A Systematic Approach to Modeling and Forecasting Based on Real Data in Machine Learning Tasks

Abstract

The article considers a systematic approach to modeling and forecasting nonlinear non-stationary data in machine learning problems. It is based on a detailed analysis of the processes under study, determination of the types of existing characteristic uncertainties, assessment of the structure and parameters of the model, and forecasting based on the constructed model. The systematic approach to modeling and forecasting combines three groups of problems on a single methodological basis: problems of data analysis and preliminary processing; problems of building models and their assessment; problems of building forecasts and their assessment. The features of nonlinear and non-stationary processes are considered. The architecture of the information and analytical system for solving forecasting problems is developed and presented. As an example of solving the forecasting problem, the problem of forecasting the indicators of electricity generation by hybrid power plants based on environmental indicators is considered. The solution to the forecasting problem is based on the methodology developed on the basis of a systematic approach to modeling and forecasting. It consists of the following stages: data collection, data research and preparation, model training on data, determining the effectiveness of the model, improving the effectiveness of the model. The following basic models were used to solve the forecasting problem: regression models based on decision trees, models based on KNN, neural network models, models based on multiple regression, models based on random forest, models based on the support vector method. To further improve the forecasting results, a two-level heterogeneous ensemble was used based on the running method at the first level and stacking at the second. As a result, the values of bias and variance of the forecast error were reduced. The effectiveness of predictive solutions was assessed by the following indicators: MAE, MSE, RMSE, MAPE.