Research on the Prediction Model of Ch4 Concentration in Adjacent Spaces of Urban Underground Gas Pipeline Network

Abstract

Monitoring of combustible gas in underground space is an important means to effectively prevent gas explosion accidents. However, considering the endurance of the equipment, the sampling interval is typically between 5 minutes and 2 hours, with only 20 seconds of actual sampling, meaning that 93.75%~99.72% of the time is a monitoring blank period, which poses an explosion risk. Therefore, accurately predicting the CH4 concentration trend during this blank period is crucial. To this end, a CH4 concentration prediction model is proposed in this paper to solve the problem of the current monitoring equipment not working continuously. First, the CH4 concentration data from 542 natural gas leakage events and 1434 biogas accumulation events in H City were analyzed, examining the effects of data interpolation resolution, sliding window length, and periodic characteristics on prediction results. For natural gas leakage events, a CH4 concentration prediction model based on GA-BP time series (GA-BPM) is proposed. For biogas accumulation events, the periodic characteristics of CH4 concentration are extracted and fused using Fourier transform and Fourier series, establishing a periodic characteristic-enhanced XGBoost (PCE-XGBM) prediction model.The results show that the best prediction performance occurs when the data interpolation resolution is 5 minutes and the sliding window length is 13; The periodic variations in biogas CH4 concentration are significant, with periods of 12 hours, 1 day, 1 month, and 1 week. Compared to 25 other CH4 concentration prediction models, the XGBoost model (XGBM) achieved the highest prediction accuracy. After incorporating periodic features, the model’s accuracy improved by 48%. Verification using 25,544 biogas accumulation data points demonstrated an accuracy rate of over 95%, meeting the needs for engineering applications.

Keywords: underground space, CH4 concentration, XGBoost algorithm, time series prediction