Time-series forecasting is a type of machine learning and statistical analysis technique that uses historical time-series data to infer future trends or states. This technique typically uses time-series data records and analyzes trends, periodicity, seasonality, and anomalous patterns to build predictive models that estimate potential data changes over a future period. Time-series forecasting is widely used in economic analysis, energy management, weather forecasting, industrial monitoring, and resource allocation.

The development of time series forecasting has a long history, and its theoretical foundation can be traced back to time series analysis methods in statistics, such as classic methods like autoregressive models (AR), moving average models (MA), and ARIMA. These methods predict future states by establishing statistical patterns in data changes over time. With the development of machine learning and deep learning, researchers have begun to use neural network models to capture more complex nonlinear time dependencies, driving time series forecasting from statistical modeling to data-driven methods.

In 2021, researchers from Beijing University of Aeronautics and Astronautics, the University of California, Berkeley, and other institutions published a paper... Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting This paper proposes the Informer model, addressing the challenges of high computational cost and difficulty in modeling long-distance dependencies in long-sequence time series prediction tasks using the Transformer model. It introduces methods such as ProbSparse Attention and distillation mechanisms to improve the efficiency of the Transformer in handling long-sequence tasks. This research has become one of the most important representative works in the field of Transformer-based long-sequence time series prediction.

Modern time series forecasting models typically learn trends, seasonalities, cyclical changes, and relationships between variables from historical data. Depending on the task, models can employ statistical methods, machine learning methods, recurrent neural networks (RNNs), convolutional neural networks (CNNs), Transformers, and, more recently, large-scale time series foundational models. Time series forecasting primarily addresses the problem of estimating future states in dynamic systems, enabling computers to conduct analysis and decision-making based on existing data. Currently, this technology is widely used in financial market analysis, weather forecasting, power load forecasting, traffic flow forecasting, supply chain management, and industrial equipment maintenance, providing data support for the planning and optimization of complex systems.