Abstract: Aiming at the problem that most traditional methods for multidimensional time series analysis relied on manually establishing temporal dependencies to explore the  implicit rules  in historical data, we proposed  an autoregressive neural network method. Firstly, the neural network composed of convolution neural network (CNN) and bidirectional long short-term memory (LSTM) was used to capture the complex dependencies existing in multidimensional input features and time series, and the linear relationship was extracted by combining the traditional autoregressive method. Secondly,  compared with several classical models on two datasets in different domains, the experimental results showed that the model had the best prediction performance and could  successfully capture the repeated patterns in the data. Finally, the  ablation experiments verified the efficiency and stability of the model framework.

Key words: multidimensional time series, neural network, autoregressive model