Physics-Informed Implicit Representations of Equilibrium Network Flows
Authors: Kevin D. Smith, Francesco Seccamonte, Ananthram Swami, Francesco Bullo
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | numerical experiments from several AC power networks and water distribution systems that indicate IFN can significantly outperform these baselines on the flow estimation task.5 Numerical Experiments We studied the transductive task of predicting unlabeled flows, given that some labeled flows in the same network are known. |
| Researcher Affiliation | Collaboration | Center for Control, Dynamical Systems, and Computation, University of California, Santa Barbara, CA 93106 USA, {kevinsmith, fseccamonte, bullo}@ucsb.edu U.S. Army Research Laboratory, Adelphi, MD 20783 USA, ananthram.swami.civ@army.mil |
| Pseudocode | Yes | Algorithm 1 Evaluating the implicit flow network. |
| Open Source Code | Yes | Code is available at https://github.com/Kevin Daly Smith/implicit-flow-networks. |
| Open Datasets | Yes | AC Power We selected 6 standard power network test cases... We used the MATPOWER toolbox [36] to solve the power flow equations... Water Distribution We selected 3 sample water distribution networks from the ASCE Task Committee on Research Databases for Water Distribution Systems database [37], representing municipal water distribution systems in Fairfield, CA, Bellingham, WA, and Harrisburg, PA. Each network contains the topology of the distribution system, as well as the characteristics of pipes and other network elements and nodal demands. We used the WNTR package [38] to compute the flow rates through each pipe (f), net inflow rate at each node (u), and edge weights associated with each pipe. |
| Dataset Splits | No | For each network, we randomly selected a fraction of the edges to be labeled edges, and we trained IFN and baselines on the labeled edges. Then we evaluated the RMSE of the flows predicted for the unlabeled edges Eu to compute the testing error. See Appendix B in the supplementary material for full details. The text implies training and testing splits but doesn't clearly mention a separate validation set. |
| Hardware Specification | No | The paper does not provide specific hardware details (exact GPU/CPU models, processor types, or memory amounts) used for running its experiments. |
| Software Dependencies | No | We used the MATPOWER toolbox [36] to solve the power flow equations, then recorded the active power flows on each branch (f), computed the net active power injections at each node (u), and selected relevant electrical parameters as edge attributes (series reactance, tap ratio, and voltage magnitude at the two incident nodes). and We used the WNTR package [38] to compute the flow rates through each pipe (f), net inflow rate at each node (u), and edge weights associated with each pipe. (No version numbers specified). |
| Experiment Setup | Yes | For both water and power, the IFN layer uses a derivative-constrained perceptron as the inverse flow function (k = 128, p = q = 1 2) with a Re LU activation function. For power, we set dmin = 0.4 and dmax = 2; and for water, dmin = 0.2 and dmax = 20. and In Bil-MLP and Bil-GCN, q is the output of either a 2-layer MLP or GCN model with edge attributes as inputs (we use 64 nodes in each hidden layer with Re LU activations). |