Generating Synthetic Temperature Surveys for Wells Through Subsurface Spatial Machine Learning Modeling and Time Series Forecasting

Abstract

Abstract Temperature profiles are key element to ensure a healthy well life specially down-hole tubular integrity. Temperature surveys are routinely conducted and compared with a base well temperature survey, recorded soon after a well was completed. This paper demonstrates novel approach to synthesize virtual temperature surveys by data-driven meta model combining subsurface geospatial modeling and time series forecasting. The proposed model anticipates the next temperature survey profile utilizing historical temperature surveys from same well along with different wells with similar temperature behavior. Historical surveys for same well are fed as vectors into a time series forecasting model that predicts the upcoming temperature profile. While historical surveys for offset wells are fed into a 3D spatial correlation model that computes the temperature values at target well location. The final step is a probabilistic meta model trained to combine the outcomes of the two models to accurately predicted temperature profile for a target well. This methodology has been piloted and used to predict temperature profiles for entire wells in the area using a dataset of historical surveys summing up to 29000 measurements. Predicted temperature values were evaluated using Root Mean Squared Error between predicted and actual temperature values. First step is a 3D Kriging spatial correlation which was applied to predict the temperature profile at target well location. Leave-One-Out evaluation was used where, at each round, one well survey is left out of training set, and used as evaluation. Average RMSE from all rounds was 4.43. The second step is time series prediction where LSTM RNN model was applied with 70-20-10 data division. The modeling cycle included tuning of LSTM hyperparameters in addition to optimizing the number of historical surveys used as predictor of next survey. RMSE for training, evaluation and test was 2.96, 3.6 and 4.2. In the third step, predicted temperature surveys using Kriging and LSTM were input into a probabilistic model that makes the final prediction. Final model error was 3.2 using Leave-One-Out evaluation approach. This paper includes more comprehensive evaluation of the synthetic temperature surveys’ accuracy and ability to predict abnormal behavior. This paper presents a new approach to proactively assess the well healthiness and lower the cost of conventional temperature surveys by using data driven models to predict temperature profiles. These models not only capture historical trends of changing temperature inside the wellbore, but also address spatially distributed environmental trends that are not measured frequently such as formation properties and fluid distribution.