Deep Learning Based Interval Forecasting Model And Its Application In Time Series Analysis

In recent years,with the rapid development of the big data technology in China,huge amounts of time series data have been collected from the fields of finance,power,and industry.The technique of time series prediction can be used for deeply mining the variation regularity and trends of time series data,and for making relatively scientific inferences about future development and changes of the related fields.Time series prediction models have been widely applied in the fields of electric power prediction,stock analysis and prediction,and air quality index prediction.The methods of time series prediction can be listed as the point prediction and the probability interval prediction.The point prediction method can only carry out the value of the mean prediction,and usually the prediction results are not accurate enough,and there are some sense of uncertainty in the prediction,which is difficult to estimate.And currently,the probabilistic interval prediction methods have some problems such as the need for distribution assumptions,high computational complexity,and intervals crossing.In order to optimally estimate the uncertainty level of the time series prediction,this paper proposes a novel differentiable multi-interval loss function with distributionfree based on the high-quality interval principle of the prediction interval,and the relationship between the intervals at different confidence levels as well as the Sigmoid function softening technique.In this paper,a multi-interval joint prediction model is proposed by combining the new loss function with deep learning,and the validity and practicability of the model are verified by numerical simulation data and public data.The experimental results show that the multi-interval loss function can not only effectively solve the problem of interval crossing,but also can efficiently and reasonably quantify the uncertainty of the prediction.Compared with the selected referential models,the performance of the predicted interval from the proposed method is raised at the most by more than 20%.