Accurate prediction of reservoir production dynamics remains a key challenge in the oil and gas industry, especially for complex, high-dimensional time-series data. Conventional models fail to capture temporal dependencies, while existing hybrid models suffer from high parameter complexity and lack automated hyperparameter tuning, increasing training difficulty. To address these issues, this study proposes a novel hybrid model, TCN-LSTM-AVOA, combining a temporal convolutional network (TCN) with a long short-term memory network (LSTM) and incorporating the African Vulture Optimization Algorithm (AVOA) to enhance forecasting accuracy. The model not only captures complex temporal relationships and nonlinear features in reservoir data but also facilitates automated tuning of critical hyperparameters (e.g., the number of TCN kernels, LSTM units, batch size, and learning rate), which significantly enhances its robustness. Compared to eight benchmark models (back propagation neural network (BPNN), LSTM, convolutional neural network(CNN)-LSTM, TCN-LSTM, LSTM-AVOA, CNN-AVOA, TCN-AVOA), TCN-LSTM-AVOA achieves superior performance on a two-dimensional, three-phase heterogeneous reservoir, yielding a root mean square error (RMSE) of 7.0806, mean absolute error (MAE) of 3.4780, coefficient of determination (R2) of 0.9975, and mean absolute percentage error (MAPE) of 1.81%. This work demonstrates a more accurate and efficient methodology for reservoir production prediction, with significant potential for oilfield production optimization and resource management.
Xin et al. (Tue,) studied this question.