Data Center Cooling Control Using Causal Signal Injection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning techniques for data center energy management require extensive data and time for training, leading to short periods of optimized operations and potential 'catastrophic forgetting' due to infrastructure changes.
Innovation Solution
The method involves injecting randomized controlled signals into data center operational controls, monitoring responses, and computing confidence intervals to select optimal signals for improving energy efficiency, thereby enabling faster and more robust learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If advanced machine learning techniques such as convolutional neural networks are applied to develop predictive models, then energy reduction for cooling is achieved, but data requirements and training time increase significantly
Solution Approach 1:
A digital twin is introduced as an intermediary virtual model that simulates data center cooling systems. This digital twin generates synthetic training data and enables predictive modeling without requiring extensive real-world operational data, thus reducing the data requirement while achieving energy optimization goals.
Solution Approach 2:
The system performs preliminary actions by pre-training machine learning models using synthetic data generated from the digital twin before deployment. This pre-training phase prepares the models in advance with sufficient training data, eliminating the need for extensive real-world data collection and long training periods during actual operation.
2Measurement precision
If machine learning models are trained on years of operational data, then model accuracy is improved, but the system becomes vulnerable to catastrophic forgetting when infrastructure changes occur
Solution Approach 1:
The system implements dynamic model updating where the machine learning model is continuously retrained and adapted as infrastructure changes occur. Rather than relying on static historical data, the model dynamically adjusts to new conditions, maintaining accuracy while adapting to equipment refreshes and infrastructure modifications without catastrophic forgetting.
Solution Approach 2:
A feedback mechanism is established where the system continuously monitors infrastructure changes and performance metrics, then uses this feedback to trigger model retraining or adjustment. This closed-loop feedback ensures the model remains accurate and adaptable by incorporating new information from changing infrastructure conditions.
3Productivity
If computer equipment is refreshed every three years, then system performance is improved, but machine learning model accuracy degrades requiring re-training
Solution Approach 1:
The system performs preliminary actions by pre-training models with diverse synthetic data that covers potential future infrastructure configurations. This preparation in advance allows the model to be more resilient to equipment refreshes, reducing the frequency and impact of required re-training events.
Solution Approach 2:
The system implements continuous feedback monitoring that detects when infrastructure changes occur during the three-year refresh cycle. When changes are detected, the feedback mechanism triggers targeted model adjustments or re-training, maintaining accuracy without requiring full model retraining regardless of infrastructure changes.
Data Source
Figure 1A
Figure 1B
Figure 2
AI summary
Methods for active data center management by injecting randomized controlled signals in the operational controls of the cooling infrastructure of the data center and receiving response signals corresponding with the injected signals. The injected signals are used to adjust the operational controls of the cooling infrastructure, and the response signals relate to the operational conditions in the data center. Based upon the response signals along with independent and external variables, the randomized signals are continually injected into the cooling infrastructure and fine-tuned based upon the response signals. Optimum or improved parameters for controlling the cooling infrastructure of the data center are determined based upon the signal injections and corresponding responses.