Time Series Forecasting Based On Incremental Integrated LSTM Model

Time series data widely exist in many fields such as economy,society and industry,such as stock big data,meteorological big data,thermal power big data and so on.Since these massive time series data represent the operating mechanism behind the objects they belong to,how to excavate and model such time series big data so as to predict and control the future evolution of the belonging objects has always been a hot spot in our society.However,the data in reality not only has massive and complex characteristics,but also changes dynamically over time,which puts extremely high requirements on the model's fitting ability and dynamic adaptability.On the one hand,the characteristics of massive,complex and non-linear time series data makes it difficult for the traditional statistical regression methods to fit the data,which cannot meet the accuracy requirements of practical engineering.On the other hand,emerging deep learning methods often use offline batch processing for training,and the high-speed dynamic changes of time-series data streams will cause the model to gradually fail over time,that is,the prediction accuracy of the model will continue to decline.As re-training the updated data on the neural network often requires a very high time cost,it is difficult for the existing deep learning methods to quickly implement the dynamic evolution of the model.Due to the limitations of the traditional statistical regression model and deep learning method,this paper proposes an incremental integrated LSTM model to solve the problem of online time series prediction.Since incremental learning can continuously learn new knowledge from the new data set on the basis of retaining the learned knowledge from the old data set,this paper introduces the idea of incremental learning to dynamically fit the complex time series data with dynamic changes.In this paper,the real-time data stream is sliced into several sub-data sets.By using LSTM's time memory characteristics and strong nonlinear fitting ability to train the sub-data sets,several weak learners are obtained.In order to reduce the training time of neural network and improve the efficiency and effect of the model,the transfer learning method based on sample and model are used to perform transfer training among multiple weak learners.Finally,the integration strategy is used to combine the weak learners to update the strong learner incrementally so as to predict the real-time data online.The innovation of this paper mainly lies in the following three points:1.Proposed an integrated LSTM neural network model that can be updated incrementally and integrates the ideas of integrated learning and transfer learning.Compared with the traditional LSTM model,our model can realize the parallel training and online update of the neural network.2.A shared fusion algorithm combining the original neural network layer and the new neural network layer is proposed,which can extract valid information from the original neural network and fuse it with the current new input information.3.We propose a weak learner buffer pool strategy to limit the infinite growth of the number of weak learners in the model's time series data stream prediction process.This model also provides a new online time series prediction method.This paper implements the proposed theoretical model on a public air quality time series data set.The experimental results show that,compared with the traditional offline LSTM neural network,our method not only has a certain degree of improvement in the model fitting degree and prediction loss,but also can greatly improve the training efficiency of the model,and realize the incremental prediction of time series.