<p>Modeling multivariable time series remains a critical research area. Despite the numerous time series prediction models introduced, the development of models that balance interpretability and accuracy remains an open challenge. In this work, we propose a novel Spatial-Temporal Fuzzy Cognitive Map based adaptive Graph Learning model (STFCM-AGL) to achieve interpretable multivariate time series predictions. First, leveraging high-order fuzzy cognitive maps, we introduce granularity adjustment strategies to obtain dynamic interval weights such that the model’s capability can be improved in complex system. Second, the Adaptive-GraphSAGE module is designed based on the graph structure of the fuzzy cognitive map to capture subtle changes in node features and to fully leverage the deeper relationships between concepts. Moreover, leveraging node embeddings enriched with spatial information, the temporal convolution module enhances the model’s time-series representation capabilities. Finally, the parameters of the fuzzy cognitive map, graph neural network, and temporal convolution module are jointly learned within an end-to-end framework. Experimental results demonstrate that the proposed model outperforms state-of-the-art baseline methods across real-world datasets from transportation, finance, and environmental domains, while maintaining the interpretability of the foundational fuzzy cognitive graph model.</p>