Prediction of Loess Collapsibility Coefficient through the HHO-BP Neural Network

Abstract

The coefficient of collapsibility is the key parameter involved in the computation of the loess collapsible deformation. However, influencing by many factors, theoretical estimation of this parameter becomes extremely difficult. In order to quickly and conveniently predict the loess collapsibility coefficient according to soil index properties, through data mining technology, a model was proposed based on BP neural network optimized by HHO (Harris Hawk Optimization) algorithm. The proposed model was validated using a database acquired from engineering practice, which indicated that at least five soil index properties were necessary for accurate predictions. Also, compared with prediction models based on BP (Back Propagation) and PSO (Particle Swarm Optimization)-BP neural network, the proposed model gained faster iteration rate, higher prediction accuracy and lower error value. Sensitivity analysis based on the connection weights shows that the most important index properties affecting the coefficient of collapsibility are plasticity index and void ratio, followed by dry density and degree of saturation, and finally water content.