Accurate prediction of the gas-oil ratio (GOR) curve is essential for ensuring the efficient and safe development of CO2-driven reservoirs. Actual GOR curves exhibit complex characteristics of long-term trends and short-term fluctuations. It poses significant challenges to the existing univariate LSTM models in rapid response to changes and anomaly detection, limiting their effectiveness in such applications. To address these issues, this paper proposes a multi-input LSTM model that integrates constrained time series variables of water cut curves and introducing a sliding time window simultaneously. The model also corresponding benefits from the more information of constrained time series and the balance between long- and short-term data. The results indicate that incorporating water cut as a constrained time series variable improves the accuracy of GOR prediction, where the coefficient of determination (R2) increased from 0.23 in the univariate model to 0.78, demonstrating a significant enhancement in the model's ability to explain data variability. The mean absolute error (MAE) decreased from 231.4 to 135.0, and the root mean square error (RMSE) was reduced from 517.5 to 273.9, further confirming the enhanced predictive capabilities of the model. The improved performance of the model can be attributed to two main factors. First, the impact results show that changes in water cut status typically precede significant increases in GOR. It provides the model with early signals of GOR changes and revealing the causal relationship between them, which serves as a critical basis for prediction. Second, the introduction of water cut into the multi-input model expands the input dimensions and deepens the understanding of the interactions among different features in time series data, helping to improve prediction accuracy. The proposed multi-input LSTM model enhances the accuracy of predicting overall trends in the GOR curve and increases sensitivity to fluctuations. It provides engineers with predictive tools for early warning of CO2 breakthroughs and offers important means to reduce the uncertain risks associated with CO2 breakthroughs in oil fields.