A Generative Adversarial Network with Attention Mechanism for Time Series Forecasting

Time series forecasting has widespread applications in everyday life, e.g. in the domains of weather forecasting, transportation planning, and electricity demand prediction. Due to the high dimensions, uncertainties, and changes of data in time series, accurately predicting future trends of time series remains a challenge. In this paper, we propose a new time series forecasting model, the Generative Adversarial Network Model with Attention Mechanism (GANAM), which uses a global residual network to enhance the attention mechanism module and acts as a generator to learn the deep temporal features and then uses the discriminator of the generative adversarial network for improving the forecasting capability. In addition, the classical time series model can lead to an accumulation of errors in the multi-step prediction process. The proposed model also can effectively reduce this problem by utilizing the generative adversarial network with an attention mechanism. The experiments on various real-world datasets have demonstrated the effectiveness of our model.