Author: Ali Rajaei
Affiliation: Delft-AI Energy Lab, Department of Electrical Sustainable Energy, Delft University of Technology, the Netherlands
Contact: a.rajaei@tudelft.nl
Date: April 2025
This repository accompanies the research paper:
Rajaei, Ali, Olayiwola Arowolo, and Jochen L. Cremer.
"Learning-Accelerated ADMM for Stochastic Power System Scheduling with Numerous Scenarios."
IEEE Transactions on Sustainable Energy, 2025.
The increasing share of uncertain renewable energy sources (RES) in power systems necessitates new efficient approaches for the two-stage stochastic multi-period AC optimal power flow (St-MP-OPF) optimization. The computational complexity of St-MP-OPF, particularly with AC constraints, grows exponentially with the number of uncertainty scenarios and the time horizon. This complexity poses significant challenges for large-scale transmission systems that require numerous scenarios to capture RES stochasticities.
This paper introduces a scenario-based decomposition of the St-MP-OPF based on the alternating direction method of multipliers (ADMM). Additionally, it proposes a machine learning-accelerated ADMM approach (ADMM-ML), facilitating rapid and parallel computations of numerous scenarios with extended time horizons. Within this approach, a recurrent neural network approximates the ADMM sub-problem optimization and predicts wait-and-see decisions for uncertainty scenarios, while a master optimization determines here-and-now decisions. A hybrid approach is also developed, which uses ML predictions to warm-start the ADMM algorithm, combining the computational efficiency of ML with the feasibility and optimality guarantees of optimization methods.
The numerical results on the 118-bus and 1354-bus systems show that the proposed ADMM-ML approach solves the St-MP-OPF with 3β4 orders of magnitude speed-ups, while the hybrid approach provides a balance between speed-ups and optimality.
This repository contains:
- β Pyomo-Gurobi implementation of the Stochastic AC multi-period OPF
- β Scenario-based decomposition using ADMM
- β ML model for approximating ADMM subproblems
- β Training and evaluation pipelines for ADMM-ML and hybrid solutions
- β
Training data generation with
$\epsilon$ -greedy exploration