Time Series Forecasting Based On Gated Neural Networks And Reinforcement Learning

With the advent of the era of big data,data presents new features of a sharp increase in volume,increasingly complex types,and reduced value density.However,the requirements for data processing performance have not been reduced.How to predict and analyze it has been a hot topic in recent years.At present,the main time series prediction methods are divided into traditional linear prediction models which are based on the assumption of stationarity,but most of the actual data are non-linear.And the artificial neural network method,although it can process non-linear data more effectively,but its versatility is relatively poor.In order to solve the problem that many current deep learning methods cannot always find the optimal network parameters when predicting time series data with weak periodic samples,and the accuracy is not high enough,we proposed a new gated recurrent unit with reinforcement learning which can be called GRURL.The network uses the reward mechanism to automatically adjust network parameters to the best state while making predictions.It also finds the methods that works best for GRURL in data preprocessing,overfitting problems,and gradient optimization.we compare the GRURL model with BP network,LSTM network,and general GRU network,observe whether the performance is improved,and evaluate the error of each prediction.Taking a small sample of vehicle sales data with weak periodicity as an example,the experimental results show that the improved GRURL network has lower prediction errors than BP,LSTM,and GRU networks,and is more suitable for predicting such data.At the same time,a series of A-share time series data was used to conduct extended experiments to verify the universality of this network.Finally,considering that most neural network models are currently built and trained on the IDE,it is difficult for others to use directly except the writer himself.Therefore,a visual front-end and back-end system was designed and implemented,using tfjsvis as the front end to display system functions and the tensorflow.js framework as a back-end to encapsulate a trained model which can implement functions like java web.