Multi-chiller scheduling using reinforcement learning with transfer learning for power consumption prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing chiller scheduling strategies fail to consider the impact of one control strategy on another and neglect additional parameters like modulating chiller temperatures, leading to inefficient power consumption in multiple-chiller plants.
Innovation Solution
Implementing reinforcement learning with transfer learning to schedule chiller ON/OFF sequencing and leaving chilled water temperature (LCWT) scheduling, using a pre-trained transfer learning model and deep reinforcement learning agent to predict power consumption and optimize chiller operation based on design parameters, leaving and returned chilled water temperatures, and ambient temperature.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If reinforcement learning with transfer learning is implemented for multi-chiller scheduling, then power consumption is minimized and energy efficiency is improved, but system complexity and computational requirements increase
Solution Approach 1:
The transfer learning model is pre-trained on source chiller data before deployment to the target chiller. This preliminary training action allows the model to learn general patterns from multiple chillers, and then adapt to the specific target chiller with limited data, reducing the overall computational complexity and data requirements while achieving optimal power consumption minimization
Solution Approach 2:
The system changes the parameter representation by using state vectors that combine multiple chiller parameters (temperatures, loads, binary states) into a unified state representation. This parameter transformation enables the reinforcement learning agent to process complex multi-chiller system states efficiently, balancing the trade-off between capturing system complexity and minimizing computational burden
2Productivity
If deep reinforcement learning agent is used for chiller sequencing and temperature modulation, then chiller operation efficiency is optimized, but computational resources and training time increase
Solution Approach 1:
The deep reinforcement learning agent is trained in advance using simulated environments and transfer learning from source chillers. This preliminary training action allows the agent to learn optimal control policies before actual deployment, reducing the need for extensive real-time training and enabling efficient chiller operation from the start
Solution Approach 2:
The system creates a simulated copy of the chiller plant environment for training the reinforcement learning agent. This virtual copy allows extensive training to occur without affecting real chiller operation, separating the training time burden from actual operational time while maintaining high operational efficiency upon deployment
3Loss of energy
If multiple control parameters including leaving chilled water temperature modulation are considered, then energy optimization is improved, but control strategy complexity increases
Solution Approach 1:
The system merges multiple control decisions (chiller ON/OFF sequencing and leaving chilled water temperature modulation) into a single unified reinforcement learning control framework. This consolidation allows the agent to optimize both parameters simultaneously based on the same state representation and reward function, improving energy optimization while managing complexity through integration rather than separate control loops
Solution Approach 2:
The system transforms the control strategy by representing multiple control parameters (chiller states, temperatures, loads) as a unified state vector. This parameter transformation allows the reinforcement learning agent to process and optimize multiple control variables simultaneously, achieving better energy optimization without proportionally increasing control complexity
Data Source
AI summary
Conventionally, chiller power consumption has been optimized by using a Cooling Load based Control (CLC) approach which does not consider the impact of one control strategy on the other. Embodiments of the present disclosure provide reinforcement learning (RL) based control strategy to perform both chiller ON/OFF sequencing as well as setpoint leaving chilled water temperature (LCWT) scheduling. A RL agent is trained using a re-trained transfer learning (TL) model and LCWT, return chilled water temperature of target chillers and ambient temperature of building are read for determining required cooling load to be provided by target chiller(s) based on which the target chillers are scheduled for turning ON/OFF. Transfer learning-based approach is implemented by present disclosure to predict power consumed by a chiller at some setpoint by using a model trained on similar chillers which were operated at that setpoint since the chillers are usually run at a single setpoint.