Using Data Mining Methods (Neural Network) in Tehran Fuel Consumption Analysis in Public Transportation

Abstract

To predict demand for oil products using statistical methods such as regression models, auto regression and moving average, from the beginning, the form of the functional relationship between the dependent variable (consumption of the product) and the independent variables (total population, urban population Rural population, gross national product, value added of industries and mines, road and rail transportation section, agricultural sector, product prices, number of diesel fuel buses, number of mini- buses and capacity of power plants ) Be clear. In most cases, for simplicity, linear, quadratic or logarithmic relationships are assumed, although this simplification may lead to false results. Determining the functional relationship between the consumption of the product and the factors affecting it is a very complicated problem and it is simply not possible. Therefore, the use of intelligent systems such as neural networks, which in recent years has been considered by many experts, is reasonable. In this study, multi-layer neural networks with retro-reflection learning were used to predict Tehran's gas consumption. In this study, in order to find the predicted values closer to the desired output values and to select the best solution with the lowest MSE, the network problem Neural networks were solved for

various values of the secret layer neurons and educational, evaluation and experimental sets. Since manual selection of numbers in addition to being time-consuming is a possible mistake due to the fact that some values are not selected, in order to avoid this problem, the optimal number of secret layer neurons, the percentage of data related to the training, evaluation and testing sets It was found by MATLAB software and the problem was solved for each of the MLP, MLFF, and MLCF states, which ultimately was the best answer with the lowest MSE of MLFF mode with 8 neurons in the hidden layer, and taking 65 % Of the data for the training set, 16% for evaluation and 19% for the test. Eventually, by the time series of the network, it was trained to predict the next seven courses, and the gasoline consumption rate was projected until 1400.