Abstract:
In the existing multivariate time series prediction methods, traditional models cannot capture short-term mutation during long time series, which leads to the delay of prediction trend and large prediction error. This paper proposes a short-term information enhancement model, termed as clockwork triggered long short term memory (CWTLSTM) neural network. This new model groups neurons in the network and assigns different activation frequencies to each group. Each group of neurons is activated only when the time step is equal to an integer multiple of their specified period. According to whether the group period is 1, the network is divided into backbone network chain and short-term input enhancement chain. When the short-term input enhancement chain is activated on the time step close to the output position, the operation result of the input information at that point is transmitted to the backbone network chain in one direction to enhance the weight of short-term input data. Thus, this model can quickly respond to the data fluctuation caused by short-term mutation information on the basis of storing long-term information. Compared with LSTM, XGboost and CWRNN models, the prediction performance of CWTLSTM is verified by air pollution data set and cement cooler data set. It is shown via experiment results that the proposed model has good performance in reducing forecasting error and forecasting future trend. In the experiment, the parameter sensitivity of the model to the periodic allocation strategy is also analyzed and the role of CWTLSTM in short-term information enhancement is illustrated.