Knowledge-graph-driven environmental monitoring with cross-regions knowledge transfer

Traffic flow prediction is a critical task in intelligent transportation systems. However, cross-city data-driven prediction models encounter numerous challenges. A principal issue is data scarcity in underdeveloped cities, resulting from inadequate data collection mechanisms. Additionally, these models frequently rely exclusively on direct spatio-temporal traffic data, often deficient in thorough extraction and lateral exploration of external information from source cities, thereby introducing risks of negative transfer. This paper presents KGD-Transfer, a knowledge graph-driven cross-city traffic flow prediction framework that utilizes the extensive semantic information embedded within knowledge graphs. To address the challenge of data scarcity, we advocate for the construction of Fused Meta-Knowledge ( \({{\textbf {FM}}_k}\) ${\mathbf{FM}}_k$ ) from multiple source cities to facilitate precise traffic prediction. Our knowledge fusion network (Ka-net) deeply integrates heterogeneous knowledge graphs with traffic data. Following this, we apply Neural Controlled Differential Equations (NCDEs) to extract intricate spatio-temporal features from the integrated knowledge. Furthermore, the Model-Agnostic Meta-Learning (MAML) framework is utilized to efficiently minimize the risk of negative transfer in cross-city learning. This approach enables the transfer of fused source knowledge by employing non-shared parameters to perform deep feature matching across cities, capitalizing on the spatio-temporal commonalities within the \({{\textbf {FM}}_k}\) ${\mathbf{FM}}_k$ . Experimental results from four real datasets illustrate that the inclusion of contextual information from knowledge graphs markedly enhances the prediction model’s understanding and reasoning capabilities. KGD-Transfer outperforms advanced baseline methods in both short-term (10 min) and long-term (60 min) predictions, demonstrating superior accuracy and generalizability.