In the last decade, the underground gas storage business has transformed from the typical seasonal operation model to a more market-oriented operation that allows production and injection of gas at any time of the year. To satisfy the new operational demands an efficient and reliable means of forecasting the wellhead pressure of gas storage wells in dependency of injection and withdrawal rates is required. The application of artificial intelligence technology, namely neural networks, for such specialized tasks could be a viable alternative to sophisticated system simulation models.
The proposed approach is based on purely data-driven models to predict the wellhead pressure of gas wells in underground gas storage. This method comprises the necessary steps in the gathering and processing of the required data for further utilization in different prediction models. The aim of the automated data processing is to assure that a high-quality set of information is fed into the prediction models. These models are fully data-driven and include feed forward and Elman recurrent networks, nonlinear autoregressive models with exogenous input (NARX) and fully recurrent networks. All models are trained on an identical data set and their predictive performance is evaluated on the already known training information and a statistically independent test data set.
The training process of the different models showed that a high-quality data set is a keystone for the successful application of such predictive models. The data used to train are measured directly at the wellhead and then resampled to hourly time steps by using a combination of Gaussian kernel regression and median filtering. This is coupled with a rigorous quality control procedure to ensure that the resampled data still carries all the essential information while erroneous data is cancelled out. The network performances are compared by different error metrics applied to the training data and to statistically independent test data sets. The most significant finding is that all recurrent models exhibit much better predictive capabilities than the feed forward network with the fully recurrent network being the best forecasting model. The ability of these network types to manage temporal dependencies greatly enhances their performance. The achieved results confirm the applicability of recurrent neural networks to predict wellhead pressures in a range of about 1 percent of the maximal operating pressure.
In comparison to the briefly outlined approach of applying neural networks, the classical way of modelling such a plant would be to use a sophisticated system of different simulation applications including a reservoir simulator coupled to surface network modelling software. The preparation and maintenance of such models is a tedious and time-consuming task, but badly maintained or outdated model are not fit for purpose when required. Once a forecasting system based on neural network technology is trained and validated against real data, keeping it up to date is a much less expensive job. The trained model then allows forecasting numerous different operating scenarios of the gas storage plant in much shorter time.
The ability to model the short-term behaviour of an underground gas storage facility with a satisfactory accuracy opens many ways of optimization. As an example, this enables the gas storage operator to trade spare storage capacity in an efficient and profitable way on the gas spot market. Furthermore, an optimized plant operation, which reduces wear, would potentially decrease unscheduled downtimes and consequently result in reduced operating expenditures. All in all a better understanding of the plant’s operating capabilities substantiated by a reliable forecasting system allows utilizing it to full capacity.
Have you signed up for October's Gastalk webinar? Do not miss out on hearing presenter, Susan Sakmar, and leading analyst from Wood Mackenzie, Murray Douglas, discuss the biggest stories this month in Europe! Sign up for FREE today!
Image courtesy of RAG Austria
Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.