ОЦЕНКА МЕТОДОВ МАШИННОГО ОБУЧЕНИЯ ДЛЯ ПРОГНОЗИРОВАНИЯ СЕЗОННЫХ ВРЕМЕННЫХ РЯДОВ

Abstract

Time series with seasonal variability is widely used to describe processes in variousfields, such as trade, analysis of financial markets, forecasting of passenger air transportation,and description of climatic changes. Recently, this approach has been widely used to describetechnological processes as well. In this regard, applying predictive models in control systems ofcomplex technical objects has become possible. Machine learning methods can be effectivelyused to build predictive models of series of this type. In this case, only historical data accumulatedover several periods of seasonal observation is used as input data for constructing theforecast. Knowledge of other parameters, as a rule, is not required. The article considers creatinga predictive time series model with seasonal variability, describing a technological process,the inlet flow of a wastewater treatment plant being chosen as a model. The general methodologyof model building, requirements for the input data sets, and algorithms of preprocessing toform samples used for model training and testing are described. Classical methods (SARIMA,Holt-Winters Exponential Smoothing, ETS), as well as new algorithms (Facebook Prophet,XGBoost, Long Short Term Memory), were used to build the predictive model. The implementationof the algorithms is done in the Python language, and recommendations for the use of existinglibraries and functions of this language are given in the work. The comparative analysis ofthe accuracy of the obtained models is given on the calculation of a set of statistical metri cs.Analysis of methods performance is also carried out since the time it takes to create a modeland get a forecast plays an important role when running the model in real production conditions.The best method for solving the set task for application in real-time control systems waschosen based on the sum of estimates. In conclusion, recommendations for improving forecastaccuracy were given, and future research directions were outlined.