Enhanced Deep Learning Method for Natural Gas Pipeline Flow Prediction Based on Integrated Learning
Yunhao Li, Changjing Sun, Qiang Li
Abstract
Urban gas pipelines must contend with situations such as road construction and excavation for house building, where short-term emergencies leading to large-scale leaks pose significant risks to both people and the environment. To enhance the response cycle for detecting leaks in urban natural gas pipelines, this paper proposes a real-time flow prediction model for gas pipelines. This model is an improved version of the Long Short-Term Memory (LSTM) neural network, utilizing an ensemble learning algorithm. It processes the instant flow data from preprocessed historical flow meters as input and fine-tunes the neural network’s hyperparameters through grid search. The LSTM, with its inherent temporal memory function, serves as a weak predictor within the ensemble, which is then strengthened through a weighted combination using the Adaboost ensemble learning algorithm. The findings indicate that our approach, in comparison to a singular LSTM network, yields lower Mean Squared Error (MSE), Mean Absolute Error (MAE), and Symmetric Mean Absolute Percentage Error (SMAPE). The enhanced LSTM model with ensemble learning significantly improves time-series forecasting accuracy, exhibiting robust generalization and stable predictive performance, thus providing critical insights for real-time monitoring and intelligent alarm systems in urban gas networks.