Time Series Prediction Based On Deep Network And Its Interpretability

Time series forecasting is an important research direction,and in various fields of reality,improving the accuracy of time series forecasting can improve the return and reduce the risk,thus bringing great practical significance.At present,the mainstream time series forecasting methods include statistical learning methods,traditional machine learning methods and deep learning methods.Due to the complicated temporal pattern of time series and complex correlation with external factors,traditional methods cannot efficiently extract the characteristics of time series.Deep learning models,especially the Long Short-Term Memory(LSTM)based models with attention mechanism present the art-of-the-state performance due to its excellent capability of feature extraction and representation.However,deep learning methods are difficult to visualize and have poor interpretation.In this thesis,more reliable time series forecasting model and its interpretion model are studied.It is found that the classic LSTM model cannot capture the long-time dependence of time series,which focuses more on the output of the final hidden layer.In order to distill correlation information with large time span,a dual-attention mechanism-based LSTM model is developed,which introduces attention mechanisms in both temporal and feature dimensions to the classic LSTM model.Sufficient experiments are implemented,and results on three datasets show that the proposed model outperforms classic LSTM model and single-attention mechanism LSTM model on both forecasting accuracy and robustness.The model is interpreted by visualizing the attention weights of the proposed model,and the results show that the dual-attention mechanism can capture the much larger time-span correlation information.A multi-state forecasting algorithm based on LSTM and Hidden Markov Model(HMM)is proposed for time series with multiple pattern shift.The algorithm first uses the traditional HMM with high solvability for the shift modeling of hidden states.Then,the states are modeled by LSTM models.The proposed model is tested on several datasets,and the results verify its effectiveness.Moreover,an interpreting method is designed to characterize the shift of hidden states of LSTM network,which can visualize the hidden states of LSTM model.