Prediction Of Remaining Useful Life Of Supercapacitors Based On Improved Temporal Convolutional Networ

Supercapacitors are used in many fields,such as providing peak power demand and storing braking energy.These applications require high reliability and service life of supercapacitors.For high power,high cycle and maintenance-free energy storage systems,the characteristics of supercapacitors meet the requirements of the above applications.The energy density and power density of supercapacitor will decrease due to its aging and imperfect manufacturing process,which will eventually lead to its failure.In order to avoid the breakdown of the energy storage system,it is important to improve the reliability of the energy storage system to complete the replacement before device reaches end of life state.Accurate tracking of aging degree of supercapacitor is the premise of accurate intervention before device failure.Aging of supercapacitor is a nonlinear and complex process involving multiple variables.In order to emphasize the focus of research,this paper abstracts and simplifies the physical model of supercapacitor,and introduces the equivalent capacitance parameter to describe the aging state of supercapacitor.The cycle life test method was used to charge and discharge supercapacitors at constant current,and the capacity decay data sets of supercapacitors were collected and recorded to provide data support for predicting the remaining useful life of supercapacitors.Due to the advantages of artificial neural network such as self-learning,associative storage and optimal solution tracking,it can well adapt to the characteristics of nonlinear and high complexity of supercapacitor aging.Therefore,this paper applies classical artificial neural networks,such as back propagation neural network,gated recurrent units recurrent neural network and long short-term memory recurrent neural network,to the analysis and prediction of the remaining useful life of supercapacitor,and simulates and selects the prime model.On this basis,the model is upgraded to a stacked bidirectional long short-term memory recurrent neural network from the aspects of comprehensive weight adjustment and network capacity expansion.Experimental results show that the performance of the upgraded model is better than that of the basic model.Even though the recurrent neural network has excellent performance in time series,the problems of gradient explosion and gradient disappearance caused by the multiplication effect in the process of error back propagation still exist.In order to improve the above defects,this paper proposes time convolution network.The time convolutional network breaks the traditional recurrent network framework,and its residual modules are composed of two layers:extended convolution,weight normalization layer,Relu nonlinear function,and Dropout regularization stack.As a result,the conditions for generating gradient disappearance become more stringent,thus weakening the above problems.Furthermore,the existence of jumping connections can overcome the degradation problem of stacked network layer.In order to prevent the model from over-fitting,the early stop technique is used to limit the training degree.Simulation results show that time convolutional network can further improve the accuracy of tracking aging state of supercapacitor.In order to optimize the robustness of the model,this paper selects the initial weights and thresholds of the time convolution network using improved particle swarm optimization algorithm.The inertial weightω,the maximum flight step c₁ of individual extremum and the maximum flight step c₂ of global extremum are optimized into adaptive variables that balance local search and global search capability.In order to improve the probability of the algorithm converging to the global optimal solution,the simulated annealing algorithm is introduced in this paper.The probability of accepting a solution worse than the current one is calculated using Metropolis criterion,and the probability of jumping out of the local and approaching the global optimal solution is improved.Finally,simulation results compare the unoptimized,traditional and improved time convolution networks,and verify the high precision and robustness of the proposed method.