Time Series Prediction Based On Deep Learning And Its Interpretability

Time series prediction has been a popular research topic for decades of booming machine learning,and developing time-series prediction models plays an important role in explaining complex real-world elements.As the fields of natural language processing and image processing develop better and better,people also hope that time series prediction techniques can contribute to real production life.In recent years,deep learning techniques have been constantly updated and developed,new theories and models have emerged,and the arithmetic power of computers has increased,which provides a reliable theoretical support and practical basis for solving complex time series prediction problems.However,in order to apply the models to practical production life,people mostly require the interpretability of the models,but most deep learning models are still complex black boxes inside,which means their internal logic and operation are hidden from users,and the decision basis of the models cannot be fully understood,and the reliability is greatly reduced.Therefore,how to build interpretable time-series prediction models is the focus of today’s deep learning timeseries prediction tasks.This thesis takes multivariate water level data as the research object,and proposes a deep learning-based TCN-LSTM model and Attention-TCN-LSTM model to perform time series prediction tasks for the complex and variable characteristics of water level data.The TCN-LSTM model uses a temporal convolutional network as a feature extractor to extract important features for temporal data,which improves the traditional deep learning The attention-TCN-LSTM model is based on the TCN-LSTM model with the addition of a multi-headed self-attention mechanism module,which can run in parallel with the time-series convolutional network module to The attention score of features is calculated to reduce the influence of unimportant data and abnormal data on the model and enhance the prediction accuracy and robustness of the model.In addition,for the black box characteristics of deep learning models,this thesis quantifies the input variables and implemented functions of the Attention-TCN-LSTM model from various aspects such as model distillation,variable sensitivity analysis,and parameter sensitivity analysis,which enhances the understanding of the model decision basis and improves the overall reliability of the model.In this thesis,the performance of the proposed two water level time-series data prediction models is evaluated on a real water level dataset and compared with the classical baseline model.The experimental results validate the superior performance of the TCN-LSTM model and the Attention-TCN-LSTM model,and provide an interpretability analysis of the models,which helps make the water level prediction models easier to understand for the reader as well as enables model designers to better improve the model performance and increase the reliability of the models.