Recurrent Neural Network Based Gating for Natural Gas Load Prediction System

Prediction of natural gas consumption is an important element in gas load management aimed to better utilize the facilities of a gas distribution system. The major challenges faced by developers of prediction systems are the variety and volatility of consumer profiles, strong seasonal dependency and dependency on climatic conditions, and lack of extensive and reliable historical data. In this paper, the problem of seasonal dependency is tackled with a recurrent neural network used as a gate for a statistical mixture model. Historical consumption data along with climatic conditions and other auxiliary descriptors are combined with expert delineation of heating season boundaries to provide training data. The resulting gating system is capable of reliable identification of the start and end of the heating season and, combined with the statistical models, of accurate predictions of gas load.