Hybrid Convolution Based Online Multivariate Time Series Forecasting Algorithm

Proposed is an online multivariate time series forecasting algorithm based on mixed convolution to address the issue of low prediction performance of existing algorithms in real-time online forecasting tasks for data streams. Firstly, a mixed convolution layer is used to extract local changes in time series in the time dimension and the correlation between variables in the variable dimension. This helps the model better capture cross-time and cross-variable dependencies of time series, while focusing on local modeling of time series to improve the model’s adaptability to frequent changes in data distribution. Secondly, the model adopts a linear structure to meet the real-time requirements of forecasting tasks, learning the complex variations of the time series through deep convolutional neural networks and using linear transformation to achieve real-time output of forecasting results. Experimental results demonstrate that the proposed algorithm outperforms six mainstream multivariate time series forecasting algorithms on four datasets in terms of prediction performance.