基于DQN算法的泵站供水系统节能控制优化
Optimization of Energy-Saving Control for Pumping Station Water Supply System Based on DQN Algorithm
陈财会 1张天 1黄健康 1金典 1王卓悦 1张小磊1
作者信息
- 1. 哈尔滨工业大学<深圳>土木与环境工程学院,深圳 518055
- 折叠
摘要
针对手动调节泵站中水泵运行的转速和启停会造成严重的能量浪费问题,引入基于深度Q网络(deep Q-learning network,DQN)的强化学习算法,通过获取当前泵组运行的状态,自动优化水泵组工作时各个水泵的运行参数,在各个水泵均处于高效区的前提下,提高水泵组的整体效率.对水泵组状态优化问题分别进行了数学描述和马尔可夫决策过程描述.同时定义了水泵组运行时的状态空间、动作空间和即时奖励值,构建DQN网络,并以深圳市M水厂为算例,在由Gym构建的自定义仿真环境中进行验证.相较于人工调控,DQN算法调控降低了 8.84%的损失能耗,一年可节省吨水电耗达 1.27×10-2 kW·h/t,实现了节能减排,具有良好的经济效能.同时,DQN算法可通过在线学习的方式适应供水环境的变化,具有自主性、实时性、可推广性等优点.
Abstract
To address the issue of significant energy waste caused by manually adjusting the running speed and start-stop of the pumps in the pump station,the deep Q-learning network(DQN)algorithm was introduced to automatically optimize the operation of each pump during the operation of the pump unit by obtaining the current operating status.Operating parameters,on the premise that each pump was in the high-efficiency area,improved the overall efficiency of the pump unit.The state optimization problem of the pump unit was described mathematically and using the Markov decision process.At the same time,the state,action and reward value of the pump unit were defined,and a DQN network was established.As an example,Shenzhen M WTP was validated verified in a custom simulation environment built by Gym.Compared to manual regulation,DQN algorithm regulation reduced the loss of energy consumption by 8.84%,and could save electricity consumption per tons of water by 1.27×10-2 kW·h/t a year,which contributing to energy saving,emission reduction,and good economic.At the same time,the DQN algorithm,which had the advantages of autonomy,real-time,and generalizability,could adapt to changes in the water supply conditions through online learning.
关键词
泵站供水/优化调度/DQN算法/马尔可夫决策过程/节能减排Key words
pumping station water supply/optimal controling/DQN algorithm/Markov decision process/energy saving and emission reduction引用本文复制引用
基金项目
深圳市可持续发展科技专项(KCXFZ20201221173602008)
哈尔滨工业大学(深圳)课程教学项目(HITSZERP21003)
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据