Deep Differential Neuro Evolution For Chaotic Time Series Prediction

Chaos is widespread in nonlinear systems such as climate,energy,hydrology,and finance,and the presence of chaos often makes nonlinear systems more difficult to analyze and predict.In a chaotic system,many factors affect the system's motion and are difficult to define.Besides,it is difficult to obtain complete system information when observing a nonlinear system,which will undoubtedly bring difficulties to the prediction of the nonlinear system.Besides,the traditional neural networks model for chaotic time series prediction suffers from low prediction accuracy and difficulty in determining the network topology.In recent years,the problems of chaotic time series prediction have also drawn the co ncern of developers in the domain of deep learning.To improve the prediction accuracy of chaotic time series,this paper carries out the research of chaotic time series prediction models and algorithms based on chaos theory and deep learning,and the main research contents are as follows.(1)This paper uses a convolutional neural network(CNN),gated recurrent units neural network(GRU),and attention mechanism to develop a deep hybrid neural network.It is applied for predicting chaotic time series.In data preprocessing,this paper combines normalization and phase space reconstruction techniques in chaos theory to process the data.In the deep hybrid neural network model,CNN is used to obtain the spatial feature information of the system from the reconstructed phase space of the chaotic time series.Then,the spatial feature information is combined with the original chaotic time series and GRU is used to extract the spatio-temporal feature information of the system from the combined series.In this paper,an attention mechanism with a nonlinear activation function is designed to capture the key spatio-temporal information.The attention mechanism can distribute the weight of key spatio-temporal feature information,and the prediction model will predict according to the obtained weighted spatio-temporal feature information.(2)This paper uses the idea of neuroevolution to optimize the topology of the proposed model,including the number of filters of the CNN,the number of hidden neurons of the GRU,and the time step required for prediction.For the problem of slow convergence speed and operation speed of the differential evolution algorithm(DE),this paper uses adaptive variational operators and dynamic chaotic crossov er operators to improve the DE algorithm.(3)The feasibility of the proposed prediction algorithm and model is verified by the Lorenz theory dataset,and its applicability in different scenarios is verified by using the monthly average sunspot dataset and coal mine gas concentration dataset.According to the experimental results,the improved DE algorithm not only has a more rapid aggregation rate but also can accelerate the optimization search time.Besides,the improved DE algorithm is optimized to achieve stable and high precision prediction performance in chaotic time series prediction.The chaotic time series prediction model proposed in this paper has strong adaptability and can be extended to different nonlinear problems,which holds important academic and functional relevance to the research and understanding of nonlinear systems.