Production Decline Rate Prediction for Offshore High Water-Cut Reservoirs by Integrating Moth–Flame Optimization with Extreme Gradient Boosting Tree

Abstract

The prediction of production decline rate in the development of offshore high water-cut reservoirs predominantly relies on the traditional Arps decline curves. However, the solution process is complex, and the interpretation efficiency is low, making it difficult to meet the demand for rapid prediction of production decline rates. To address this, this paper first identifies the key influencing factors of production decline rate through comprehensive feature engineering. Subsequently, it proposes a novel prediction method for the production decline rate in offshore high water-cut reservoirs by integrating Moth–Flame Optimization with Extreme Gradient Boosting Tree (MFO-XGBoost). This method utilizes seven dynamic and static influencing factors, namely vertical thickness, perforated thickness, shale content, permeability, crude oil viscosity, formation flow coefficient, and well deviation angle, to predict the production decline rate. The forecasting outcomes of the MFO-XGBoost method are then compared with those of standard RF, standard DT, the standalone XGBoost model, and the calculated results from the exponential decline model. Additionally, the forecasting capability of the MFO-XGBoost method is benchmarked against Particle Swarm Optimization–XGBoost (PSO-XGBoost) and Bayesian Optimization–XGBoost methods for predicting the production decline rate in offshore high water-cut reservoirs. The findings from the experiments show that the MFO-XGBoost method can achieve accurate prediction of the production decline rate in offshore high water-cut reservoirs, with a coefficient of determination (R²) reaching 0.9128, thereby providing a basis for strategies to mitigate the production decline rate.

1. Introduction

In the advancement of offshore reservoirs with a high water cut, the production of offshore oilfields typically experiences stages of increasing production, stable production, and declining production [1,2,3,4]. As offshore high water-cut oilfields enter the middle and late stages of development, many mature oilfields face the challenges of accelerated natural decline and increased difficulty in maintaining stable production [5,6,7]. Consequently, studying the production decline patterns and adopting corresponding stable production measures has become crucial. Regarding the research on oilfield production decline patterns, over 20 types of production decline models have been proposed by researchers both domestically and internationally [8,9,10,11]. Analysis reveals that most studies are based on Arps decline curves to determine the connection between the rate of production decline and the time series data [12,13,14,15]. Currently, in the advancement of offshore reservoirs with a high water cut, the prediction of production decline rate using traditional Arps decline curve theory involves a complex solution process and low interpretation efficiency, making it difficult to meet the demand for rapid prediction of production decline rates.

In recent years, data-driven machine learning (ML) techniques have increasingly emerged as a vital tool for the rapid prediction of development indicators in oil and gas field development. These methods can not only process large amounts of dynamic and static data but also provide more accurate predictions of development indicators through complex nonlinear modeling capabilities [16,17]. The swift progress of artificial intelligence (AI) technology is significantly enhancing the development of oil and gas resources like never before [18,19,20,21,22]. Alimohammadi H. et al. [23] employed a multivariate time series approach for predicting production in unconventional resources. Amr S. et al. [24] used ML models to forecast the production of horizontal wells. Bhattacharyya S. et al. [25] used ML models to predict the decline in oil cut of Bakken shale wells. Chaikine I. et al. [26] used ML models to forecast the production of wells under multi-stage hydraulic fracturing conditions, and Chaikine I. et al. [27] also employed ML methods to forecast the output of multi-stage horizontal wells. Chen X. et al. [28] used the LSTM method to predict the productivity of horizontally drilled shale gas wells after volumetric fracturing. Gao Q. et al. [29] applied artificial intelligence technology to predict the production of unconventional natural gas. Artificial intelligence technology is full of unlimited vitality in the oil and gas field [30]. However, existing studies still exhibit limitations in three key aspects: noise robustness, generalization capability under small-sample conditions, and feasibility for real-time deployment.

Therefore, addressing the issue that traditional production decline rate prediction methods in the oil and gas field can no longer meet the demand for rapid prediction, this paper develops a lightweight MFO-XGBoost framework to reduce on-site deployment costs and employs the Extreme Gradient Boosting (XGBoost) algorithm to predict the production decline rate of offshore high water-cut reservoirs. Furthermore, it utilizes the Moth–Flame Optimization (MFO) algorithm to automatically tune the hyperparameters of XGBoost, proposing an MFO-XGBoost method for predicting the production decline rate in offshore high water-cut reservoirs. By identifying the key influencing factors of production decline rate through comprehensive feature engineering, this method uses seven dynamic and static influencing factors, namely vertical thickness, perforated thickness, shale content, permeability, crude oil viscosity, formation flow coefficient, and well deviation angle, to predict the production decline rate in offshore high water-cut reservoirs. The prediction results of the MFO-XGBoost method are then compared with those of standard RF, standard DT, the standalone XGBoost model, and the calculated results from the exponential decline model. Furthermore, the prediction performance of the MFO-XGBoost method is benchmarked against PSO-XGBoost and Bayesian-XGBoost models for predicting the production decline rate in offshore high water-cut reservoirs. The results demonstrate that the MFO-XGBoost method can achieve accurate prediction of the production decline rate in offshore high water-cut reservoirs, with a coefficient of determination (R²) reaching 0.9128, thus providing a basis for strategies to mitigate the production decline rate.

2. Principle of the MFO-XGBoost Method

2.1. XGBoost Algorithm

The XGBoost [31] algorithm, building upon the foundation of gradient boosting trees, combines multiple decision trees into a strong classifier, effectively addressing the limitations of gradient boosting trees in terms of efficiency and scalability.

Given a dataset $D=\left\{\left(x_i,y_i\right)|i=1,2,\dots ,m,x_i\in R^p,y_i\in R\right\}$ consisting of m samples and p features, and given $k\left(k=1,2,\dots ,K\right)$ regression trees, where F is the space of regression trees, the model can be represented as follows:

$${\widehat{y}}_i=\sum _{k=1}^K{f}_k{x}_i,f_k\in F$$

(1)

The objective function is defined as:

$$O_{\mathrm{bj}}=\sum _{i=1}^{m}l\left(y_i,{\widehat{y}}_i\right)+\sum _{k=1}^K\mathsf{\Omega }\left(f_k\right)$$

(2)

where ${\widehat{y}}_i$ represents the predicted value and $y_i$ represents the actual value.

By incorporating a regularization term $\Omega \left(f_k\right)$, the result of the t-th iteration of the model can be expressed as shown in Equation (3):

$${\widehat{y}}_i^{\left(t\right)}=\sum _{j=1}^t{f}_k\left(x_i\right)={\widehat{y}}_i^{\left(t-1\right)}+f_t\left(x_i\right)$$

(3)

Substituting Equation (3) into Equation (2), the objective function at the t-th iteration, denoted as $O_{\mathrm{bj}}^{\left(t\right)}$, is obtained as shown in Equation (4):

$$O_{\mathrm{bj}}^{\left(t\right)}=\sum _{i=1}^{m}l\left[y_i,{\widehat{y}}_i^{\left(t-1\right)}+f_t\left(x_i\right)\right]+\mathsf{\Omega }\left(f_k\right)+\sigma$$

(4)

Carrying out a second-order Taylor expansion of the objective function while including the regularization term $\Omega \left(f_k\right)$, as defined in Equation (2), yields Equation (5):

$$\left\{\begin{array}{c}{O}_{\mathrm{bj}}^{\left(t\right)}\cong {\sum }_{i=1}^m[{\partial }_{{\widehat{y}}_i\left(t-1\right)}l\left(y_i,{\widehat{y}}_i^{\left(t-1\right)}\right)f_t\left(x_i\right)+{\displaystyle \frac{1}{2}}{\partial }_{{\widehat{y}}_i\left(t-1\right)}^{2}l\left(y_i,{\widehat{y}}_i^{\left(t-1\right)}\right)f_t^2\left(x_i\right)]+\mathsf{\Omega }\left(f_k\right)+\sigma \\ \mathsf{\Omega }\left(f_k\right)=\gamma T+{\displaystyle \frac{1}{2}}\lambda \Vert {\omega }^2\Vert \end{array}\right.$$

(5)

where T and w indicate the quantity of leaf nodes and the corresponding leaf weight value of the tree, respectively. $\gamma$ is the tree complexity penalty coefficient. $\lambda$ is the leaf weight penalty coefficient. $\sigma$ is the other items.

2.2. Moth–Flame Optimization Algorithm

Genetic algorithms (GAs) and Bayesian optimization are widely used. However, the moth–flame optimization (MFO) algorithm offers the following advantages: (1) In high-dimensional spaces, MFO reduces iteration counts by 30–50% compared to GA [32], and (2) its spiral search mechanism achieves better exploration–exploitation balance than grid search [33,34].Therefore, MFO [35] was selected to optimize the hyperparameters of the XGBoost model. The algorithmic steps of the Moth–Flame Optimization are as follows:

(1) Initialization: Randomly generate the initial positions of the moth population, calculate the fitness value of each moth, and initialize the flame positions with the moth positions.

(2) Flame update: In each iteration, merge the moth and flame populations, sort them based on their fitness values, and select the top N optimal solutions as the new generation of flames.

(3) Update of the moth position: Assign each moth to a flame and update its position accordingly. Continue this process repeatedly until either the maximum number of iterations is achieved or the best solution for the model’s hyperparameters is identified.

2.3. Workflow of the MFO-XGBoost Model for Predicting Production Decline Rate in Offshore High Water-Cut Reservoirs

The workflow of the MFO-XGBoost method for forecasting the production decline rate in offshore high water-cut reservoirs is demonstrated in Figure 1. The detailed procedures are outlined below:

(1) Data Preprocessing: Perform data preprocessing on the collected data, including the detection and removal of missing values, outliers, and duplicate entries.

(2) Feature Engineering: Employ feature engineering techniques to reduce the dimensionality of the dataset’s features, enhancing the practicality for application in actual oil and gas fields and alleviating the pressure of data acquisition for various oilfields.

(3) Dataset Partitioning and Initialization: Shuffle the dataset and divide it into a training set (70% of the samples) and a testing set (the remaining 30% of the samples). Select an appropriate fitness function, initialize the moth population, and calculate the fitness value for each moth.

(4) Position Update and Iteration: Update the position information of moths and flames. Decrease the number of flames and update the positions of moths using the corresponding formulas. The iteration process continues until the optimal values for the model hyperparameters are found or the maximum number of iterations is reached.

(5) Model Training and Prediction: Assign the optimal hyperparameters obtained to the XGBoost model and retrain it. Apply the trained XGBoost model to the testing set to obtain the final prediction results of the MFO-XGBoost method.

2.4. Evaluation Metrics

For regression problems, the following metrics are typically employed:

(1) Mean Absolute Error (MAE): The mean absolute error is calculated by averaging the absolute differences between the predicted values and the actual values. It reflects the overall deviation of the predictions, and a lower MAE indicates better performance.

$$\mathrm{MAE}={\displaystyle \frac{1}{n}}\sum _{i=1}^n|y_i-{\widehat{y}}_i|$$

(6)

(2) Mean Squared Error (MSE): The Mean Squared Error is calculated by averaging the squares of the differences between the predicted values and the actual values. It amplifies the impact of larger errors, and a lower MSE indicates better performance.

$$\mathrm{MSE}={\displaystyle \frac{1}{n}}\sum _{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2$$

(7)

(3) Root Mean Squared Error (RMSE): The RMSE is derived by taking the square root of the MSE. This process brings the metric back to the same scale as the original data, and a smaller RMSE signifies improved performance.

$$\mathrm{RMSE}=\sqrt{\mathrm{MSE}}=\sqrt{{\displaystyle \frac{1}{n}}\sum _{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}$$

(8)

(4) Coefficient of Determination (R-squared, R²): The coefficient of determination represents the model’s ability to explain the variance in the target variable. A higher R² value indicates a better fit, with values closer to 1 being desirable.

$$R^2=1-{\displaystyle \frac{{\sum }_{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}{{\sum }_{i=1}^n{\left(y_i-\overline{y}\right)}^2}}$$

(9)

In Equations (6)–(9), $y_i$ represents the true value of the production decline rate for the i-th sample in the offshore high water-cut reservoir production decline rate dataset, ${\widehat{y}}_i$ represents the predicted value of the production decline rate for the i-th sample in the offshore high water-cut reservoir production decline rate dataset, and n is the total number of samples in the offshore high water-cut reservoir production decline rate dataset. In Equation (9), $\overline{y}$ represents the mean of the true values in the offshore high water-cut reservoir production decline rate dataset.

3. Application Example

3.1. Overview of the Studied Oilfield Block

The oilfield block investigated in this study is the P oilfield block, an offshore high water-cut reservoir located in China. This block comprises 442 effective production wells, including 75 horizontal wells and 367 directional wells. The fluids in this oilfield block are characterized by high viscosity, high density, high gum content, and low asphaltene content. The formation in this block exhibits a high initial oil saturation pressure, a small pressure difference between initial and bubble-point pressures, a low dissolved gas–oil ratio, and significant variations in formation crude oil viscosity. The formation water in this block is of the sodium bicarbonate type, with a total formation water salinity ranging from 3000 to 8000 mg/L.

This oilfield block is situated within a normal reservoir temperature and pressure system. However, it exhibits weak edge water energy and insufficient natural energy, necessitating the early implementation of artificial energy supplementation for development. Upon commencement of production, the oilfield block achieved a daily oil production of 0.92 × ${10}^4{\mathrm{m}}^3$, with an overall water cut of 84.7%, a daily water injection rate of 6.19 × ${10}^4{\mathrm{m}}^3$, and a cumulative water–oil ratio of 0.89. Detailed parameters for a portion of the offshore P oilfield are presented in

Table 1. Partial parameter list of a certain offshore P oilfield.

Parameter	Value
Reservoir Oil Saturation Pressure	6.890–13.720 MPa
Reservoir Oil Viscosity	9.1–944.0 mPa·s
Average Relative Density of Natural Gas Samples	0.769
Pressure Coefficient	1.00
Pressure Gradient	0.977 MPa/100 m
Temperature Gradient	3.0 °C/100 m

.

3.2. Data Acquisition and Processing

In this study, a sample dataset was constructed based on the P offshore high water-cut oilfield block, which includes 442 effective production wells, comprising 75 horizontal wells and 367 directional wells. Information regarding the rate of production decline and both the dynamic and static factors affecting this decline rate was gathered for these 442 wells. The influencing factors include vertical thickness, perforated thickness, porosity, oil saturation, shale content, permeability, crude oil viscosity, mobility, formation flow coefficient, reservoir heterogeneity coefficient, the coefficient that describes the change in permeability at a boundary between two different reservoir zones, production pressure difference, and well deviation angle. The angle-based well deviation angle was calculated through the following steps:

(1) Obtain vertical thickness ($\alpha$) data.

(2) Obtain measured depth ($\beta$) data.

(3) Calculate the radian-based well deviation angle ($\psi$) using Equation (10):

$$\psi =\arccos {\displaystyle \frac{\alpha }{\beta }}$$

(10)

(4) Calculate the angle-based well deviation angle (${\psi }^{\prime }$) using Equation (11):

$${\psi }^{\prime }={\displaystyle \frac{\psi }{\pi }}\times 180$$

(11)

The collected sample dataset was then subjected to missing value detection. A missing value heatmap was plotted, as shown in Figure 2, to visualize the distribution of missing values within the dataset, where dark blue indicates missing values and yellow indicates no missing values. As evident from Figure 2, the reservoir heterogeneity coefficient, the coefficient that describes the change in permeability at a boundary between two different reservoir zones, and production pressure difference exhibited significant missing data. Consequently, these three features were excluded from the sample dataset. Furthermore, any other sample points containing missing values were also removed. Subsequently, the sample dataset underwent duplicate value and outlier detection, revealing no duplicate or outlier entries. After data cleaning, the statistical summary information of the sample dataset is presented in

Table 2. Mathematical statistical summary of production decline rate sample dataset for offshore high water-cut oil reservoirs.

	Count	Mean	Std	Min	25%	50%	75%	Max
Vertical Thickness/m	209	65.112	23.115	6.5	50.7	64.312	83.8	135.3
Perforated Interval/m	209	55.824	21.337	2.4	42	54.987	70.8	118.8
Porosity/%	209	26.154	1.951	20.274	25.119	26.225	27.326	30.479
Oil Saturation/%	209	66.516	7.054	48.874	63.042	66.758	71.302	84.1
Shale Content/%	209	11.785	2.32	6.168	10.376	11.591	12.912	19.813
Permeability/mD	209	942.283	335.933	293.048	711.526	943.44	1103.695	2256.5
Crude Oil Viscosity/(50 °C) mPa·s	209	175.938	143.768	48.21	80.98	124.8	193.4	830.8
Mobility/(mD/(mPa·s))	209	7.942	4.641	1.056	4.426	7.158	10.189	23.278
Reservoir Flow Coefficient/(mD/(mPa·s))	209	450.779	333.069	6.372	211.434	395.233	581.682	1622.779
Deviation Angle/°	209	46.201	16.94	0	35.211	47.027	55.087	89.01
Production Decline Rate	209	0.1	0.06	0.026	0.06	0.081	0.123	0.304

. Table 2 describes the mathematical distribution of some features in the sample dataset, including the total count, average, standard deviation, lowest value, 25th percentile, median, 75th percentile, and highest value for every column.

Feature engineering was performed on the cleaned sample dataset of production decline rate in offshore high water-cut reservoirs. Given the limitations of individual feature engineering methods, such as strong reliance on linear assumptions, sensitivity to noise and small samples, and a lack of dynamic robustness verification [36], this paper adopts a comprehensive feature engineering approach that integrates five methods: linear regression, Grey Relational Analysis (GRA), Shapley Additive Explanations (SHAP), Pearson correlation coefficient, and Mean Decrease Impurity (MDI). The workflow of this integrated approach is illustrated in Figure 3. The main steps of this method are as follows: First, the sample dataset of production decline rate in offshore high water-cut reservoirs is input into linear regression, GRA, SHAP, Pearson correlation coefficient, and MDI to perform their respective calculations and obtain the corresponding results. Second, the top 70% of features are selected based on the feature engineering evaluation criteria of each method. Then, a majority voting approach is applied, where features identified by three or more methods are considered the finally selected key influencing factors. Finally, the selected key influencing factors are output as the dominant factors for the production decline rate in the offshore high water-cut reservoir sample dataset.

The results of linear regression, GRA, SHAP, Pearson correlation coefficient, and MDI are shown in Figure 4, Figure 5, Figure 6, Figure 7, and Figure 8, respectively. In linear regression, GRA, SHAP, Pearson correlation coefficient, and MDI, a larger fitting coefficient with the production decline rate, a higher grey relational grade, a larger SHAP value, a larger absolute value of the Pearson correlation coefficient, and a higher importance score, respectively, indicate that the feature is more likely to be a key influencing factor. The key influencing factors identified by each method are then subjected to majority voting, and the results are shown in Figure 9. As shown in Figure 9, the key influencing factors for the production decline rate in the offshore high water-cut reservoir sample dataset are vertical thickness, perforated thickness, shale content, permeability, crude oil viscosity, formation flow coefficient, and well deviation angle.

3.3. MFO-XGBoost Model Training and Validation

The preprocessed and feature-engineered sample data of production decline rate in offshore high water-cut reservoirs was subjected to Z-score standardization. The data was then divided randomly into a training set and a testing set in a 70:30 ratio.

The MFO algorithm was employed to optimize the hyperparameters of the XGBoost model using the training set. The fitness function was set as the average coefficient of determination (R²) obtained from 10-fold cross-validation of the XGBoost model in each iteration. The population size was set to 50 with 200 iterations. The six hyperparameters optimized were: n_estimators (number of trees), learning_rate (step size for weight updates in each iteration), max_depth (the greatest depth of a tree), gamma (the minimum reduction in the loss function needed to justify an additional split on a leaf node), subsample (the proportion of samples to be randomly selected for each tree), and colsample_bytree (the proportion of features to be randomly selected for each tree). After 200 iterations, the coefficient of determination (R²) of the MFO-XGBoost method on the testing dataset reached a maximum value of 0.9128 and converged at the 180th iteration. The iterative optimization process of the optimal hyperparameters for MFO-XGBoost is shown in Figure 10, and the optimal hyperparameter values for the MFO-XGBoost method are listed in

Table 3. Optimal hyperparameter values for the MFO-XGBoost model.

Hyperparameter	Value
n_estimators	159
learning_rate	0.065
max_depth	3
gamma	0.001
subsample	0.527
colsample_bytree	0.711

.

The MFO-XGBoost method, built using the optimal hyperparameter combination, was used to predict the production decline rate in offshore high water-cut reservoirs. The calculated results from the exponential decline model were taken as the true values of the production decline rate. The results are shown in Figure 11. The results indicate that the MFO-XGBoost method achieved a coefficient of determination (R²) of 0.9542 on the training set and 0.9128 on the testing set. This demonstrates that the MFO-XGBoost method exhibits a good predictive performance for the production decline rate in offshore high water-cut reservoirs.

3.4. Comparison of Prediction Performance of Different Models

To further validate the superiority of the MFO-XGBoost model in predicting the production decline rate of offshore high water-cut reservoirs, the prediction results of the MFO-XGBoost method were compared with those of standard RF, standard DT, and the standalone XGBoost model, using the calculated results from the exponential decline model as the true values of the production decline rate. Additionally, MAE, MSE, RMSE, and R² were employed as evaluation metrics to quantify the prediction performance of each ML model. The results are shown in Figure 12, Figure 13 and Figure 14 and summarized in

Table 4. Comparative analysis of standard ML models vs. MFO-XGBoost.

		DT	RF	XGBoost	MFO-XGBoost
	MAE	0.0216	0.0287	0.0109	0.0101
Training Dataset	MSE	0.0011	0.0014	0.0002	0.0002
	RMSE	0.0328	0.0379	0.014	0.0134
	R2	0.7261	0.6336	0.9503	0.9542
	MAE	0.0225	0.0161	0.0141	0.011
Test Dataset	MSE	0.0008	0.0004	0.0004	0.0002
	RMSE	0.0285	0.0211	0.0191	0.0146
	R2	0.6668	0.8174	0.8503	0.9128

.

As evident from Figure 12, Figure 13 and Figure 14 and Table 4, among the standard RF, standard DT, and XGBoost models, XGBoost exhibited the best prediction performance, which inherently stems from its flexible nonlinear modeling capability and targeted optimization strategies. Compared to the standalone XGBoost model, the hyperparameter optimization component of the MFO-XGBoost method significantly improved the prediction performance.

Based on the same dataset, two other prediction models, Bayesian Optimization–XGBoost (Bayesian-XGBoost) and Particle Swarm Optimization–XGBoost (PSO-XGBoost), were also used to predict the production decline rate in offshore high water-cut reservoirs and were compared with the MFO-XGBoost method. The results are shown in Figure 15 and Figure 16 and summarized in

Table 5. Comparative analysis of standard machine learning models vs. MFO-XGBoost.

		MFO-XGBoost	Bayesian-XGBoost	PSO-XGBoost
Training Dataset	MAE	0.0101	0.0144	0.0055
	MSE	0.0002	0.0004	0.0001
	RMSE	0.0134	0.0187	0.0072
	R2	0.9542	0.9105	0.9868
Test Dataset	MAE	0.011	0.013	0.0117
	MSE	0.0002	0.0003	0.0002
	RMSE	0.0146	0.0176	0.0158
	R2	0.9128	0.873	0.8981

.

As shown in Figure 15 and Table 5, compared to the MFO-XGBoost method, the Bayesian-XGBoost model showed average prediction performance. In the testing set, there were several wells where the predicted production decline rate differed significantly from the calculated results of the exponential decline model, leading to poorer overall prediction performance. As shown in Figure 16 and Table 5, the PSO-XGBoost model exhibited overfitting, with good prediction performance on the training set but a significant decline in prediction performance on the testing set. Thus, the MFO-XGBoost method demonstrated the best prediction performance for the task of predicting the production decline rate in offshore high water-cut reservoirs. The underlying mechanisms for the superior prediction performance of the MFO-XGBoost method are as follows:

(1) MFO optimizes the critical hyperparameter combinations of XGBoost, making the model better adapted to the specific dataset’s feature distribution and noise level, thereby reducing prediction bias.

(2) By optimizing parameters such as subsample and colsample_bytree, MFO controls the randomness of data sampling, enhancing the model’s robustness. Adjusting max_depth balances the tree complexity, avoiding overfitting to the training data.

(3) The hyperparameter space of XGBoost typically exhibits non-convex and multi-modal characteristics. The population intelligence of MFO can effectively traverse this complex space and find better solutions.

4. Conclusions

To address the challenge of predicting production decline rates in offshore high water-cut oil reservoirs, this study proposes an MFO-XGBoost intelligent prediction framework. The key findings are as follows: (1) Multimodal feature fusion identified, for the first time, seven dominant control parameters: vertical thickness, perforated thickness, shale content, permeability, crude oil viscosity, formation flow coefficient, and well deviation angle. This provides a theoretical basis for targeted regulation of high water-cut reservoirs. (2) Three models—MFO-XGBoost, Bayesian-XGBoost, and PSO-XGBoost—were applied to predict production decline rates using actual data from the offshore P oilfield. The test set coefficients of determination (R²) were 0.9128, 0.873, and 0.8981, respectively, demonstrating the MFO-XGBoost model’s superior performance in predicting decline rates for offshore high water-cut reservoirs.

5. Limitations and Future Work

The current study has certain limitations due to its reliance on a specific dataset. The machine learning prediction model was validated only in the offshore P oilfield, a high water-cut reservoir in China, and has not been tested across diverse geological or operational environments. While this confirms the model’s effectiveness and accuracy under the studied conditions, future work should incorporate datasets from other geological settings to investigate potential performance variability. The authors plan to conduct additional experiments if supplementary data become available, further assessing the model’s generalizability.