Scholarlink

Managing gas grid infrastructure requires real-time control decisions that balance multiple competing objectives while maintaining safety constraints. In this work-in-progress/lessons-learned paper, we present lessons learned from pilot deployments of a multi-objective reinforcement learning approach for gas grid inflow control, using Generalized Policy Improvement - Prioritized Dyna (GPI-PD) to maintain a diverse set of policies for operator selection. Our system leverages operational time series data from past gas grid operations to train ML regressors capable of predicting the next-timestep \(s_{t+1}\) based on the current state \(s_t\) and action \(a_t\) , where \(s_t\) denotes the gasgrid’s current condition (e.g. linepack and pressure) and \(a_t\) the control action (e.g., inflow adjustments) at time t. The GPI-PD algorithm maintains multiple policies optimized for different objective weightings, allowing operators to select appropriate control strategies based on current operational priorities and constraints using the industry-proven multi-criteria decision-making framework Qualicision AI. Through real-world pilot testing, we identify key challenges in deploying RL for safety-critical infrastructure, including labeling efforts, simulation scalability, policy interpretability requirements, and the importance of maintaining operator trust through transparent multi-policy selection. Our results highlight practical considerations for future deployments of RL in critical infrastructure management.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-objective Reinforcement Learning for Safety-Critical Gas Grid Management: Lessons from Real-World Pilot Deployment

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Alexander Schiendorfer,
  • Thomas Schmidt,
  • Karima Outafraout,
  • Michelle Baschin,
  • Yuguang Hei,
  • Tom Streubel,
  • Pascal Kätzel,
  • Lilia Michailov,
  • Rudolf Felix,
  • Lars Harpeng

摘要

Managing gas grid infrastructure requires real-time control decisions that balance multiple competing objectives while maintaining safety constraints. In this work-in-progress/lessons-learned paper, we present lessons learned from pilot deployments of a multi-objective reinforcement learning approach for gas grid inflow control, using Generalized Policy Improvement - Prioritized Dyna (GPI-PD) to maintain a diverse set of policies for operator selection. Our system leverages operational time series data from past gas grid operations to train ML regressors capable of predicting the next-timestep \(s_{t+1}\) based on the current state \(s_t\) and action \(a_t\) , where \(s_t\) denotes the gasgrid’s current condition (e.g. linepack and pressure) and \(a_t\) the control action (e.g., inflow adjustments) at time t. The GPI-PD algorithm maintains multiple policies optimized for different objective weightings, allowing operators to select appropriate control strategies based on current operational priorities and constraints using the industry-proven multi-criteria decision-making framework Qualicision AI. Through real-world pilot testing, we identify key challenges in deploying RL for safety-critical infrastructure, including labeling efforts, simulation scalability, policy interpretability requirements, and the importance of maintaining operator trust through transparent multi-policy selection. Our results highlight practical considerations for future deployments of RL in critical infrastructure management.