Research On Machine Learning Based Medium And Short Term Natural Gas Load Prediction For Urban Areas

In recent years,the global gas industry has garnered significant attention,and the proportion of gas in primary energy has been steadily increasing.Accurately predicting gas load is crucial for ensuring stable gas use,achieving economic benefits,and enabling precise scheduling.As such,the purpose of this paper is to achieve precise gas load prediction by undertaking the following tasks:To address the challenge of large seasonal fluctuations in gas load caused by factors such as holidays and extreme weather,this paper analyzes the correlation between load abrupt changes and commonly used features such as temperature and date.Additionally,noise augmentation techniques are applied to enhance the dataset to tackle the issue of imbalanced samples in the dataset.And the Boruta algorithm is used for feature selection to eliminate redundant features.In order to address the issue of large short-term fluctuations in gas load,this paper adopts the seasonal and Trend decomposition based on Loess to decompose the gas load sequence to reduce its volatility.In addition,to solve the problem of imprecise scheduling caused by the difficulty in capturing time series features,a hybrid neural network composed of a one-dimensional convolutional network and a series of gated recurrent unit networks is proposed to extract the time series features.Furthermore,to address the problem of significant accuracy decline in the later period when predicting weekly loads using existing models,this paper replaces the expert network of the multi-gate hybrid expert network framework with a hybrid neural network for prediction.To address the problems of low prediction accuracy due to limited data volume,difficult feature extraction for mid-term gas load,this paper uses the dual grey model and the adaptive boosting tree model to predict the mid-term load(monthly total)separately.Specifically,the first grey prediction solves the problem of traditional grey prediction models’ difficulty in tracking periodic sequences by spanning a year.Furthermore,the second grey prediction compensates for errors to improve accuracy.And the adaptive boosting uses the monthly average load data obtained by averaging the daily load,overcoming the problem of small mid-term load data.To further improve the noise immunity and prediction accuracy of the model,the final prediction result is made using stacking ensemble learning combined with the above two models.Finally,experiments are conducted on real city gas load data sets for daily load prediction for short-term load,weekly load prediction,and total monthly load prediction for medium-term load,and the results prove the effectiveness and practicality of the model proposed in this paper.