Research On Dynamically Correlated Time Series Prediction Model

In recent years,with the rapid development of Smart-City,more and more sensors are deployed in various locations of the cities to continuously monitor the surrounding environment and traffic.Each sensor continuously collects a large number of time series data in its specific geographical space.In geospatial,the data in the same area are monitored by multiple sensors,and there is a complex correlation between them;in the time dimension,the sequence data of each sensor changes irregularly with time.In addition,road conditions around the sensor,weather changes,POI and so on will also affect the sensor readings.Therefore,the prediction of sensor data can help to control urban air pollution and reduce traffic congestion,so as to serve the city and effectively improve people’s lives.Problems of the traditional sensor prediction research are as follows:(1)firstly,there is scale mismatch between the sensor monitoring range and external weather data;secondly,there is cross domain influence between multi-source heterogeneous data and sensor data.(2)The modeling of time series is mostly nonlinear.The sequence data of each sensor is highly dynamic,the change of data with time has two characteristics: nonlinear and linear.(3)In the spatial dimension,there is correlation between multiple sensors,and this correlation is dynamic.For the above problems,the research in this paper is as follows:Firstly,a multi-source heterogeneous data grouping fusion method based on multi feature clustering is proposed.High frequency weather data in grid form from official sources is used,and a method based on image convolution is used to extract grid weather data to capture its impact on time series;a method based on multi feature clustering is used to group and fuse the external features and time series features that affect time series.Secondly,a time series prediction method AMED(ARIMA and multiple encoder decoder)based on ARIMA(autoregressive integral moving average)model and multiple Encoder-Decoder is proposed.The ARIMA model is used to fit the time series to obtain the residual sequence and a part of the prediction sequence which contain the nonlinear characteristics of the time series.Then the multiple Encoder-Decoders is used to process and predict the multiple feature groups to obtain a part of the prediction sequence.Then,we use a Convolution Neural Network(CNN)and ARIMA to convolute and predict the historical time series data around the site.Through a softmax function,the final prediction result is obtained by weighted summation of the three results.Finally,taking Beijing air quality prediction and San Francisco Bay area traffic flow prediction as the target,the experiments are carried out to compare the Mean Absolute Error(MAE)of prediction under different feature grouping methods,and verify the superiority of the grouping method proposed in this paper;by comparing the Mean Absolute Error(MAE)and Root Mean Square Error(RMSE)of prediction under different time intervals of several models,the prediction results of the proposed model are proved to have better performance and can be applied to a variety of scenarios.