TimeCNN: Refining inscross-variable interaction on time point for time series forecasting.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Transformer models show promise for time series analysis but struggle with dynamic, multifaceted cross-variable correlations.
• Existing models often fail to capture both positive and negative correlations that evolve over time in multivariate time series.

Purpose of the Study:

• To propose TimeCNN, a novel model designed to refine cross-variable interactions for improved multivariate time series forecasting.
• To address the limitations of current Transformer-based models in handling dynamic and complex inter-variable relationships.

Main Methods:

• Introduced TimeCNN, a model featuring a timepoint-independent approach where each time point utilizes a unique convolution kernel.
• This allows for independent modeling of relationships among variables at each specific time point.
• The approach effectively captures both positive and negative correlations and adapts to their temporal evolution.

Main Results:

• TimeCNN demonstrated superior performance across 12 real-world datasets compared to state-of-the-art models.
• Achieved significant reductions in computational requirements (approx. 60.46%) and parameter count (approx. 57.50%).
• Inference speed was 3 to 4 times faster than the benchmark iTransformer model.

Conclusions:

• TimeCNN offers an effective solution for multivariate time series forecasting by accurately modeling complex cross-variable dynamics.
• The model provides substantial efficiency gains in computation and speed, making it a practical advancement.
• Code availability will be ensured through public release on GitHub.