Analysis and Improvement Approach of the Impact of Data Disturbance on LSTM Prediction Algorithm

Abstract

If the sampling data with noise or outliers are used to train the long-short-term memory (short as LSTM) network, whether the perturbations and outliers in sampling data affect the training performance and prediction accuracy of LSTM networks is a key problem. This paper analyzed the impact on the LSTM prediction effect when using perturbations and isolated/patchy outliers involved data for network training and prediction. The results showed that the prediction accuracy decreases as the amplitude of the perturbations and the range of outliers increase. In order to overcome this effect, an improved method of Pre-set Outliers Tolerant Filter is proposed, and an Outliers-Tolerant Multi-LSTM model, in short, the OTML model, is established. The prediction effect of the proposed model is compared with that of the LSTM model without the filter and that of the LSTM model with a mean filter. Comparison results showed that the OTML prediction model proposed in this paper can eliminate the influence of random noise and isolated outliers thoroughly. When meeting the patchy outliers whose width is smaller than the radius of the window, the OTML prediction model can filter them out too, so as to realize the high-fidelity prediction of the data.