Time Series Forecasting Method Based On Deep Learning And Application To Internet Of Things

With the rapid development of the Internet of Things(IoT)technology,its applications are widely used in various scenarios and generate massive time-series data.Time series prediction is based on discovering the pattern of historical data,therefore ouput the state of the next time slot.Mining the patterns behind these data are crucial to various industries.However,due to the non-stationary and non-linear characteristics of the time series data,the existing solutions cannot meet the requirements of the diversification and differentiations in IoT data forecasting.Rapid improvement of hardware's process capacity stimulates the development of deep learning,and its tremendous achievements in natural language processing and other fields provide effective solutions for IoT series data prediction.Therefore,how to use deep neural networks to solve the time series forecasting problems in specific IoT application scenarios is the main research content of this article.The main contribution of this research includes:1.Propose a time series prediction framework based on LSTM feature fusion.With the generation of massive amounts of data,time series have changed from low-dimensional to complex high-dimensional.Traditional time series forecasting methods cannot discover the correlation between multi-dimensional features,and most of them focus on short-term time correlation,which results in perditions inaccuracies and inefficency in specific IoT scenarios.To overcme such challenge,this paper proposes a network framework suitable for multi-feature time series prediction.This network framework can not only discover the time correlation in the time series,but also complete the feature extraction and capture the correlation between multiple features.And use this framework to solve the problem of cow estrus detection in the application scenario of the IoT smart animal husbandry.The application results show that our proposed solution outperforms exiting detection algorithms in terms of accuracy and efficiency.2.A spatio-temporal sequence prediction framework based on multi-component spatio-temporal graph convolutional neural network is proposed.The spatio-temporal data in IoT scenarios contain geographic location information.In order to fully mine and utilize the spatio-temporal correlation and the multiple temporal dimension patterns contained in the time series data,this paper proposes a multi-component fusion forecasting framework suitable for spatio-temporal data.The framework is based on the spatio-temporal graph convolution,uses graph convolutional neural networks to discover spatial correlations,uses time-gated convolution to extract temporal correlations,and finally predict by fusing multiple components and other related features.This paper uses the framework to solve the problem of traffic flow prediction in the application scenarios of IoT smart transportation.The application results show that the multi-component fusion prediction framework proposed in this paper is better than a single component,and the fusion of related features such as climate can further improve the predicted outperforms.