One of the often-ignored challenges of digital transformation is the impact on Greenhouse Gas Emissions (GHG) and energy consumption, especially in an ESG environment. When dealing with abstract concepts such as ‘the cloud,’ or when systems are out of sight, out of mind, it is easy to forget that the physical data centers that make up ‘the cloud’ consume well over 2% of electrical energy in the USA and have surpassed total equivalent emissions of the entire aviation industry. The present actual value is no longer available as the major data center owners no longer report energy consumption. The location and makeup of energy supplying the datacenter has an impact with present estimate range from 3.1 kWh per gigabyte of storage (ACEEE) to 7 kWh (Carnegie Mellon University). The development, training and operation of AI and Machine Learning also has an impact with a direct report by Microsoft (The Carbon Footprint of AI, October 26, 2020, Microsoft) as 626,000 lbs CO2 for a small AI system. Small machine learning model development takes an average of 65,000 lbs CO2 in the cloud and 35 lbs CO2/hour to run – per model, in addition to any storage.
On the other hand, the same reports identify that local storage on a PC or laptop takes 0.000005 kWh per gigabyte (GB). The training of a similar model on a local system, with the national average of 954 lbs-CO2/kWh and assuming 5kWh for the data center would be ~14 hours. If we take my development laptop, a training system would use negligible energy (< 1 gigabyte) for storage and total energy under full loading was measured at 34 watts (0.034 kWh), or 0.48 lbs CO2 for 14 hours.
When we begin to look at IoT applications, what is the actual energy and emission usage? Is there a focus on data to the cloud, software on the cloud, and data control? Is there research into the data center location utilized for energy and emissions? When we took a look at how we would evolve the ECMS continuous monitoring system these were concerns. It was challenging to determine the overall emissions and kWh, but we did the research – and it was difficult enough that it is not unreasonable to assume that most IoT companies do not consider this outside of storage and cloud use costs.
As systems expanded, the original systems ECMS-E1 and ECMS-32, were primarily run by the customer/user internal servers or cloud servers. Based on work that was initiated with Onics Energy we determined that the average on-site server was drawing 100 Watts plus the use of a cell modem (2 watts) then cloud communication and storage. This was an edge system, but pretty much overkill. As we expanded wind turbine ECMS-E1s we moved to fanless industrial edge computers with measured demand under full load at less than 15 watts plus the ECMS system at 2.5 watts and cell modem at 2 watts. This included edge storage, operating the expert software, and cloud-based storage and anomaly AI, resulting in (based on datacenter location) <4 kWh/GB. Without backup, the amount of data for a single channel (ECMS-E1) is 100 MB full datasets, or 3200 MB (3.2 GB) for a 32-channel ECMS system. Assuming a dataset per hour over 8,760 hours per year, the local system would consume 170 kWh regardless of the number of channels. The cloud system would be 3,500 kWh for a single channel and 112,130 kWh for a loaded ECMS-32. Usage is being rounded upwards for purposes of identifying maximum impact. Actual impact is much smaller based on recommended use strategies.
The GHG emissions associated with the ECMS-E1 system would be 3,670 kWh/machine and 2.6 Tonnes CO2 per year. The ECMS-32 would result in 79.4 Tonnes CO2, or 2.5 Tonnes CO2/machine per year. In a recent installation we applied one ECMS-E1 and 3 fully developed ECMS-32 (96 machines monitored) resulting in emissions of ~240 Tonnes CO2. Of course, this is worst-case with changes to the amount of data stored and frequency of testing. On-site storage and analysis results in less than 25 Tonnes CO2 for the same application.
The trade-off is the detection and correction of found defects or waste such as 42 kW loss in a medium-sized belted application resulting in 260 Tonnes CO2 improvement through simple corrections. When we add in lubrication detection, alignment issues, soft foot, driven equipment problems, operations improvements, etc. the simple GHG payback is usually measured in a week or so. However, the approach does require that the maintenance program is incorporated as part of ESG goals and/or the energy management program.
A different choice would have been a complete move of software to a cloud-based machine learning system. If we go from the baseline values provided in the articles mentioned at the beginning of this article, or assume a lighter value for a learning period of a week (168 hours) at 4 kWh/GB at 2.5 GB of data for training resulting in 1,680 kWh and 1.2 Tonnes CO2 per system (2,645 lbs CO2). If storage is maintained and grows over the course of a year, including data used for training future models, the value of that use continues to increase. However, if we assume just the 2.5 GB, then storage would result in 39.7 tonnes and the machine learning model would emit 136 tonnes CO2 for a grand total of 176 tonnes. This can be curbed through the use of edge computing in order to reduce the associated carbon footprint with IoT devices or reducing the model operating time, as well as a review and selection of site and types of machines (servers) being used.
Strategies can be applied to reduce the machine learning model operating time, such as only running the model at a longer frequency, such as once per hour, once per day, or other. Most cloud-based offerings are actively reducing their footprint based on policy and public pressure and will also provide some type of tool for measuring and monitoring your usage. As we are involved in a growing number of ESG programs, MotorDoc has been actively involved in monitoring and applying a variety of approaches to streamline our systems, development, and partner development.
Finally, understanding the impact, let alone costs, of operating, storing and maintaining data, and running machine learning/AI models for predictive maintenance and other digital transformation, especially with corporate ESG goals, is critical. This should be a consideration as you review systems, select tactics, and determine usage of IoT systems. Presently, over 90% of information stored is not used, and that includes IoT device data. When an IoT system is used, and is developed for efficiency, then the results should be a rapid payback in both local energy and, more importantly, cloud energy and emissions. It is equally important to remember that while cloud storage moves IT responsibility, resources and hardware off-site, it is dramatically more ‘expensive’ from an energy and emissions perspective.
References and Resources:
The Carbon Footprint Of AI – Sustainable Software (microsoft.com)
AI and climate change: The mixed impact of machine learning | TechTarget
