Prediction Of Seismic Response Of Liquid Storage Tanks Based On Neural Networks

Tanks studied are cylindrical tanks, which are standing on the ground withoutanchor. By the development of petroleum chemical industry, light industry, and nuclearelectricity, they are widely used. Now they have been the most important establishment.Tanks are usually used to storage liquids which are easily to burn or explode. Once metearthquake, they are too easily to result in severe by-disaster, which will bring ouractivities and circumstance serious calamity. Study of tanks' seismic response has beenone of the most attractive problems. After neural networks brought and its successfullyapplications in many fields, a new method of tanks' seismic response appeared.This paper worked in prediction of seismic response of liquid storage tanks basedon BP neural networks, aiming at 8 floating roof tanks. The main work studied asbelow:1. Presented BP neural networks to predict the seismic responses of liquid storagetanks based on the current internal and external studies and development;2. Introduced neural networks and BP neural networks and problems which areneeded to notice. According to the standing cylindrical tanks calculating method incode GB50191-93, obtained seismic responses of 8 standard floating roof tanks withdifferent heights of liquid stored, and intensities, including base shear, base moment,and axial compressive stress of tank wall. The data obtain were used as training data ofBP neural networks model.Primarily established a prediction model of seismic responses of liquid storagetanks under MATLAB Neural Networks Toolbox Functions. After testing the reliabilityof the model by predicting seismic responses of 8 groups of random conditions chosencertain height of liquid stored, and intensity, found that relative prediction errors of1000m3 tank, 2000m3 tank, 10000m3 tank and 50000m3 tank are too big to be accepted.After analyzing, considered it was affected by the boundary factor and ratio ofdiameter and height of liquid stored. By decreasing these four tanks' height differencesof liquid stored, the model was re-founded by training the newly obtained sample data,and a new model with relative prediction error distinctively reduced was obtained. Themaximum of relative prediction error is -5.48%.3. At last, summarized the results, and indicated the problems found and themethod for improving. Advised that use the method of combine Genetic Algorithmsand neural-fuzzy to study the prediction of seismic response of liquid storage tanksmore deeply.