Nonlinear Forecast Of Soilless Cultivation Ceramsite Property Based On BP Neural Network

Soilless cultivation technology is a new type of plant cultivation technology.It has broad prospects for future development.As the substrate of soilless cultivation,ceramsite can store and release water and fertilizer required for the growth of plants.Its density and water absorption rate have a decisive influence on soilless cultivation technology.Research and practice have proved that the type and proportion of raw materials of ceramsite are variable,and the properties are also affected by many factors such as granulation and sintering process.It shows highly nonlinear mapping relationship between the properties of ceramsite and the above factors.This relationship cannot be described by conventional mathematical relations or mathematical models,so it is difficult to accurately forecast the properties of soilless cultivation ceramsite.In this thesis,a three-layer BP neural network model was established to solve the difficult fitting problem of highly nonlinear mapping relationship between the properties of soilless cultivation ceramsite and influence factors,such as the proportion of raw materials and sintering parameters.Levenberg-Marquardt algorithm,Bayesian Regularization algorithm,Scaled Conjugate Gradient algorithm were used to train the BP neural network.Genetic algorithm was embedded into the BP neural network to realize the global optimization of weight and threshold.The nonlinear fitting effect using the four BP neural network models was compared and analyzed to determine the optimal model.On this basis,the optimal model were used to forecast the properties of fly ash(or coal gangue)ceramsite,and its reliability is verified by experiments.The main conclusions are as follows:(1)Four BP neural network model can better fit the highly nonlinear mapping relationship between the apparent density or water absorption and six influence factors of the fly ash ceramite.The nonlinear fitting effect is very good with good fitting accuracy and goodness.(2)Four BP neural network model can better fit the highly nonlinear mapping relationship between the apparent density or water absorption of the coal gangue ceramite and five influence factors.The nonlinear fitting effect is also very good with good fitting accuracy and goodness.(3)When using Bayesian Regularization algorithm,the BP neural network nonlinear fitting has the smallest root-mean-square error,the largest coefficient of determination,and the largest correlation coefficient of the master samples fitting.The slope of the master samples fitting line is closest to 1.It has the best nonlinear fitting effect by Bayesian Regularization algorithm,followed by the Levenberg-Marquardt algorithm,and Scaled Conjugate Gradient algorithm and Genetic algorithm.(4)The convergence speed is faster when using Levenberg-Marquardt algorithm and Genetic algorithm,than the speed when using Scaled Conjugate Gradient algorithm and Bayesian Regularization algorithm.(5)BP neural network model based on Bayesian Regularization algorithm was selected for nonlinear forecast.The model can accurately forecast the properties of fly ash(or coal gangue)ceramide according to the preset proportion of raw materials and sintering parameters,and the relative error between the forecast value and the test value is within ±10%.So,The forecast accuracy of model is very high.