. in model 3 is calculated by separating the model into per month as in Fig. 7 and

IV. Results and Discussion
The performance of each network model in modeling the rainfall-runoff relationship for the Cikapundung watershed was calculated by mean square error (MSE), Nash-Sutcliffe Efficiency (NSE) and correlation coefficient (R2). The training performance results of the model will be displayed in the Table 1 and will be represented in graphical form as in Fig. 10-15. The following are the results of the hydrograph and dependable flow graph of 3 models:

V. Conclusion
The ability of artificial neural network models to predict can be adjusted by changing existing parameters. In this case, the use of the number of neurons in the hidden layer, the more neurons used, the iteration process will be shorter and get faster results. But too much use of neurons can cause the network to become over fitting. Then it must be adjusted to the existing data input in order to avoid over fitting. In this study, the use of neurons was effective with 50 neurons for each hidden layer.
From the three models that have been formed it can be concluded that the third model has a correlation value and a good error value to be used. These results it can be seen that ANN has a fairly good ability to replicate random discharge fluctuations in the form of artificial models that have almost the same fluctuations and can also be applied in rainfall runoff modelization even though the results of the test results are not very accurate because there are still irregularities. With the dependable flow modeling this can be a parameter for decision making what will be the solution to water use in the Upper Basin Citarum. The results of modeling this artificial neural network can be used by the stakeholder to be used as decision makers adjusted to the real conditions in the Upper Basin Citarum, especially in Cikapundung and Citarik.