Resource usage forecasting model in cloud computing using Informer architecture

The object of the study is the process of forecasting and monitoring the load of cloud infrastructure. It is determined that one of the effective ways to implement this process is the use of transformer-based neural networks for processing multivariate telemetry time series. Classical architectures, in particular Informer, provide high forecasting accuracy but require significant computational resources and long training times, which complicates their integration into real-time monitoring systems. The aim of the research is to develop and experimentally evaluate a modified Informer architecture optimized for faster training and more efficient resource usage while maintaining acceptable accuracy. The proposed model takes seasonal patterns into account and implements efficient self-attention with positional and seasonal embedding representations. The model was tested in a system implementing a closed MAPE (Monitoring–Analysis–Planning–Execution) loop, which made it possible to assess its performance in automatic resource management for Kubernetes. A comparative analysis with the classical Informer architecture was carried out. The optimized model achieved an eightfold acceleration in training and a tenfold reduction in the number of parameters, while maintaining over 88 % of the explained variance and comparable accuracy in forecasting CPU, memory, network, and disk metrics. This makes it suitable for deployment in industrial cloud resource monitoring and management systems in real-time operation.