Research On Short-Term Time Series & Time-Space Prediction Algorithm And Application Based On Machine Learning

In the fields of natural science and social production,a large number of decisions are needed,and decision-making problems cannot be separated from predictions.Forecasting is an exploration of future developments based on historical data.Forecasting problems can be derived from any type of data set.There is a type of data called time series data.Time series prediction based on time series data studies how to effectively mine the potential relationship of data development in time dimension and predict the future development of data.In practical applications,some problems only consider the time dimension information can’t meet the problem solving conditions and the growing multi-dimensional data application requirements.The spatial and temporal prediction problem of increasing the spatial dimension information considers both the temporal characteristics and the spatial characteristics.In this paper,two prediction models are proposed to solve the main problems in time series prediction and space-time prediction.details as follows:The main reason for the accuracy and stability of the time series prediction model is the non-stationary characteristics of the time series data.The prediction of road traffic flow is a typical time series forecasting problem.This paper will use short-term traffic flow forecasting as the main application,and propose a short-term traffic flow prediction algorithm in non-stationary environment.First of all,this paper uses the idea of seasonal model to process traffic flow data and eliminate its non-stationarity.Support vector regression(SVR)is chosen as the basic method of prediction.The kernel function is used to transform low-dimensional traffic flow data into high-dimensional space.The linear decision function is constructed to transform the regression prediction problem into convex quadratic programming problem.Get the optimal solution.The hyperparameter optimization problem of machine learning model has always been a major problem that plagues the development of machine learning applications.For this reason,this paper proposes an automatic parameter optimization method.The hyperparameter optimization process of the model is regarded as a black box function optimization process.The maximum value of the function is obtained by Bayesian optimization method.The Gaussian process is used to construct the prior probability model of the objective function,and then the Bayesian theory is used to calculate Posterior probability.After multiple iterations,the optimal hyperparameter combination is obtained,so that the prediction model can obtain the best learning effect.The experimental results show that the proposed model has good prediction accuracy and stability.The difficulty in time-space prediction is mainly the effective extraction of temporal and spatial structure information.In this paper,short-term rainfall prediction based on radar data space-time extrapolation is the main application.The short-term space-time prediction is studied.A deep neural network model for extracting spatiotemporal structure information is proposed for short-term rainfall regression prediction.For the radar image time series data samples,this paper firstly uses a set of convolutional neural networks(CNN)to extract the spatial information contained in the radar image to complete the meteorological spatial feature coding.On top of this,the extracted feature map is used as the input of the long-term and short-term memory network(LSTM),and the correlation of the timing of the radar feature image is extracted by LSTM,and the feature learning of the time dimension of the radar weather map is completed.The fusion of convolutional neural network and LSTM effectively extracts the spatio-temporal features of radar images.Finally,the full-connection layer output is used to complete the rainfall regression prediction,and a complete radar extrapolation short-term precipitation prediction framework is constructed.Experiments based on real data acquisition verify that the proposed algorithm has good prediction accuracy.