L-IPM: Learning Interior-Point Central-Path Projection for OPF

Code and trained models for "Learning Interior Point Method Central Path Projection for Optimal Power Flow". L-IPM runs MATPOWER's MIPS interior-point solver for the first K = 3 Newton iterations, lets a grid-informed LSTM (GI-LSTM) project the iterate toward the converged solution, screens the projection for feasibility, and lets MIPS finish from the projected point.

Requirements

• MATLAB (tested with R2023b) and MATPOWER 7.0.
• Python 3.11 with tensorflow==2.21, numpy, scipy, pandas, h5py, openpyxl, matplotlib (e.g. conda create -n lipm python=3.11 && conda activate lipm && pip install tensorflow==2.21 numpy scipy pandas h5py openpyxl matplotlib).
• All MATLAB scripts locate the repository through mfilename('fullpath'); the only machine-specific setting is the Python interpreter, read from the environment variable LIPM_PYTHON (e.g. C:\...\envs\lipm\python.exe) or edited at the top of the script.

What is in the folder

| lib/, mips/ | MATPOWER 7.0 library and MIPS solver. | | mips.m | logs every iterate through ipm_logger, contains the L-IPM hook after iteration K, and records per-iteration wall time. | | ipm_logger.m | Collects the IPM iterates during data generation and during L-IPM runs. | | lipm_project.m | The hook body: sends the first K iterates and the load vector to Python, receives the projected state, applies the acceptability screen, and either replaces the MIPS iterate or falls back to classical IPM. | | lipm_infer.py | Python side of the bridge: loads the trained GI-LSTM once, standardizes inputs, runs the forward pass, clips/floors the output, and returns the screen residuals. Set SYSTEM, OPF, MODEL_FILE at the top to pick the model. | | case3.m, case24_ieee_rts.m, case118.m, case2869pegase.m | Test systems (MATPOWER format). | | Data_Generation.m | Generates the training data: Latin-hypercube load scenarios (seed 42), classical MIPS solves, first iterates + converged solution per scenario. | | export_case_data.m | Writes <system>_<opf>_meta.mat (network data and MIPS variable layout) needed by the Python code. | | gi_lstm_<system>_<opf>.py | GI-LSTM training script for one system/formulation; running it with no arguments reproduces the deployed model (gilstm_<system>_<opf>/). | | gilstm_<system>_<opf>/ | Trained model: gi_lstm.weights.h5, scalers.npz, results.json (metrics, architecture, training time), <tag>_lstm_pred.mat (test-split predictions and run ids). | | IPM_vs_LIPM_24bus.m | Classical IPM vs L-IPM comparison. | | timeing_periter.m | Wall time of every Newton iteration of classical MIPS (per-iteration timing table). | | bench_infer.py | Pure GI-LSTM inference latency (graph-compiled forward pass, median of 300 calls). |