A novel investigation on oxidation behavior of crude oil during in-situ combustion based on neural networks

In-situ combustion (ISC) was regarded as a highly prospective thermal recovery technique, and its development performance was closely related to the oxidation behavior of crude oil. To accurately predict and evaluate crude oil oxidation behavior, high-pressure differential scanning calorimetry was employed to acquire heat release data at different heating rates. A dataset was then constructed using heating rate and temperature as the input variables and oxidation heat flow as the output variable. On this basis, prediction models based on a standard back propagation neural network (BP), a genetic algorithm-BP neural network (GA-BP), and a particle swarm optimization-BP neural network (PSO-BP) were established. The correlation coefficient (R) and root mean square error (RMSE) were utilized to assess the performance of three models. Furthermore, after determining the optimal network structure and training function, the particle swarm optimization (PSO) was introduced to refine the initial weights and biases of the model. Accordingly, the final optimized PSO-BP prediction model was constructed for accurate prediction and evaluation of crude oil oxidation behavior. The results showed that in comparison with the standard BP and GA-BP models, the PSO-BP model achieved greater fitting accuracy and lower prediction errors on the training and test sets, showing that PSO effectively enhanced the capability of the model to predict the nonlinear oxidation process of crude oil. The optimized PSO-BP model still fitted the experimental results well at a heating rate of 15 ℃/min, which verified its favorable generalization ability and stability. The proposed method enabled rapid and accurate prediction of crude oil oxidation behavior under different operating conditions, while significantly reducing the workload required for thermal analysis experiments.