Comparison of different machine learning algorithms to estimate liquid level for bioreactor management

Estimating the liquid level in an anaerobic digester can be disturbed by its closedness, bubbles and scum formation, and the inhomogeneity of the digestate. In our previous study, a soft-sensor approach using seven pressure meters has been proposed as an alternative for real-time liquid level estimation. Here, machine learning techniques were used to improve the estimation accuracy and optimize the number of sensors required in this approach. Four algorithms, multiple linear regression (MLR), artificial neural network (ANN), random forest (RF), and support vector machine (SVM) with radial basis function kernel were compared for this purpose. All models outperformed the cubic model developed in the previous study, among which the ANN and RF models performed the best. Variable importance analysis suggested that the pressure readings from the top (in the headspace) were the most significant, while the other pressure meters showed varying significance levels depending on the model type. The sensor that experienced both headspace and liquid phases depending on the level variation incurred a higher error than other sensors. The results showed that the ML techniques can provide an effective tool to estimate digester liquid levels by optimizing the number of sensors and reducing the error rate.

Automated summary

Machine learning techniques were used to improve the accuracy of liquid level estimation in an anaerobic digester and optimize the number of sensors required. Four algorithms were compared, with artificial neural network and random forest models performing the best. Variable importance analysis showed that pressure readings from the top were the most significant, and the significance of other sensors varied depending on the model type. Sensors experiencing both headspace and liquid phase variations had higher errors. The results demonstrate the effectiveness of ML techniques in estimating digester liquid levels by optimizing sensor numbers and reducing error rates.