Study on Model of Natural Gas Hydrate Formation Based on Extremely Randomized Trees

Abstract

Abstract It is easy to form hydrates in the development of natural gas. It is of great significance to study the formation of prediction models to guide the safe production of oil and gas fields. Based on computer intelligence algorithms, a prediction model of natural gas hydrate formation based on Extremely randomized trees was established and compared with the BP Neral Network model. To objectively evaluate the predictive power of the model, an extensive database of more than 1000 hydrate formation conditions was established. The results show that the number of optimal decision trees for the Extremely randomized trees model is 6, and the decision tree depth is 32. The BP Neral Network model has a flat error distribution with a maximum error of 6.37%. The error distribution of the Extremely randomized trees model is abrupt, with a maximum error of 3.39%, with higher stability and accuracy. In terms of pure water, the BP Neral Network model performs well only in a small number of conditions due to over-fitting, but the Extremely randomized trees model can avoid over-fitting by using the large number theorem, showing a stronger advantage.