Interpretable Long-Term Forecasting Based on Dynamic Attention in Smart City

Abstract

Accurate prediction is of great significance to the construction of a smart city. However, current models only focus on mining the relationship among sequences and ignore the influence of the predicted sequences on future predictions, so we propose a Dynamic Attention Neural Network (DANN) based on encoder-decoder, which combines encoder context vectors and newly generated decoder context vectors to jointly dynamically representation learning, then generates corresponding predicted values. DANN processes data via the Bi-directional Long Short-Term Memory (Bi-LSTM) network as the fundamental structure of the network between encoder and decoder. What’s more, in order to produce a new feature representation with low redundancy, gate mechanism network module is used to adaptively learn the interdependence of multivariate feature data. The relevant experiments show that compared with baseline models, DANN has the most stable long-term prediction performance, which reduces the problem of error accumulation to a certain degree.