Slug Forecasting and Production Optimization Using Deep Learning

Abstract Subsea fields characterized by multiphase reservoir, deep wells, and pipeline-riser setups, are prone to slugging. Sustained slugs arriving at surface facilities, cause several operational challenges leading to reduced production or field shutdown. Reliable forecasting of slugging behavior is crucial in mitigating slugging and optimizing field behavior. The goal of this work was to assess potential of deep learning models for slugging forecasting and production optimization. Historical data from several field transmitters which record signatures specific to slugging events was collected and studied. Domain experts classified normal flow, unstable flow, and slug flow conditions in the field based on those data. Several deep learning-based models were trained on the supervised data to generalize correlations for slug identification and forecasting. Experiments were conducted on the labelled data to identify best suitable prediction model. A state-of-the-art temporal fusion transformer model was selected for time series forecasting based on experiments. The real-time and historical data possess noisy signals and skewed distribution depending upon frequency of slug events. The deep learning model was found to overcome these challenges and successfully learn the complex interaction in time between various field variables and correlate with slugging patterns. Not only was the model able to classify between slug and no-slug events, but it was further able to forecast slugging critical variables such as slug frequency, amplitude, slug volume, etc. The selected model outperformed traditional statistical models such as Non-Parametric Time Series (NPTS) and other deep neural network models such as DeepAR, and Simple Feed Forward. The model was able to match the flow behavior during slugging events with continuously improved performance after multiple training exercises on additional data. This novel solution can be used for active monitoring of slugging conditions in field, generating alerts and provide forecasting on slug critical information. Visibility in the future will enable understanding of slugging behavior and empower the operators to take proactive mitigation measures based on predicted slugging signatures. This will improve the operational efficiency and hence minimize the negative impact of slugging on productivity, asset integrity and the environment.