Time series forecasting plays a pivotal role in various fields, such as economics, meteorology, energy, and others. This paper presents the architecture of a hybrid deep learning model designed for forecasting time series values. Traditional models, developed based on machine learning methods and statistical approaches, demonstrate certain achievements and successes in addressing a broad range of tasks across various applied fields. However, despite their potential strength and versatility, these models face several limitations in situations where data exhibit high complexity and unpredictability. In conditions where data are characterized by high dimensionality, unstructured nature, or contain a significant amount of noise and anomalies, traditional approaches may lose the ability to adequately interpret and extract useful information from such data. To overcome these limitations, a hybrid model combining the Prophet forecasting model, recurrent neural networks (RNNs), and convolutional neural networks (CNNs) has been proposed. Convolutional neural networks have the capability to detect patterns that are inaccessible to statistical methods, while recurrent neural networks are effective for processing sequential data. Thus, combination of these architectures allows for improved forecasting quality. Experiments conducted on various datasets have shown that the proposed hybrid model outperforms some conventional statistical methods and individual machine learning models in terms of forecast accuracy. The developed model is promising for applications requiring a high level of predictive accuracy in conditions of complex temporal dependencies.
A.V. Gulyaev (Mon,) studied this question.