Developing a Recommender System for Ammonia Production

Alexander Verkholomov25 May 2024

Problem

A chemical plant wants more output without new equipment. Operators run the process by hand and do not know the best setpoints for gas flow, steam, air and temperatures. The EuroChem team had three months to reach a demo. The data was poor: the key metric (ammonia output) is summed from several sensors with different errors and buffer tanks, analyzers are few, signals are noisy, and effects of a parameter arrive with delays of up to 4 hours.

Example from the article: at the "Ammonia 4" unit of NAK Azot, models predict hydrogen production, the CO2 content after absorption and ammonia output, and the operator sees recommended air flow on a screen.

Solution in brief

Split the goal into several model-plus-optimizer pairs, each for one measurable intermediate target. Build lagged features from correlation plots, fit a hybrid linear plus tree model, and search the controllable inputs with a gradient-free black-box optimizer. Post-process the raw optimum (smoothing, limits, step changes) before showing it to the operator.

Questions this case answers
  • How to recommend setpoints of controllable inputs for a slow process whose target is not measured directly, when signals are noisy and effects arrive with delays and only a few sensors exist
  • How to optimize controllable inputs of a black-box model without gradients and turn the raw optimum into a recommendation that a human operator can safely follow
  • How to build a recommender system that tells plant operators which process parameters to set to raise output
  • How to model a production metric when there is no direct sensor and the signal is noisy and delayed

Tools and methods

  • Ridge regression + CatBoost, stacked by linear regression — hybrid model for linear and conditional dependencies
  • LightGBM — digital twin of the CO2 analyzer
  • Nevergrad (NGOpt) — gradient-free optimizer over a serialized black-box model
  • Lag features from correlation-vs-shift plots — handle delays in a slow process
  • Post-processing layer — smoothing, regulatory bounds, step changes, temporary shutdown logic
Continue reading — it's free

Google or a one-time email code. No password.