This paper proposes a novel framework for multivariate time series forecasting called MTGNN, which integrates graph neural networks (GNNs) to capture spatial and temporal dependencies. The framework consists of three core components: a graph learning layer, a graph convolution module, and a temporal convolution module. The graph learning layer automatically extracts uni-directed relations among variables, while the graph convolution module captures spatial dependencies and the temporal convolution module captures temporal patterns. The framework is trained in an end-to-end manner, allowing it to simultaneously learn the graph structure and the GNN. The proposed model outperforms state-of-the-art baseline methods on three of four benchmark datasets and achieves on-par performance with other approaches on two traffic datasets. The model is effective in handling both small and large graphs, short and long time series, with and without externally defined graph structures. The framework is designed to address two main challenges: unknown graph structure and the need to learn both the graph structure and the GNN in an end-to-end framework. The model uses a curriculum learning strategy to improve performance on multi-step forecasting tasks. The results show that the proposed model achieves state-of-the-art performance on both single-step and multi-step forecasting tasks. The model's effectiveness is validated through ablation studies and case studies, demonstrating its ability to capture spatial-temporal dependencies in multivariate time series data.This paper proposes a novel framework for multivariate time series forecasting called MTGNN, which integrates graph neural networks (GNNs) to capture spatial and temporal dependencies. The framework consists of three core components: a graph learning layer, a graph convolution module, and a temporal convolution module. The graph learning layer automatically extracts uni-directed relations among variables, while the graph convolution module captures spatial dependencies and the temporal convolution module captures temporal patterns. The framework is trained in an end-to-end manner, allowing it to simultaneously learn the graph structure and the GNN. The proposed model outperforms state-of-the-art baseline methods on three of four benchmark datasets and achieves on-par performance with other approaches on two traffic datasets. The model is effective in handling both small and large graphs, short and long time series, with and without externally defined graph structures. The framework is designed to address two main challenges: unknown graph structure and the need to learn both the graph structure and the GNN in an end-to-end framework. The model uses a curriculum learning strategy to improve performance on multi-step forecasting tasks. The results show that the proposed model achieves state-of-the-art performance on both single-step and multi-step forecasting tasks. The model's effectiveness is validated through ablation studies and case studies, demonstrating its ability to capture spatial-temporal dependencies in multivariate time series data.