Estimation of fast pyrolysis product yields of different biomasses by artificial neural networks

In this study, the yields of different biomasses and wastewater sludges obtained in the drop-tube-reactor fast pyrolysis system were estimated by feedforward artificial neural networks (ANN) models using a total of 174 experimental data. The performances of 14 developed models in estimating the yields were investigated by using 7 data sets consisting of 21 input parameters with different data sizes, hidden layers, and neuron numbers. The best and average MSE values obtained from ANN application for bio-oil (BO) output of 14 models are listed from smallest to largest. Models numbered as 5-10-14 with lower top 3 average MSE values were selected as better models in the ranking. Among the three models, the ANN architecture has 1 hidden layer, 20 neurons, and 75-15-15 % data division for training-testing-validation. ANN architecture performance for BO output was applied to two different datasets for biochar (BC), BC-BO, and BC-BO-BG (biogas) products within the scope of models 5-10-14 and their performances were examined with MSE and R2 statistical parameters. The lowest and highet MSE values were 3.91 and 6.99 for BO and BC estimations in the first database, but they were 3.67 and 10.72 for BO and BC-BO estimations in the second database, respectively.