A deep belief network based precipitation forecast approach using multiple environmental factors

Abstract

Precise precipitation forecast can better reflect the changing trend of climate, provide timely and efficient environmental information for management decision, as well as help people to make preparations for the incoming floods or droughts. However, existing approaches have limited ability to forecast future precipitation in different regions. In order to addess the problem, this paper proposes a big data based approach for precipitation forecasting based on deep belief nets, called DBNPF (Deep Belief Network for Precipitation Forecast). The proposed approach can not only learn the hierarchical representation of raw data using a highly generalized way, but also make a more accurate description of the rule underlying different kind of environmental factors. A set of dedicated experiments with hydrological multivariate time series from four typical areas of China is conducted to validate the feasibility and robustness of the model. In the experiments, environmental factors, filtered by factor analysis, are used as input vector, and the next 24 hours precipitation is used as the output vector. We compare DBNPF with other traditional machine learning approaches. The experimental results show that the proposed approach is more robust than other approaches and can also improve the forecast precision.