Research On Prediction Model Of Charge State Of Fuel Cells Based On Pruning Neural Network

At present,energy and environmental problems are becoming more and more serious,and vigorously developing energy-saving and new energy vehicles is an effective way to solve energy and environmental problems,and it is also a powerful measure to achieve national ecological civilization construction.Unlike traditional fuel vehicles,new energy vehicles use fuel cell stacks as one of all power sources or power sources.Accurate and real-time acquisition of the state of charge(SOC)of the fuel cell stack is essential for battery management systems.Function.In order to improve the accurate estimation of battery SOC,this paper proposes a fuel cell stack state-of-charge prediction algorithm based on pruned neural network.The BP(Back Propagation)neural network is optimized by improving the fast convergence characteristics of the Particle Swarm Optimization(PSO)Particle Swarm Optimization algorithm,and then the network model is pruned,the model topology is compressed,the calculation complexity is reduced,and the model training speed is increased.Precision.The main research content includes the following three aspects:1.The evaluation function compares and analyzes the dynamic and static characteristics of the fuel cell stack to determine the composition vector of the input neurons.Analyze the impact of dynamic characteristics such as battery discharge voltage and static characteristics such as battery manufacturers on the accuracy of SOC through two evaluation methods:dynamic characteristics and static characteristics,and select reliable and significant feature factors that affect the experimental results as training samples to train the model.Effectively improve SOC prediction results.2.Improved particle swarm algorithm.The traditional particle swarm optimization algorithm is prone to the problem of premature convergence falling into the local optimal solution during particle iteration.At the same time,inappropriate learning factors,inertial weights and other parameters may lead to inaccurate final convergence results.Therefore,this paper proposes an improved method,using Logistic function to chaotically initialize the spatial position of the particles;introducing FWA(Fireworks Algorithm)algorithm to increase the randomness in the particle iteration process;establishing a new parameter formula for dynamic tuning;and proposing a cyclic single-dimensional optimization strategy,While maintaining the diversity of the particle swarm,it avoids the particles from crossing the optimal solution.Experiments show that the improved particle swarm optimization algorithm can effectively improve the convergence speed and convergence accuracy.3.Optimized the initial parameters and topological structure of BP neural network.Using the fast convergence characteristics of particle swarm optimization algorithm,a rough search is performed in the global range,and a set of better initial weights and thresholds are obtained to train the model.Then use the pruning algorithm to compress the topology of the BP neural network model,remove the redundant connection weights,and save the sparse format model.By integrating the particle swarm optimization algorithm,while maintaining the BP neural network model refinement ability,it also improves the training accuracy of the model.The experimental results show that the SOC prediction algorithm based on PSO optimized BP neural network can shorten the training cycle time of the model and improve the prediction accuracy.