Multivariate time series forecasting is widely used in fields such as meteorology and transportation. However, due to limitations in data collection, transmission, and storage, the Multivariate time series data used for prediction often has varying degrees of missingness, which affects the accuracy of the forecast results. Considering that traditional methods of imputing missing values may lead to the accumulation of errors, we propose a Multivariate time series forecasting for Missing (TSFM) aimed at addressing missing data issues, focusing on minimizing the impact of missing values on the capture of spatiotemporal features. The model includes three key modules: the mask-updating temporal convolution module, the spatial-correction graph convolution module, and the spatiotemporal reconstruction prediction module. In the mask-updating temporal convolution module, an updated component of the missing mask is utilized to minimize the impact of missing values on the capture of long and short-term temporal dependencies. The spatial-correction graph convolution module introduces local spatial biases to correct errors in spatial feature extraction caused by missing values. The spatiotemporal reconstruction prediction module integrates global and local spatiotemporal features to reconstruct erroneous spatiotemporal dependencies resulting from missing values. Experiment on three real-world datasets with data missing rates set at 0.2, 0.4, 0.6, and 0.8 indicate that the TSFM model achieved higher forecasting accuracy compared to eight other benchmark models, even when there is an 80% missing rate in the data, the model can still make predictions with minimal error.