Research map / Probabilistic and scientific modeling / Dynamical systems
FlowNet: Modeling Dynamic Spatio-Temporal Systems via Flow Propagation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FlowNet, a physics-inspired spatio-temporal forecaster that models node interactions as conserved flow-token transfers with learned adaptive spatial masks, and reports gains over ten baselines on PEMS04F, DeepBase, and SINPA.
Problem
Spatio-temporal forecasting methods based on graph convolution or spatial attention rely on similarity-driven connectivity and fixed or noisy neighborhoods, neglecting asymmetric, flow-mediated interdependencies between nodes.
Approach
Flow tokens act as information carriers that are allocated from source to destination nodes via Flow Allocation Modules, obeying a conservation law (source depletes, destination gains), with a learnable node-specific Adaptive Spatial Masking radius restricting interactions, and a cascaded design using hyper-connections and a Mixture-of-Linears MLP.
Key results
- On PEMS04F short-term, FlowNet reaches MAE 18.48 and RMSE 29.03, versus 18.73/29.29 for the second-best STAEformer.
- On SINPA short-term, FlowNet reaches MAE 39.03 and RMSE 76.04, versus 59.04/103.75 for GWNET.
- On DeepBase short-term, FlowNet reaches MAE 0.43 and RMSE 0.93, versus 0.49/1.02 for GWNET.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FlowNet instead of the strongest reported baseline | Spatio-temporal forecasting · PEMS04F, DeepBase, and SINPA short- and long-term tasks | MAE | Helped |
| Add retained flow tokens (ablation) | FlowNet · PEMS04F short- and long-term forecasting | MAE: 0.08; 1.32 | Helped |
| Add retained flow tokens (ablation) | FlowNet · PEMS04F short- and long-term forecasting | RMSE: 0.13; 1.74 | Helped |
| Add allocation flow tokens for inter-node exchange (ablation) | FlowNet · PEMS04F short- and long-term forecasting | MAE: 5.54; 4.90 | Helped |
| Add allocation flow tokens for inter-node exchange (ablation) | FlowNet · PEMS04F short- and long-term forecasting | RMSE: 7.91; 7.26 | Helped |
| Enforce conservation during flow exchange (ablation) | FlowNet · PEMS04F short-term forecasting | MAE: 0.24 | Helped |
| Enforce conservation during flow exchange (ablation) | FlowNet · PEMS04F short-term forecasting | RMSE: 0.33 | Helped |
| Enforce conservation during flow exchange (ablation) | FlowNet · PEMS04F long-term forecasting | MAE: 0.83 | Helped |
| Enforce conservation during flow exchange (ablation) | FlowNet · PEMS04F long-term forecasting | RMSE: 1.44 | Helped |
Surprising findings
- Removing the retained-flow term has the least impact on short-term forecasting but a larger impact on long-term forecasting.
- Breaking the conservation law hurts short-term forecasting more than long-term forecasting.
Limitations
- Flow allocation incurs O(N^2) complexity, limiting scalability to large node counts.
- Less efficient than STGNN-based methods for long-term forecasting and large systems.
- Evaluated on only three datasets, with no theoretical results.
Experimental setup
Datasets: PEMS04F DeepBase SINPA
Benchmarks: PEMS04F DeepBase SINPA
Metrics: MAE RMSE
Baselines: Autoformer PatchTST Crossformer iTransformer FEDformer STGCN GWNET SCINet STTN STAEformer
Builds on: Transformer PatchTST hyper-connections Tobler's First Law of Geography
Closest papers by meaning
- Analytical Construction on Geometric Architectures: Transitioning from Static to Temporal Link Prediction (ICML 2025 · similarity 0.84)
- Structural Inference of Dynamical Systems with Conjoined State Space Models (NeurIPS 2024 · similarity 0.84)
- Topology-Aware Conformal Prediction for Stream Networks (NeurIPS 2025 · similarity 0.83)
- Rethink GraphODE Generalization within Coupled Dynamical System (ICML 2025 · similarity 0.83)
- Graph Mamba Operator: A Latent Simulator for Interacting Particle Systems (NeurIPS 2026 · similarity 0.83)
- Neural Persistence Dynamics (NeurIPS 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.