June 07, 2023
In the summer of 2019, a study was carried out in search of optimization opportunities in oilfield service scheduling and logistics.
An experimental framework was developed involving three water haulage providers—one large, one midsize, and one small, family run operation. Within each company, the study controlled for human incentives by paying some team members on an hourly basis, others by the job, and finally some by the barrel.
The study also examined the use of oilfield SCADA data to help predict demand and pre-locate logistics assets in a proactive manner, within various physical constraints—such as equipment availability and human resource availability by shift and location
Information was compiled over 3 months, producing a unique, well-organized, and well-controlled data set.
A team of data scientists was engaged to scour the data for patterns and insights. In a gorgeous conference room in Houston—removed from their usual corporate cubbies and the associated blinders—they began interrogating the data.
The team explored a wide range of factors, including location dimensions, loading and discharging operations, shift characteristics, time of day, traffic congestion, local curfews, human incentives, and the impact of operational safety protocols.
An early hypothesis linked scheduling decisions to 'tribal knowledge' within operators brains about routing options, relative to their home or yard location.
There was palpable excitement when the data supported this and other ideas that were bubbling up within the team.
In line with good experimental practice, the team tested its pattern-derived models on different data subsets, operational clusters, and individual transactions to see if they could accurately predict individual logistic movements.
To the team’s disappointment, no reliable predictions could be made. The model was spot-on for certain movements and way off on others. The results seemed no better than random.
Many cups of coffee later—after the finger pointing had died down—the team decided to conduct more extensive field investigation (led by the author of this post).
Over the course of several months—after completing the requisite field safety protocols—visits were made to various environmental service facilities in the Haynesville, Permian, Marcellus, Utica, and Bakken basins.
Observations were made throughout 5,000 miles of water truck ride-along journeys, and while an estimated 10,000 truck miles were being dispatched—including produced water, freshwater, frac water, vacuum services, and other service work.
The author observed an essential humanity across the drivers, sharing stories about their lives and careers during long hours on the road. Most were older men but a few women and drivers in their early 20s showed the vocation of field work crossing demographic lines.
Day shifts started at 5am, while night shifts began at 5pm. Safety rules require drivers to take a 30-minute break within the first 8 hours of their shift. This was not reflected in the study data—an early indicator that the realities of field operations would need to be captured in greater detail if the team was to improve its predictions.
Observing field dispatchers highlighted flow rates and tank levels dictating the prioritization and scheduling of drivers. A few 'hot’ jobs were given to drivers as their first task on a shift (reinforced by text messages and calls) but, after that, drivers were generally left to decide their own route and schedule for completing the remainder of the tasks assigned to them by the dispatcher.
Sometimes a driver wouldn't be able to finish all the trips, other times they would call in and take on an extra task when they would otherwise be finished early.
The data-driven model should have been able to predict these outcomes because each trip was tracked in real time and the model could use Google Maps to compute the expected end of one job and start of the next. Why the model was failing to do so remained a mystery.
As it turned out, the answer lay in the 'squawk box' that lives in each truck’s cab.
Drivers communicate with each other about traffic conditions, politics, sports, law enforcement activities, and—most importantly—free food!
It was easy to tune out the radio chatter while riding alongside drivers, but that was a mistake.
An important source of scheduling randomness was introduced by the variance in donut availability at one disposal site versus pizza being served at a yard, or the cookies, coffee, chicken wings, or barbecue being dished up at a particular truck stop.
The human element introduced by birthday celebrations, anniversaries, awards, or just someone feeling generous made a meaningful difference to how a driver managed their route around high priority 'hot’ jobs.
The team found cases where the sub-optimal routing impact was dramatic. The driver’s mental 'algorithm' was strongly influenced by the combined stimulus of an upcoming compulsory break and the availability of free food. It even overrode personal objectives such as attending a kid’s recital or making it to a doctor’s appointment.
Small wonder the model had a hard time predicting routing decisions on a day-to-day basis.
Back in the conference room, the team debated how to align data science with the reality in the field.
In theory, economic incentives such as per-barrel pay should have ensured “perfect” alignment between the dispatcher, driver, and well site data.
In practice, the team learned how powerful tribal knowledge can be when trying to match data science methods with human behavior. There are limits to economic optimization in the real world.
It's hard enough to get younger workers to take up a field service career, away from urban areas and requiring a lot of training and practice to become safe and effective.
Optimization efforts that dictate routes and reduce job satisfaction have the potential to turn newer employees away from the job. Excessive scheduling instructions might even cause them to tune out the radio or have a negative impact on their alertness and safety.
The study team chose to pursue fleet-level optimization instead, and the early results are encouraging.
With new challenges such as renewable fuels, in-field refueling options, changes to safety regulations, and even the possibility of autonomous trucks all converging, there are many layers of scheduling optimization opportunities to pursue.
And that's just in trucking.
When you consider ongoing improvements in the vessel, rail, and pipeline logistics that connect to those trucks—including the challenge of breaking bulk goods for last mile distribution—the opportunity for data science to positively impact profitability and decarbonization is immense.
We must, however, remember that some of us love chicken wings while others love donuts—and those differences really matter.