The perils of hot caustic
Ammonia plants use a combination of steam reforming and shift reactors to convert air, steam, and a hydrocarbon feedstock to a stream containing hydrogen, nitrogen, and CO2. The CO2 needs to be removed before the H2/N2 stream is sent to the synthesis loop. The CO2 removal is accomplished using a caustic scrubbing system.
I was developing an overall ammonia plant model for a customer in Japan. About mid-way through the project my Japanese colleagues came to the US for a review meeting in which we tried to match the model results against plant operational data (temperatures, pressures, flow meter readings, chemical analyses).
Everything was matching up rather well… Except for the chemical analysis of the CO2 scrubber underflow. The plant’s analyses were consistently much lower in dissolved CO2 than what my model was calculating. We went round and round trying to figure out what might be wrong with the model physical properties, the control settings for the absorption and stripping columns, etc. We just couldn’t see anything wrong with the model. We could force it to match the plant’s underflow analysis but we then had much too much CO2 getting into the synthesis loop. So my colleagues called the plant back in Japan and asked if there was any way their analysis could be incorrect. They were told (rather vehemently) that there was no way that the analyses could be in error… That plant personnel had taken the underflow sample every shift for nearly twenty years and that the analysis was always the same.
That’s where the light began to dawn. My Japanese colleagues and I had naively assumed that the analysis was the output from some inline automatic analyzer. When I was told that plant personnel “took the sample” I began to suspect that this was a case of sampling error. The underflow from the Benfield unit would be rather hot and at some pressure. I asked my colleagues to call back and ask how the sample was taken and whether it was kept at the same pressure until it was analyzed.
While they checked back with the plant, I reran the model including an underflow sample stream that I flashed to atmospheric pressure. And the model’s predicted sample composition matched the plant analysis perfectly. And then my colleagues got off the phone and confirmed that the sample was taken by opening a valve and catching a stream of hot caustic in a bucket. So the problem was solved. When the pressurized caustic solution was dropped to atmospheric pressure a large part of the CO2 flashed off and that totally changed the sample composition. And the analysis was always the same because it was always flashed to the same state.
We also discussed the point that this sampling exercise was both dangerous to the operator and totally pointless. I later found out that the plant discontinued the practice as soon as my colleagues got back to Japan.
Model validation is a critical step that must be performed before a model can be safely used to study or improve a process. But validation is very much an art. Most large scale process models assume steady-state operation but no large process is every really at steady-state so trying to decide when you are “close enough” is a challenge. In addition, the operational data that you must validate against always has errors. In my experience over modeling dozens of plants, it works out about 50/50. In other words, if you have a significant discrepancy between the model and the plant data about half the time you’ve done something wrong in the model and about half the time the plant data is wrong. That ratio obviously depends on how careful you are in your initial model development and how well run the plant is.