NGCU: A New RNN Model for Time-Series Data Prediction

Abstract

With the rapid development of machine learning, a possibility is provided for high-precision prediction of time-series. This paper proposes a new unit which is called New Gate Control Unit (NGCU) based on Recurrent Neural Networks (RNN). The proposal of NGCU is mainly used for prediction of time-series data. NGCU alleviates the problems of gradient disappearance and explosion of traditional RNN. Compared with Long Short-Term Memory (LSTM) and Gate Recurrent Unit (GRU), NGCU improves not only the computational complexity of gating unit but also the sensitivity of model learning. To verify the accuracy, efficiency and feasibility of NGCU, this paper uses RNN, LSTM and GRU to conduct comparative experiments, and uses three different data of air quality, Hang Seng Index, and gold future price to prove the generalization of NGCU. Mean Absolute Error (MAE), Mean Squared Error (MSE), Explained Variance Score (EVS), R 2 and training time are used to evaluate experimental results. Among the three different data prediction results, the R 2 of NGCU is 0.9736, 0.9872, and 0.9231, respectively. And NGCU's MAE, MSE, EVS are also the best. Compared with LSTM and GRU, NGCU has the least training time, which is 323.5261s, 53.3257s, and 43.4814s respectively.