Multivariate time series (MTS) forecasting has important application value in fields such as finance, healthcare, and meteorology. In recent years, forecasting models based on the Transformer architecture have received widespread attention. However, existing Transformer-based models often treat each variable sequence as an independent single variable sequence, ignoring the dependency relationships between variables. In addition, the self-attention mechanism faces limitations of quadratic time complexity and high memory usage when dealing with long sequence time-series forecasting (LSTF). Therefore, we propose the Intention-based Multivariable Time Series Forecasting Transformer (IPSTT). Specifically, IPSTT captures the temporal and inter variable dependencies of MTS through a spatial-temporal Transformer. We also use probsparse self-attention mechanism to reduce the time complexity of LSTF. In addition, pre trained GRU provides predictive intent for spatial-temporal Transformers. The experimental results show that the proposed method achieves a 50% and 20% improvement compared to the conventional methods on benchmark datasets and aviation datasets.