Research On Prediction Of Natural Gas Price And Consumption Based On Machine Learning Methods

Energy is the material basis for the survival and development of human society,and has a particularly important strategic position in the national economy.Whether in social,economic,or environmental aspects,energy is crucial to the sustainable development of any country.More than one-fifth of the global primary energy is natural gas,which is universally considered as clean fossil fuel.Price is one of the important market signals,whose changes are bound to bring about a notable impact on natural gas trade and global energy markets.To predict natural gas prices accurately can be beneficial to optimize decision management and avoid potential risk,as well as energy planning and regulatory decisions.As the increase of natural gas demand,the consumption of natural gas is also an important measure of a country’s economic power.To predict the consumption of natural gas accurately can promote the resources’ saving and utilization,effectively reduce the cost,and better make productive planning and infrastructure construction.Therefore,the accurate predictions of natural gas prices and consumption not only provide an important guidance for effective planning and the implementation of energy policy,but also have a far-reaching influence on economic planning,energy investment,environmental protection,etc.It is an important work to achieve the goal of carbon peak and carbon neutralization.This thesis focuses on the predictions of natural gas prices and consumption in the United States by machine learning.The main contributions are as follows:(1)A machine learning approach called gradient boosting machines is introduced to forecast natural gas prices.At present,this method has not been applied in the field of natural gas price prediction.The spot prices of natural gas in Henry Hub are investigated over the period from January 2001 to December 2018,covering several factors including heating oil price,crude oil price,natural gas rotary rig count,marketed production,consumption,underground storage,and imports.Gradient boosting machines algorithm is used to respectively forecast daily,weekly,and monthly natural gas spot price.Through a series of data pretreatment operations,from daily data,weekly data,monthly data three different dimensions of the gradient intensifier parameters repeatedly adjusted.The empirical results show that the proposed model based on gradient boosting machines possesses an outstanding predicting performance and an advantage in terms of the natural gas price prediction.(2)Four data-driven price prediction models of natural gas based on machine learning are discussed,which are artificial neural network,support vector machine,gradient boosting machine,and Gaussian process regression.Through model evaluation and parameter selection,the performance of these four representative machine learning methods in natural gas price forecasting is analyzed and compared.Among them,the nonlinear autoregressive model with exogenous input is selected in the artificial neural network.The quadratic kernel function with small root mean square error is selected as the kernel function of support vector machine;The least square loss function suitable for regression model is used in gradient boosting machine.The square exponential kernel with small root mean square error is selected as the kernel function of Gaussian process regression.Four prediction models are built up by adopting the monthly spot price data from January 2001 to December 2018 in Henry Hub.The input variables include heating oil price,crude oil price,natural gas rotary rig count,heating degree days,cooling degree days,marketed production,consumption,underground reserves,imports,and historical price.The empirical results show that all the proposed four machine learning prediction methods are effective for predicting natural gas prices.Overall,the prediction performances of artificial neural network and support vector machine are better than that of gradient boosting machine and Gaussian process regression.More precisely,artificial neural network is the best,while gradient boosting machine is worse than others but can directly identify important features and the importance of predictor.(3)A new hybrid prediction model,which combines deep learning with denoising technology based on compressed sensing,is proposed for predicting natural gas price.In the proposed model,the denoising technology based on compressed sensing is used as a pre-processing step to denoise the original data of natural gas spot price to get the cleaned data.Then,a powerful deep learning algorithm is used to model the clean data and then give the final prediction results.The natural gas spot price data in Henry Hub are used as sampled data and the empirical results show that the proposed hybrid model that sufficiently associates the advantages of both deep learning and compressed sensing is superior to the single model in terms of both level and direction prediction.Meanwhile,such a denoising approach can significantly improve the prediction ability of the deep learning model.(4)Data-driven models for natural gas consumption prediction are studied,which are based on four machine learning methods,including artificial neural networks,support vector machine,gradient boosting machine,and Gaussian process regression.Four prediction models are established using monthly data of American natural gas consumption from January 2001 to December 2018.Industrial production index,personal income,population,heating degree-days,cooling degree-days,marketed production,net imports,net inventory extraction,price and historical consumption are considered as the input variables.The proposed models are also used to separately predict the natural gas consumption in the fields of residence,commerce,industry,electric power,and transportation.Finally,the empirical results show that the four machine learning prediction methods are effective for predicting natural gas consumption.In general,the prediction performances of Gaussian process regression and artificial neural network are better than that of support vector machine and gradient enhancer.Gaussian process regression is slightly better than other methods.