Abstract
Abstract Precise forecasting of oil production rates is essential for effective reservoir management and economic decision-making in the petroleum industry. Traditional techniques such as Decline Curve Analysis and Numerical Reservoir Simulation are limited in capturing the complexity of unconventional reservoirs. This study presents a framework that applies two machine learning models: Random Forest (RF) and Long Short-Term Memory (LSTM). The models were tested on production data from a large oil field in Kazakhstan for the period 2003 to 2020. The RF model identified Water Cut as the main factor, while the LSTM model captured time-dependent patterns in the data. The R2 and MSE results for the training set were 0.98 and 281 for RF, 0.93 and 807 for LSTM, and 0.95 and 512 for the Hybrid model. When it comes to the testing dataset, the results for error metrics were 0.83 and 2769 for RF, 0.94 and 899 for LSTM, and 0.88 and 53 for the Hybrid model. These results were further supported by the model's selection criteria: the Hybrid was characterized by the minimum AIC and BIC, which demonstrates its better balance between forecasting accuracy and model parsimony. To the best of our knowledge, this is among the first attempts to utilize such a hybrid model for oil production predictions in Kazakhstan. The findings demonstrate the potential of this tool as a valuable instrument in reservoir management and production planning.