Forecasting based on machine learning methods in the fuel monitoring system

Abstract

The article considers a forecasting information system based on machine learning methods for fuel monitoring. The system solves the following monitoring tasks: data analysis and evaluation, model building and forecasting values or decision-making. The system is developed based on information technologies for fuel monitoring. The information system consists of the following subsystems: information collection and storage subsystem, data preparation subsystem, data analysis and pre-processing subsystem, modeling subsystem and forecasting subsystem. An important place in the modeling and forecasting subsystems is occupied by modules for assessing the quality of models and forecast values based on quality metrics. In the forecasting subsystem, in particular, the forecasting module based on basic alternative models has a forecast value combination module, which implements seven different methods for combining forecast values. In most cases, combination helps to improve the quality of forecasts. The experimental part of the study considers the problem of predicting the volumes of possible filling of storage systems with fuel based on a report on regular data collection on the level and amount of fuel in the tanks on the ship. The following machine learning methods were used for forecasting: exponential smoothing, regression neural network models and Bayesian structural time series models. The quality assessment of the obtained forecast values was carried out using the following quality metrics: MAE, MSE, RMSE. The information system makes it possible to obtain high-quality forecasts of the amount of fuel for tanks of various types, as well as generalized indicators.