Research On Intelligent Identification Method Of Well Leakage Risk State Based On Extreme Learning Machine

Well loss is a common downhole complication in the drilling process.Once it happens,it will cause the loss of drilling cost and production time.Using logging parameters to give real-time warning of well loss risk can prevent further development of well loss and avoid greater loss of time and cost,which has practical research significance.Therefore,this paper takes an LD drilling completion in Ledong block,Yingqiong Basin,South China Sea as an example to carry out the following research.(1)Based on the mechanism of well loss,two major risk factors of well loss were obtained by accident tree analysis: formation condition,drilling fluid performance and improper drilling operation.Based on the offset well and logging information,the formation in which the drilling was completed was judged to be a high risk formation with narrow pressure window,narrow safe density window and induced fracture loss.The engineering parameters and drilling fluid parameters related to well loss were obtained by analyzing the corresponding relationship between the characterization rule of well loss and the change of real-time logging parameters.(2)The data of well loss events under drilling,circulation and static working conditions were screened,and the labels of three data sets of normal,suspected and lost well were calibrated.High-dimensional logging parameters were selected and a series of pre-treatments were carried out.K-means clustering algorithm was used to eliminate the supplementary data.Filter the noisy data;Principal Component Analysis was carried out for high dimensional characteristic parameters.Normalized the data before input to the model.(3)The effect of two parameter combinations of high-dimensional parameter group and low-dimensional parameter group after PCA treatment on the model’s early-warning effect was compared,and the accuracy of PCA parameter group was verified to be higher.Aiming at the different variation trends of loss-related logging parameters under drilling,circulating and static conditions,SSA,SSA-ELM,BP and RF were compared.The PCA-SSA-ELM early warning model was optimized by considering the three working conditions comprehensively.The average accuracy of the model under the three working conditions was 99.54%,and the average F1 score was99.52%.(4)Using MATLAB software and SQL Server database,based on PCA-SSA-ELM algorithm,a real-time risk warning software for well loss was developed,which realized the function modules of user management,data transmission and storage,real-time monitoring and warning,event input and query.