Average Load Is the Wrong Input for Choosing a Cell Chemistry
Average load is the wrong input for choosing a cell chemistry because a chemistry is selected by the worst short window in the duty cycle, not by the mean over the whole route. Three numbers do the work: the highest C-rate the pack will actually deliver, how long that peak lasts, and how often it comes back. Getting all three means binning a recorded current trace into a C-rate histogram a supplier can quote against, instead of handing over an average.
I have sat at more than one supplier conference table with a log like that open on my laptop. One of them was 47 minutes of machine data pulled off the CAN bus, a time column and a current column, roughly 28,000 rows, the fan running because the spreadsheet was heavy. The customer's spec said 0.6C continuous. That number was honest. It was also about to put the wrong chemistry on the shortlist.
An average does not care about shape. Forty-seven minutes of mostly 0.4C work with one twelve-second 3.8C climb in the middle averages to something near 0.6C. Nobody lied to anybody. The average answers how much energy the machine used. It does not answer how much current the cells have to deliver, and those are different questions with different units.
The numbers I use below are round numbers chosen to show the shape of the method, not measurements from a specific machine. Yours will come out of your own log.
Bin the trace in amps, then draw C-rate lines on top
The first move is counterintuitive if you have spent years reading datasheets. Do not bin in C-rate. Bin in amps.
Amps are set by the load. The motor, the inverter, the hill, the payload, the operator with his foot down. C-rate is amps divided by pack capacity, and capacity is the thing you have not chosen yet. If you bin in C-rate assuming a 20Ah pack and the shortlist moves to a 30Ah pack, every boundary in your histogram is wrong and you start over. Bin the measured current into fixed amp bins and the histogram becomes a property of the application. Then lay 1C, 2C, 3C lines across it for each candidate capacity. The picture does not change. The lines move.
A few practical details matter more than they sound.
Sample rate. A 1 Hz log will not show you a two-second peak. If the only thing you have is a BMS historian, check the rate before you trust the shape, because BMS logs are frequently filtered or averaged, which is exactly the thing you are trying to get away from. Where you can, pull the raw current channel from the inverter or the motor controller at its native rate.
Which current. Pack current, sign-carrying, with regen negative. Motor current is not pack current. DC-link current is neither.
Bin width. Wide enough to read at a glance, narrow enough that the top bin is not hiding a cliff. Half-C bins usually work once you have overlaid the candidate capacities. What matters more than width is that each bin carries a maximum dwell time and a count of excursions above threshold, not just time-in-bin.
Duration is what actually picks the chemistry
A histogram with time-in-bin alone is half a picture. Two duty cycles can produce identical histograms and want different chemistries, depending on whether the high bin is a thin spike or a plateau.
So each high bin needs three companions: the longest continuous dwell inside it, how many times per mission the pack crosses into it, and how much recovery time sits between crossings.
Eight seconds at 3.5C, six times in a 45-minute mission, with minutes of 0.3C in between, is a thermal-mass problem. The cell warms a little, the heat spreads, and the next peak starts from a temperature that has already come back down. A chemistry with a modest continuous rating can be the right pick there, because the steady-state number is what drives case temperature, and the impulse never lasts long enough to move it much.
Nine minutes at 2.4C with forty seconds of recovery is a different animal. That is a near-steady-state problem wearing a peak's clothing, and it will be sized by the thermal system rather than by impulse tolerance. Same bin, opposite conclusion.
That is the whole reason the average is useless here. The average sits at 0.6C whether the peaks are brief or long, and the chemistry decision turns entirely on which one it is. A shortlist built on an average tends to reach for the family with the highest continuous rating, which is the most expensive way to solve a problem the machine does not have.
Voltage sag eats the end of the discharge, and it eats it differently by chemistry
The pack does not just have to survive the peak. It has to deliver it when the state of charge is low, and that is where the two main families diverge in a way that changes usable capacity.
LFP holds a flat voltage plateau through most of the discharge. NMC slopes more steadily from full to empty. That shape difference shows up at the end of a discharge under load.
On a flat plateau, the pack sits comfortably inside its window for most of the mission and then crosses a knee where voltage falls quickly. Add a high-current peak near the end of discharge and the internal resistance drop pushes terminal voltage through the low-voltage cutoff before the pack has delivered anything close to its nameplate amp-hours. The BMS either clamps current or opens the contactor, and either way the machine loses performance at exactly the moment the operator wanted it. The flatter the curve, the more abrupt that is, and the harder it is for a voltage-based state-of-charge estimate to warn you it is coming.
On a sloped curve the sag is larger at a given current but easier to see coming. The controller can predict the approach to cutoff and start derating early, which is gentler on the machine and worse on usable capacity, because derating early means leaving amp-hours in the pack.
Neither shape is right in the abstract. What matters is that you run the histogram against the end-of-discharge condition, not just the full-pack condition, and that you specify the behavior at the C-rates and temperatures you actually have, down to cutoff. A line in a pack spec that says 3C peak without saying at 15 percent state of charge, at freezing, down to cutoff voltage is not a spec. It is a wish.
Temperature belongs on the histogram too. Annotate the log with the ambient channel or the pack thermistor and you will usually find that the peak that clears easily in July is the one that trips in January.
The histogram has a negative side
Regenerative braking adds a constraint the discharge side will never reveal. Take the same trace, look at the negative current, and bin it the same way. Charge acceptance is capped by temperature and by state of charge in a way discharge is not, and cold cells take charge slowly.
If the charge histogram exceeds what the pack will accept, the BMS clips the regen. On a vehicle that shifts work to the friction brakes and changes pedal feel. On a machine it can mean a lost function or an overspeed condition. Either way it is a requirement you set, and it filters the shortlist a second time. A family that wins comfortably on discharge peak can lose outright on cold charge acceptance, and nothing in the average load figure will tell you that.
What you hand the supplier
Not a single number. A table.
Give them the amp bins with time-in-bin, longest continuous dwell, excursions per mission, and the temperature band each occurrence fell into. Give them the charge side in the same format. Add missions per day, expected life in missions, and the ambient range the machine actually sees. State the sample rate and where the current was measured, because a supplier who knows the log is at 1 Hz will add their own margin, and one who does not will quote against a number that is smoother than reality.
Then ask them to quote against the table rather than against a continuous C-rate. The suppliers who can answer in the same format are telling you something useful about how they think. The ones who quietly convert it back to an average are telling you something too.
One limitation worth stating plainly
A single log from a single day is not a duty cycle. You want the worst grade, the worst payload, and the worst ambient, and ideally the worst combination of the three, which is often not the day with the highest peaks. If the customer will not hand over the raw log, ask for the ten highest current excursions with their durations and the temperature at the time; a partial answer beats a confident average.
But you do not need a year of data. You need the worst hour, binned honestly, with the dwell times attached. That is what turns a chemistry choice from a guess about marketing tiers into an engineering decision you can defend in the next design review.
