Machine Learning Models Informed by Connected Mixture Components for Short- and Medium-Term Time Series Forecasting

Andrey K. Gorshenin, Anton L. Vilyaev

This paper presents a new approach in the field of probability-informed machine learning (ML). It implies improving the results of ML algorithms and neural networks (NNs) by using probability models as a source of additional features in situations where it is impossible to increase the training datasets for various reasons. We introduce connected mixture components as a source of additional information that can be extracted from a mathematical model. These components are formed using probability mixture models and a special algorithm for merging parameters in the sliding window mode. This approach has been proven effective when applied to real-world time series data for short- and medium-term forecasting. In all cases, the models informed by the connected mixture components showed better results than those that did not use them, although different informed models may be effective for various datasets. The fundamental novelty of the research lies both in a new mathematical approach to informing ML models and in the demonstrated increase in forecasting accuracy in various applications. For geophysical spatiotemporal data, the decrease in Root Mean Square Error (RMSE) was up to 27.7%, and the reduction in Mean Absolute Percentage Error (MAPE) was up to 45.7% compared with ML models without probability informing. The best metrics values were obtained by an informed ensemble architecture that fuses the results of a Long Short-Term Memory (LSTM) network and a transformer. The Mean Squared Error (MSE) for the electricity transformer oil temperature from the ETDataset had improved by up to 10.0% compared with vanilla methods. The best MSE value was obtained by informed random forest. The introduced probability-informed approach allows us to outperform the results of both transformer NN architectures and classical statistical and machine learning methods.