BACKPROPAGATION METHOD TO PREDICT RAINFALL LEVELS IN ROKAN HULU DISTRICT

: Rokan Hulu Regency is one of the regencies in Riau Province and has 2 large rivers, namely the Rokan Kanan River and the Rokan Kiri River. When the level of rainfall is high, the river flow overflows and causes flooding in several sub-districts. Climate change that is happening has an impact on changes in rainfall patterns, to overcome the impacts of climate change Rokan Hulu Regency needs to take early steps before a flood disaster occurs. The method used in processing the data is the backpropagation method and data processing using the Matlab software. Data processing was carried out on rainfall data for the last 6 years, from 2015 to 2020, which was sourced from the Rokan Hulu Regency Statistics Agency. Rainfall data from 2015 to 2019 is used as input data, while 2020 is the target. The data processing stage begins with data normalization and determining network training parameters. Artificial neural network research was carried out on rainfall data with a 5-5-1 architecture. The results of testing the rainfall data show prediction results with high accuracy, where the average accuracy rate reaches 98.17% with the highest accuracy of 99.99% while the lowest accuracy is 91.53%. The prediction results show a fairly high level of accuracy so that they can be used as evaluation material.


INTRODUCTION
In today's development, technology has become something that is familiar to the community, all activities are carried out using existing technology. The purpose of developing a technology is to facilitate and assist in carrying out various human activities and activities, such as technology at the Meteorology, Climatology and Geophysics Agency (BMKG) which records rainfall levels to make it easier to find out the high and low levels of rainfall that occur. Rain is very useful in the continuity of human life and in various agricultural sectors that require water at normal intensity. When the intensity of rainfall is high, it can cause natural disasters, one of which is flooding.
Rokan Hulu Regency is one of the regencies in Riau Province and has 2 large rivers, namely the Rokan Kanan River and the Rokan Kiri River. When there is high rainfall, the rivers in Rokan Hulu Regency overflow and cause flooding in various sub-districts. The floods that occurred in Rokan Hulu Regency not only caused material losses but also claimed lives, such as the 2019 flood which caused the death of a toddler girl to be found dead in the flood currents.
Climate change has many impacts on changes in rainfall patterns, to overcome the impacts of climate change, Rokan Hulu Regency needs to take early steps before a flood disaster occurs by managing sewers, water management, and a flood forecasting system. Disasters can be anticipated with accurate information about the level of rainfall that will occur within a certain period of time. Thus a prediction system for future rainfall levels is needed. The resulting prediction of rainfall levels can be used as material for anticipation and reference for the Rokan Hulu Regency government in evaluating the impact of greater losses due to flooding. The resulting level of prediction accuracy is determined by the method, in this case a suitable and accurate method for predicting rainfall levels is a backpropagation artificial neural network.
Rainfall prediction is the use of science and technology to predict the future state of the earth's atmosphere at a certain place. Rainfall prediction is done with the help of a computer or certain software in the modeling system [1] [2].
Backpropagation is a multi-layer feed forward training method that has been proposed by Rumelhart and McClelland [3] [4]. Backpropagation is included in the popular category of artificial neural networks and has advantages in the learning or training process, the training or learning process is repeated by having good computation, especially when data is presented on a large and complex scale [5].
Previous backpropagation research has been carried out by Setti and Wanto. Research predicts the number of internet users in 25 countries with the highest level of internet usage and produces the best network architecture 3-50-1 with an accuracy rate of 92% [6]. Furthermore, Lesnussaa et al also conducted research on the backpropagation method and produced an accuracy rate of 80% [7]. Sovia et al researched backpropagation to predict bitcoin price movements and were able to conduct training and data testing based on network patterns that had been formed [8]. Purnawansyah et al used backprogation to predict the inflation rate in Samarinda City, architectural parameters 5-5-5-1 produced a good prediction error rate [9]. Muflih et al used backpropagation to predict rainfall in Wonosobo Regency, resulting in an MSE of 0.17042 when testing the network [10]. Saragih et al conducted research on predicting the value of exports in North Sumatra with backpropagation. The study uses 5 architectural models namely 4-5-1, 4-7-1, 4-9-1, 4-10-1 and 4-11-1. The best model of these 5 models is 4-7-1 with a 100% accuracy rate, with a time of 27 seconds. The error rate used is 0.001 -0.05 [11].
Based on the background that has been described, the researchers conducted a study entitled "Backpropagation Method for Predicting Rainfall Levels in Rokan Hulu Regency"

METODE
The framework for the research conducted can be seen in Image 1.

Image 1. Research Framework
Description of the steps in the research framework Image 1 is:

Identify the Problem
Identifying the problem is the initial stage that is carried out to determine the formulation of the problem that occurs in Rokan Hulu Regency.

Analyze the Problem
Analysis of the problems that have been determined so that the problems that occur can be well understood.

Rainfall Data for Rokan Hulu Regency
The data used in this study are rainfall data in Rokan Hulu Regency for the last 6 years from 2015 to 2020 obtained from the Central Bureau of Statistics for Rokan Hulu Regency.

Perform Data Normalization
The data normalization stage is the stage of transforming the rainfall data that has been obtained so that the distribution of the data becomes new data that is more normal and does not occur anomalies. Data normalization is performed using a binary sigmoid activation function (never reaches 0 or 1).

Normalized Rainfall Data (Binary Sigmoid)
Normalized rainfall data contains rainfall data for 2015 to 2020 with values never reaching 0 or 1.

Data Processing With Backpropagation Method
Data processing is based on the stages of the backpropagation method as follows: a. Perform random initial weight initialization Initialize the initial weights which are determined randomly, namely from the input layer to the hidden layer and from the hidden layer to the output layer. The feedforward stage is the forward propagation flow stage where each input layer neuron receives a signal from the input data which is then forwarded to the hidden layer neurons and from the hidden layer neurons it is forwarded back to the output layer neurons.

d. Doing the Backpropagation Stage
The backpropagation stage is the backward propagation flow stage where each neuron of the output layer with the input pattern is then forwarded to the hidden layer neurons and from the hidden layer neurons it is forwarded back to the input layer neurons. The backpropagation stage will calculate any weight changes that go to the hidden layer and the output layer. e. Calculates the minimum error value Calculating the minimum error is used to see the conditions for stopping the learning process. The termination of the learning process is determined by the maximum number of iterations or the minimum error (target) value in determining the learning termination process.

Implementation and Testing
The implementation phase is carried out by building a GUI-based system using matlab software to facilitate the process of processing rainfall data into useful information. Furthermore, testing the system is it feasible to use to predict rainfall that will occur in the future and manual calculations with the system are valid or appropriate.

Results and Discussion
The results and discussion will describe the results of the processing and testing of the data that has been carried out. The results obtained are in the form of predictions of rainfall levels and the accuracy of predictions in the future with the aim of being used as a policy reference in managing the impact of changes in rainfall disasters in the future.

RESULT ANF DISCUSSION
This research uses rainfall data for the last 6 years, from 2015 to 2020. Furthermore, rainfall data is divided into 2, namely training data and test data. There is also rainfall data in Table 1. Rainfall data in Table 1 will be analyzed using the backpropagation method to predict future rainfall levels with the following stages of the backpropagation process:

Input data
The input data is in the form of rainfall data that occurred in Rokan Hulu Regency. Initialize the input data in Table 2.
The normalization results of rainfall data presented in Table 1 can be seen in Table 3. Table 3. Rainfall Data Analize    Next, do the backward propagation calculation stage, namely from the output layer to the input layer. Calculate the error in the output layer to find out whether the target error has been met or not by using the equation: When the error value has not reached the target, then it is continued with the calculation of changes in weight values with learning rate (α) = 0.1. Use the equation: (10) Then the results of calculating the weight correction value for each output unit are presented in Table 8. Next, calculate the corrected bias value from the hidden layer using the equation: (11) 0,1 * (0,00194) = 0,000194 0,1 * (-0,00394) = -0,000394 -0,000129 0,000120 -0,000366

Renewal (Update) Weights and Bias
Calculate the change in weight and bias from the input layer to the hidden layer or from the hidden layer to the output layer from the input layer to the hidden layer. Calculate the change in weight values from the input layer to the hidden layer using the equation: (12) The following is the result of changing the weight from the input layer to the hidden layer in Table 9. Calculate the change in weight from the hidden layer to the output layer using the equation: (13) The following is the result of changing the weight from the hidden layer to the output layer in Table 10.  Table 12 is the result of ANN testing of rainfall data that occurred from 2015 to 2020. The results of testing of rainfall data show prediction results with high accuracy, where the average accuracy rate reaches 98.17% with the high-