An ensemble multi-ANN approach for virtual oxygen sensing and air leakage prediction in biomass gasification plants

A recurring challenge in the operation of biomass gasification plants is the occurrence of air leaks, which prevent the resulting lean producer gas from meeting the required standards for power generation. In order to address this issue, an ensemble model composed of multiple artificial neural networks (ANNs) was developed to predict the oxygen concentration in the gas mixture and detect anomalous operating conditions (air leakage). Throughout an extensive experimental campaign, the volumetric composition of the gas mixture from a semi-industrial scale downdraft gasifier fueled with biomass pellets was systematically measured and recorded at a constant time step of 10 s using an inline portable syngas analyzer equipped with NDIR, TCD and ECD sensors. The ensemble multi-ANN model was trained with a total of 24 representative datasets, including instances of both normal and anomalous operating conditions, using k -fold cross validation with 10 submodels. The results revealed an R 2 of 0.99 and an RMSE below 0.3, indicating that the model’s error margin is lower than that of the ECD sensor. The developed model can serve as a supervisor for the ECD sensor by performing a double verification or even potentially replacing the ECD sensor, with the model assuming the task of predicting the oxygen concentration using the data recorded by the NDIR sensor. • An ML model has been developed to detect air leaks in biomass gasification plants. • A pilot-scale gasification plant has been used to collect the whole dataset. • The model has been trained on a dataset comprising a total of 27,715 records. • The RMSE of the model is 0.3, below the precision of the sensor used to record data. • The model can be used as a virtual sensor to supervise or replace the real sensor.