policies that lead to the minimum injected water and the best ThM paper w= prepared for p!'cstnlatmn ut [he Wih SPE Western Rcgmnal Mceung held , n Anchorage AK, 22.24 May l!)% 'This paper wu selected for presentalmn hy [hc SPE Program Comm!uce followlng rcwew of miormation ccmtamcd In tin ahs!ract suhmtttcd hy the aughtx( s). Contents of the paper as presented have mu hccn rewewcc hy the Sccae[y of Pctmlcum Englncers and are suhjcct to correction(s) hy (he authr>r(s) The material, as presented, does no[ nccesswdy reflect any pclwtiofi :,f (he SocIeIY O( Pctmlcum Engineers or IIS mcmhers Papers prrsenmd at SPE meel,n~s arc subject [n puhhciwm rewew hy Edthmal Commmecs of the Srciety of Petroleum Engineers. Penmssmn to copy IS rcsmcld to m ahstmcl of noi more than 3(H) words Illuslratmns may not he cop.sd The abstract should contain consplcums acknowledgment of where and hy whom tic pa~r WM prcscn!cd.
AbstractAn optimal water injection policy maximizes oil recovery per barrel of injected water while minimizing formation darnage and maintaining reservoir pressure. Optimal water injection into low permeability, tiactr.md oil reservoirs is problematic because of highly nonlinear and complex reservoir dynamics. Likewise, current first principle models of fluid movement in fractured, low permeability rock systems are insufficient to design, operate, and predict the performance of large scale waterflood, Historically, the conflict between prudent reservoir management and meeting field injection-production targets has resulted in reservoir and well damage, injectant recirculation and irreversibly lost oil production.Here we present the next generation of "intelligent" field surveillance and prediction software based on neural networks and implemented on a PC. We demonstrate a new approach to field-wise performance prediction and optimization cf waterfloods that recognizes an oil field as a coupled, highly nonlinear system of injectors and producers. With lease-wide historical data tlom a waterflood in the Lost Hills Diatomite (Kern County, CA), we construct several neural networks which recognize that individual well behavior may depend on well history and the injection-production conditions of surrounding wells. Some of our neural networks accurately predict wellhead pressure as a function of injection rate, and vice versa, for all injectors. Other networks history-match oi I and water production on the well-by-well basis, and predict firture production on a quarterly or half-year basis. Finally, our neural networks recognize and suggest water injection oil recove2y.
