Assessment of Produced Water Injection Performance during Waterflooding of a North Sea Field Using Data Mining Techniques

Karim S. Zaki; Ahmed S. Abou-Sayed; Edwin Arden Roehl
Presented at SPE Offshore Europe Oil and Gas Exhibition and Conference, Aberdeen, United Kingdom, September 2005.


This paper presents a study in which injector performance is evaluated during water flooding operations in a North-Sea Field. It is well known that water quality and well conditions strongly affect injection performance. Poor water quality leads to plugging and bacterial growth which results in a loss of permeability and injectivity decline. The effect of poor water quality can be mitigated through the treatment and removal of harmful solids, dissolved oxygen (DO) and bisulfites (BI). Long-term remideation can be achieved through various stimulation techniques. However, technology limits and frequent treatment plant upsets can negate the effects of these mitigations and frequent stimulations would result in significant costs. Optimization of injection operations depends on the selection of the proper strategy and answering questions such as “Which water contaminants have the biggest impact?” and “How much would stimulation improve injectivity?” Surprisingly, reliable answers are difficult to obtain because interactions among injection variables are obscured by highly-complex, process dynamics (time-dependencies).In this study, data mining techniques were applied in order to understand factors that affect field injection performance.The study used 11 years of injection and water quality data from 14 wells in two different blocks. Dynamic models of well behavior were synthesized using a combination of “artificial neural networks” (ANN), a machine learning technique from the field of Artificial Intelligence and “multivariate state space reconstruction” from the Chaos Theory. Sensitivity analysis performed using the ANNs revealed the relative impacts of each variable on injector performance. Despite different histories for each well, the models were relatively uniform in quantifying the relationships between the variables.Findings included – a different response to acidizing in each block; delayed cumulative formation damage from each variable of 1–3 months; the impact of DO was more than twice that of particulate matter (PM)and BI; and clorides (CL) had only a small affect on suppressing bioactivity. The uniformity and clarity of these results reduce uncertainties and provide operators with detailed knowledge to help optimize their injection designs.