In a recent project request I was asked to evaluate an ongoing study. When I started into the work I quickly realized why I was asked to look at the project. I am going to be vague as this was not unusual in every kind of project we’ve seen: the scope of the project was global requiring high level inputs and outputs of the system. Instead, the detailed components and unrelated systems, and subsystems, were obtained. This generated a lot of work, but the original scope of the project could not reasonably be achieved. The production data, which was available but not requested, was not used. Instead, global assumptions and anecdotal data was used in place of real data.
Once we have the data then the project will take several days with real projections that can be defended versus months of work to date. Yep, losing sight of the scope can get very costly.
But how do you know? Not my first rodeo. While the nitty-gritty might be useful to find anomalies and reasons for outliers, but that would be once the actual useful data is reviewed.
This concept impacts everywhere, especially with the use of machine learning/AI (data science) tools and monitoring systems. We’ve observed as correllations are cited as causation when the data is coincidental and then used as part of monitoring instead of human-in-the-loop evaluation and verification by subject matter experts (SME). We’ve even observed synthetic data used to train models with assumptions and forced correllations then everyone is surprised when they crash and burn in the field.
So, let’s say we are looking for the impact of a global energy, emissions and waste stream project for a company site. What data do you require?
Let’s take it like it’s a Reliability-Centered Maintenance (RCM) project: the first objective is to figure the scope of the equipment – in this case a facility. The next part would be to identify all inputs and outputs.
For the scope of the project you will identify inputs such as energy and materials, which might include water, etc., then you observe all outputs such as product and waste (by-products). So, let’s make this simple and say the scope is a global electrical energy improvement in a manufacturing plant. The plant manufactures material (widgets) measured in tons with several by-products that are used as feeders to other processes and products elsewhere. Materials and chemicals are brought in. There is a single meter (usually multiple) for energy consumption.
Production is able to provide the inputs, the energy consumption, and the outputs. First, the inputs and outputs MUST BE BALANCED. There is always conservation of matter, with some variability. This can include the primary product, by-products and general waste. Anything that is kept in the loop (ie: water or solution recycled) does not need to be considered. What each piece of equipment in the plant does is not considered unless there is some anomaly that crashes the model. All units must be converted to the same values, in this case tons of material.
You will also have to understand the cycle-time for matching energy consumption to the processing of materials, even with continuous processes. This is especially important when the incoming materials have variable quality and result in variations in energy use to convert to the various outputs. In these cases you will also note that there are variations in tons in to tons out of primary product and each by-product.
You then look at environmental variables and confirm whether or not energy consumption has some level of correllation. If it does have a reasonable potential impact, then you include it. Often the granularity for production, energy and environmental data is performed by 24-hour day, but you can go deeper. However, in most instances I’ve found that per-day matches incremental meter data, production flow, and output monitoring for most process manufacturing. I don’t require the granularity of each component of the system, which will generate noise.
You then calculate the consumption of electricity in kWh and compare it to the production output. The result is kWh/unit of production. Now, there are other complexities such as outages, baseline, efficiencies at different production levels, etc. That is the data sciency stuff.
You then create a simple regression model and test it using something like a 70-30 train/test split. This is where you generate or compare regression models to see which one fits the best to known data. If the variations are too great, then you spend the time to research what you might have missed. You then have a way of producing a counter-factual method of comparing what energy consumption would have occurred using that conversion from materials to products/by-products with the variables of inputs and, in this example, temperature. Then the difference between the model values to real values would be the impact.
I can do this for reliability and maintenance KPI, energy projects, and the ideas for other projects including in an RCFA, etc.
In a study that we did on a biodiesel facility, we identified inputs and outputs. The lost material in-between did not make any sense and, when pressed, we found out that a portion of by-products that made up 25-35% of outputs were excluded because someone in the organization decided we didn’t need the information. We dove in, before knowing that information was being withheld, and spent hundreds of hours trying to figure out why no model worked.
One example can be found in a study performed and published in NETA, 2023: Neutral and Ground Harmonic-Induced Power Losses and Correction – NETAWORLD JOURNAL
“A Novel Approach to Industrial Assessments for Improved Energy, Waste Stream, Process and Reliability,” Howard Penrose, 199: final
About The Author
