This app allows bioreactor operators to determine their feedstock characteristics from reading only process sensors, letting them set up their physical models to optimize their process control with ease. The deep-models were trained on data generated using a modified version of the ASM1 model implementation by Flores-Alsina et al. (2015). This implementation was modified by us to simulate only one aerobic batch bioreactor instead of a full waste-water-management plant.
The models presented here were trained in Colab to predict the following variables:
- Readily Biodegradable Substrate (
$mg/L$ ) - Slowly Biodegradable Organic Matter Concentration (
$mg/L$ ) - Soluble Ammonia Nitrogen Concentration (
$mg/L$ ) - Soluble Biodegradable Organic Nitrogen Concentration (
$mg/L$ ) - Particulate (slowly) Biodegradable Organic Nitrogen (
$mg/L$ )
The models use only time-series of the first-day (15 minute intervals) of the following sensor data:
- Dissolved Oxygen Concentration (
$mg/L$ ) - Ammonia-Ion Concentration (
$mol/L$ ) - Nitrate-Ion Concentration (
$mol/L$ ) -
$H^+$ -Ion Concentration ($mol/L = 10^{-pH}$ )
This work was conducted as part of a Final Design Project for the Bioresource Engineering (B.Eng.) program at McGill University. For more details on model training and other background information please see the report by Crèche & Harms (2022).