Temporally Constrained Long Short-Term Memory Surrogate Modeling for Water-Cut Prediction in Simulated Reservoir Scenarios

Abstract

Dynamic prediction of reservoir water cut supports development-plan optimization and injection-production control. The 3000 scenarios sampled the reservoir-production-control response space, and the surrogate model learned the mapping from input conditions to dynamic responses for rapid prediction of unseen scenarios. We propose a temporally constrained long short-term memory (TC-LSTM) network that augments pointwise fitting with relative first- and second-order temporal-difference penalties. Three consecutive windows are grouped only during training to evaluate these penalties; inference retains the standard shared LSTM forward computation. Across five random-seed runs on independently split INSIM scenarios, Full TC-LSTM achieved MSE = 0.00394 ± 0.00013, MAE = 0.0432 ± 0.0007, RMSE = 0.0628 ± 0.0010, and R2 = 0.9214 ± 0.0028. Relative to LSTM, mean MSE decreased by 31.1% and R2 increased by 4.54 percentage points. Component ablation showed that grouping alone had negligible effect, first- and second-order penalties were complementary, and temporal weighting produced a further gain. Validation sensitivity supported ε = 10−3, α = 0.5, ω = (1/3, 2/3), and λ = 0.01. The results support TC-LSTM as a stable, domain-prior-regularized surrogate within the simulated parameter space.