Artificial neural network models for wall parameters on plug-1 flow characteristics through pipelines

Abstract

Dense phase pneumatic conveying is a preferable method for transportation of a friable, fragile, abrasive, or agglomerated in nature material through pipeline with comparatively less wear and tear of the system and product as well. A fine particulate material moves as a single entity due to its cohesion in vertical or thin horizontal pipes in Plug-1 flow. A model based on mechanics was created and later modified using experiments for this type of flow. In this work Artificial Neural Networks (ANNs) models are used to study the effect of wall friction coefficient and coefficient of wall cohesion on the pressure drop. Three different datasets having 50000, 250000, and 500000 data points were used to test 19, 21, and 25 ANN architectures respectively. The best architecture was found to be t50-t40-r1 architecture with Adamax optimizer, with mean absolute percentage error (MAPE) being close to 0.00402% when tested on the 500000 samples dataset with 25000 test values, 0.0043% when tested on the dataset with 250000 samples and 25000 test values, and 0.0035% on the 50000 samples dataset with 10000 test values. The s20-s20-r1 architecture with Adam optimizer was quick and gave second best results with MAPE being close to 0.009% when tested on the 500000 samples dataset, 0.00988% when tested on the dataset with 250000 samples, and 0.00408% when tested on the 50000 samples dataset. The t40-t40-t40-t40-r1 with Adamax optimizer was slow but gave the third best results with MAPE being close to 0.0166% when tested on the 500000 samples dataset, 0.00496% when tested on the dataset with 250000 samples, and 0.00480% when tested on the 50000 samples dataset.