Highly Accurate Short-Term Gas Consumption and Elapsed Time Forecasting Using Multi-Channel Deep Neural Network

Abstract

Accurate gas consumption and elapsed time forecasting can help decision-makers detect anomaly gas usage and notify users to recognize the facility fault in real-time. However, it is challenging due to its variable and complex factors. This paper proposed a novel deep model, named multi-channel DNN (MC-DNN), for gas consumption and elapsed time forecasting by applying three time-series channels, including time variable, temperature, and historical gas consumption. Significantly, the time variable channel records time information to reflect the heating process pattern; the environmental channel records the inside temperature of the furnace to reflect the environmental factors and the raw consumption channel provides the historical gas consumption trend. A fully connected deep neural network (DNN) is adopted to extract the rich hidden patterns from three parallel channels in the proposed deep model. The extracted hidden features are combined as the fusion features for gas consumption and elapsed time forecasting. Besides, two-loss functions are utilized to connect two tasks, which increase the forecasting accuracy again. The authors verified the proposed method on the real-time dataset of the forging factory using multiple evaluation metrics. Massive experiments have proved that it can accurately forecast gas consumption and elapsed time at the same time.

Keywords

• Smart factory
• gas consumption forecasting
• deep learning
• multivariate time series forecasting