Abstract: Nowadays,most researches in deep learning depend on the self-learning capacity of used neural network.Specifically,they focus on using as less human-knowledge priors as possible during the training step,which leads to totally “black-box” and is hard to clarify the training process semantically for researchers.In light of this situation,this paper proposes an improved structure of primitive LSTM (Multi-Scale Hierarchical Long Short-Term Memory,MSH-LSTM).It retains the common procedure widely used in neural network,combines the structure of neural network and human’s prior knowledge,enables the network to train purposefully under the guidance and solving the problem of “black-box” in a way,ultimately resulting in much better analytic results on time series data.To illustrate the effectiveness of MSH-LSTM,two groups of experiments (temperature and stock-price respectively) were carried out.Experimental results demonstrate that the proposed MSH-LSTM outperforms primitive ANN,LSTM and GRU without loss of network’s applicability.In temperature experiment,MSH-LSTM,primitive LSTM and primitive GRU use temporal information to get approximately results,which are better than primitive ANN.In stock price experiment,MSH-LSTM’ssuperiority is more obvious.The error of MAPE of MSH-LSTM is increaced by an average of 19.65%,24.35%,46.3% compared with that of primitive LSTM,GRU and ANN,respectively.

Key words: Hierarchical network, LSTM, Recursive neural network, Short-term forecast, Time series