Design And Application Of Feature Guided Sequence Prediction Network

Sequence prediction has been widely used in the fields of traffic forecasting,economic statistics,weather forecasting,food production and energy consumption.In oil and gas exploration,the use of seismic data to predict well data can be regarded as a sequence prediction problem.The use of sequence prediction algorithms to extract the correlation between seismic data and well data and build an accurate lithology model has guiding significance for oil and gas exploration and development.With the development of artificial intelligence,more and more machine learning algorithms are used in sequence prediction.Traditional machine learning algorithms can mine nonlinear relationships in sequences,but in complex large-scale data,features are difficult to accurately extract,resulting in unsatisfactory sequence prediction results;algorithms based on deep learning can effectively extract massive amounts of data that cannot be used the relationship expressed by mathematical formula,but when the sample is insufficient or there is a sample conflict problem,the prediction result is unreliable.The correlation between seismic data and well data is complex,and there are few label samples.Aiming at this feature,this paper takes seismic data and logging data as examples,based on the idea of feature classification and regression,constructs a feature-guided sequence prediction model,and applies it to the exploration data of the actual work area to construct the wave of the target work area.Impedance model.The main research work of this paper is as follows:(1)Propose a sequence prediction network model guided by periodic features.For time series data,existing methods cannot effectively extract the unique time characteristics contained in the data when predicting time series data.Taking the wave impedance sequence and the corresponding well side channel sequence as an example,this paper proposes a period feature-guided time sequence prediction network model.The main idea is to first use one-dimensional convolution to extract the long and short period features in the sequence.For features,LSTM is used for feature selection,and finally feature fusion is performed to extract the correlation relationship,complete the prediction of the seismic sequence corresponding to the seismic sequence in the work area,and construct the impedance model.Experimental data proves that when extracting the wellseismic relationship with the network guided by periodic characteristics,the convergence speed of the network is accelerated,the fitting ability is strengthened,the predicted result is more accurate,and the constructed wave impedance model is closer to the actual seismic data.(2)Propose a sequence prediction network model based on spatiotemporal feature guidance.For spatiotemporal data,the existing methods have limited extraction of spatiotemporal features in complex spatiotemporal data.Taking 3D seismic data as an example,this paper proposes a sequence prediction network model based on spatiotemporal feature guidance.The main idea is to use cavity convolution to extract spatial features in 3D seismic data,and further use 3D convolution to extract temporal and spatial features in seismic data.Then use LSTM to select the spatio-temporal features,perform feature fusion,complete the extraction of the association relationship,and predict the wave impedance sequence corresponding to the three-dimensional seismic sequence to construct a reliable lithology model.Experimental data shows that the network not only maintains the accuracy of blind well prediction,but also has good continuity in the lateral direction,and the horizontal slice has a high degree of consistency with seismic data.