Recently, multivariate time series (MTS) forecasting model based on deep learning have demonstrated remarkable performance. However, existing models use all series for forecasting without selection, even though not every series actually improves forecasting accuracy. Auxiliary variables with low correlation to the target variable that are uninformative can significantly decrease forecasting accuracy. Furthermore, on high-dimensional datasets, utilizing all auxiliary series, including those with low-correlation, markedly increases the computational time and memory overhead of the MTS forecasting model. To address this issue, this paper proposes an Auxiliary Variables Enhanced intra- and inter-Series Tokenization (AVEiST), which incorporates an Auxiliary Variables (AVs) module and an intra- and inter-Series Tokenization(iST) module. The AVs module selects the top k auxiliary variables with the highest correlation to the target series. The iST module generates patch tokens from the target variable and series tokens from both the target and auxiliary variables, employs attention mechanisms to capture both intra-series and inter-series features using these tokens. This approach maximizes the extraction of effective feature representations for forecasting the target variable and reduces time and memory consumption. The experimental results demonstrate that AVEiST achieves superior performance over five baseline models, including iTransformer and PatchTST, across eight datasets, significantly improving MTS forecasting accuracy.