Research On Firefly Algorithm Based On Deep Learning And Its Application In Medium And Long Term Runoff Forecast

Runoff is a random variable affected by many factors such as solar activity,atmospheric circulation,underlying surface changes and rainfall,and the relationship between various factors on runoff is difficult to establish a rigorous mathematical model to represent,so runoff forecast is a multi-factor nonlinear complex problems.With the development of machine learning technology,support vector machine for regression has been widely used in the field of runoff forecasting,and effective forecast results have been obtained.The predictive performance of support vector machine for regression depends on the choice of penalty coefficient,the insensitive loss coefficient,and the kernel parameter.The traditional method has low efficiency in selecting these three parameters,and the subjectivity and randomness are large,resulting in poor prediction performance of support vector machine for regression.In order to improve the performance of support vector regression,the firefly algorithm is used to optimize the parameters,and a support vector machine for regression prediction model based on firefly algorithm is established.The firefly algorithm is a swarm intelligence algorithm.Because of its advantages such as simple optimization model,few parameters and easy implementation,it is widely used in many engineering fields.However,the algorithm itself has the defects of being easy to fall into local optimum,premature convergence and low precision.This thesis takes the firefly algorithm as the research object,uses the runoff prediction as the application background,uses the deep learning strategy to optimize the algorithm,and applies the improved firefly algorithm to the medium and long-term runoff prediction.The main research results obtained are as follows:(1)For the improved strategy of deep learning method,one-dimensional deep learning firefly algorithm guided by the best particle is proposed.In order to make the optimal particles obtain more search opportunities,in the evolution process of each generation of the population,the algorithm allocates certain evaluation resources for the optimal particles,performs fixed-number one-dimensional deep learning,searches for excellent solutions,and other fireflies still maintain the standard update way.The experimental results of 12 benchmark functions show that the optimization performance of the algorithm is improved.(2)For the improved strategy of deep learning objects,firefly algorithm guided by general center particle is proposed.In order to strengthen the information exchange between populations,the algorithm introduces the concept of general central particle and constructs a general central particle related to all particles.After all the fireflies complete the standard search task,the general central particle guides the population for further exploration to improve the convergence speed and search accuracy of the population.Theexperimental results of 12 benchmark functions and CEC2015 complex test functions show that the proposed algorithm has better overall performance than the other six new firefly algorithms.(3)Combining the above-mentioned deep learning methods and deep learning object,deep learning firefly algorithm is proposed.In order to strengthen the guiding ability of general central particles,the stochastic model is used to replace the full-attraction model for population evolution,then the general central particle is selected as the deep learning object,and it is subjected to a fixed number of one-dimensional deep learning.Finally,the general center particle after one-dimensional deep learning is used to guide the population evolution.Experiments with 12 benchmark functions show that the optimization performance of the deep learning firefly algorithm has been significantly improved.Taking the annual runoff of Huangfuchuan hydrological station in Fugu County,Shaanxi Province as the research object,and establishes the runoff forecast model based on support vector machine for regression with taking mean square error of the actual and the predicted runoff of value as the objective function.Various improved firefly algorithms mentioned in this thesis are used to optimize the support vector machine for regression kernel parameter selection.Experiments show that the support vector machine for regression model based on deep learning firefly algorithm has obtained the best forecast results.