A dry bulk cargo port loading and unloading production process optimization method

Abstract

The application discloses a kind of dry bulk cargo port loading and unloading production process optimization method, belong to including the following steps: constructing dry bulk cargo port loading and unloading dataset, the data in dry bulk cargo port loading and unloading dataset is preprocessed, establishes multiple loading and unloading production prediction model, respectively based on the data after preprocessing is trained, and is carried out comprehensive evaluation using evaluation index, based on evaluation result determines the best prediction model;Establish multiple optimization algorithm model, respectively with shift consumption, shift loading and unloading volume, shift man-hour as prediction target, establish each single objective optimal agent model;Solving the optimal solution of process parameters under each single objective;With shift consumption, shift loading and unloading volume, shift man-hour as the core of multi-objective optimization framework is constructed, with the best prediction model as agent model, using the best optimization algorithm and constraint penalty mechanism realizes multi-index collaborative optimization, obtains the collaborative optimization optimal solution of shift consumption, shift loading and unloading volume, shift man-hour;The method of the application can improve loading and unloading operation efficiency, reduce loading and unloading operation energy consumption.

Description

Technical Field

[0001] This invention belongs to the field of dry bulk cargo port loading and unloading technology, and relates to an optimization method for the dry bulk cargo port loading and unloading production process. Background Technology

[0002] Currently, real-time control technology for dry bulk cargo port loading and unloading operations has matured, with multiple ports deploying and operating real-time control-level systems and software platforms. Among these, the Remote Control and Monitoring System (RCMS) enables core functions such as real-time status monitoring, fault alarms, historical data playback, and collision avoidance warnings for all equipment, individual machines, yards, and vessels throughout the entire process; the integrated tipping and stacking control system enables coordinated operation of tippers, belt conveyors, and stackers, supporting automatic start / stop of the operation process, uniform feeding, and fault interlocking protection; and the integrated loading and unloading control system achieves coordinated control of reclaimers, belt conveyors, and ship loaders, possessing intelligent capabilities such as automatic docking, dock transfer, material spreading, and material flow tracking. The application of these systems and software has driven the transformation of port loading and unloading control from independent single-machine control to full-process collaborative control, and provides stable and reliable data interfaces for upper-level applications. However, traditional techniques for global optimization of the loading and unloading process still have significant limitations, with production control decisions heavily reliant on the experience and subjective judgment of operators. This model not only has low decision-making efficiency, but also makes insufficient use of production data and lacks systematic and high-precision data support, making it difficult to achieve the expected optimization effect in the loading and unloading production process.

[0003] Therefore, we developed a data-driven optimization system for loading and unloading production processes, aiming to overcome the limitations of traditional optimization methods in the industry and provide a systematic and intelligent optimization solution. This system, starting from data mining, machine learning, and optimization algorithms, and combining big data analytics, achieves precise parameter optimization for the loading and unloading production process. Through automated feature selection and hyperparameter tuning methods, the system can continuously improve optimization results in complex production environments, meeting both single-objective and multi-objective optimization needs, and providing a standardized and implementable solution for parameter adjustment and decision-making in the loading and unloading production process. Summary of the Invention

[0004] To solve the above problems, the technical solution adopted by the present invention is: an optimization method for the loading and unloading process of dry bulk cargo ports, comprising the following steps: S1: Construct a dry bulk cargo port loading and unloading dataset, classify and summarize loading and unloading operation conditions, loading and unloading equipment parameters and loading and unloading operation indicators, dynamically aggregate hourly loading and unloading equipment parameter data into shift level, and generate a combined data table after data cleaning, missing value processing and time series resampling. S2: Preprocess the combined data table, including distribution analysis, data cleaning, recursive feature elimination and cross-validation, and automatically select the optimal feature subset; S3: Establish multiple loading and unloading production prediction models, train them based on preprocessed data, and use R², RMSE, and MAE evaluation indicators for comprehensive evaluation. Determine the best prediction model based on the evaluation results. S4: Establish multiple optimization algorithm models and determine the best optimization algorithm based on the complexity of the problem, the nonlinearity of the objective function, and the quality of the solution; S5: Taking shift unit consumption, shift loading and unloading volume, and shift working hours as prediction targets, the grid search cross-validation method is used to systematically optimize the core hyperparameters of the best prediction model, obtain the optimal hyperparameter combination corresponding to each target, and establish the optimal surrogate model for each single target. S6: Based on the optimal proxy model for each single objective, global optimization is performed within the feasible domain of process parameters to solve the optimal solution of process parameters under each single objective. S7: Construct a multi-objective optimization framework with shift unit consumption, shift loading and unloading volume, and shift working hours as the core, use the best prediction model as the surrogate model, and adopt the best optimization algorithm and constraint penalty mechanism to achieve multi-index collaborative optimization and obtain the optimal solution for collaborative optimization of shift unit consumption, shift loading and unloading volume, and shift working hours. S8: Based on the optimal solutions for each single objective and multiple objectives, output corresponding process parameter adjustment suggestions.

[0005] Furthermore, the process of preprocessing the data collected in the dry bulk cargo port loading and unloading dataset is as follows: First, through data distribution analysis, the mean, median, standard deviation, distribution type, and trend of each characteristic variable in the dry bulk cargo port loading and unloading dataset are evaluated. Secondly, data cleaning is performed. For a small number of missing values, mean imputation is usually used, while for data segments with consecutive missing values ​​exceeding a set threshold, a whole segment deletion strategy is implemented. For negative values ​​in numerical columns, they are replaced with zeros. For core process parameter columns, mean imputation of non-zero values ​​is used to ensure data integrity. For general numerical columns, global mean imputation is used to ensure data stability and integrity. For categorical columns, mode imputation is used to ensure data integrity. Finally, feature selection is performed using recursive feature elimination, which automatically selects the most useful features by progressively removing unimportant features.

[0006] Furthermore, the process of establishing and training the loading and unloading production prediction model is as follows: Based on the sample size, feature dimensions, and the characteristics of the loading and unloading production dataset—namely, the target variables being continuous process indicators such as shift unit consumption, shift loading and unloading volume, and shift man-hours—as well as the degree of nonlinear relationship and feature interaction complexity between equipment parameters and operational indicators during the loading and unloading production process, P1, P2, P3, P4…P n A loading and unloading production prediction model; The dataset is divided into a training set and a test set according to a preset ratio. Based on the feature set and target value in the training set, the loading and unloading production prediction model is fitted and trained. The loading and unloading production prediction models were cross-validated and evaluated on a test set using preset evaluation indicators. The optimal loading and unloading production prediction model was determined based on the evaluation results. Receive new unknown datasets, make predictions using the optimal loading and unloading production prediction model, and output an estimate of the target value.

[0007] Furthermore, the various optimization algorithm models mentioned employ particle swarm optimization, differential evolution, a hybrid particle swarm and differential evolution algorithm, dual simulated annealing, or evolutionary strategy algorithms.

[0008] Furthermore: When the optimization algorithm employs a double simulated annealing algorithm, the trained loading and unloading production prediction model is used as a fast evaluator to optimize the loading and unloading production process parameters; the optimization process is as follows: First, initial settings are made by randomly generating an initial solution within the feasible domain of process parameters and defining an objective function. The objective function takes the process parameter vector as input and outputs the predicted value of the target variable through a trained loading and unloading production prediction model. The predicted value is used as the fitness evaluation standard. In single-objective optimization, the objective function is the predicted output of the prediction model for any one of shift unit consumption, shift loading and unloading volume, or shift working hours. In multi-objective optimization, the objective function is a weighted combination of the predicted values ​​of shift unit consumption, shift loading and unloading volume, and shift working hours, with process constraint penalty terms added. Next, the search space is explored. In each step of the outer main annealing, the current process parameters are randomly perturbed to generate new candidate solutions to expand the search range. The perturbation amplitude is controlled by an adaptive temperature scheduling mechanism. In the initial stage, when the temperature is higher than the temperature threshold, it is allowed to jump with a step size greater than the threshold. As the iteration progresses, the temperature is gradually reduced and the step size is reduced to perform a fine search in the inner annealing. Then, an evaluation and acceptance strategy is implemented. After each new solution is generated, the current combination of process parameters is input into the trained loading and unloading production prediction model to obtain a fitness value, which serves as the optimization evaluation criterion. If the new solution is better than the current solution, it is accepted; if the new solution is worse than the current solution, the Metropolis acceptance criterion is used to avoid getting trapped in local optima with probability P. Next, a boundary condition check is performed to check whether each generated combination of process parameters is within the preset boundary range. If it exceeds the boundary, it is limited to a reasonable range. When the number of iterations reaches the preset value, or the change in the objective function value is close to convergence, the process stops. Finally, the optimal solution is output, including the best combination of process parameters found and the target value obtained through model prediction.

[0009] Furthermore: The expression for the Metropolis acceptance criterion is as follows:

[0010] Where ΔF = F(x_new) - F(x_current) is the difference between the objective function values ​​of the new solution and the current solution, and T_k is the temperature parameter of the k-th iteration.

[0011] Furthermore, using shift unit consumption, shift loading / unloading volume, and shift working hours as prediction targets, a grid search cross-validation method was employed to systematically optimize the core hyperparameters of the random forest regression model. The process is as follows: Single-objective optimization sub-step 1: Initialize the random forest regression model, fix the basic parameters, and set the evaluation index as the coefficient of determination R. 2 The model performance is evaluated using grid cross-validation. Multiple independent training subsets are generated through bootstrap sampling, and a decision tree is trained on each subset. Finally, the average of the prediction results of all decision trees is taken as the final output. Single-objective optimization sub-step 2: Hyperparameter search space definition. Define the hyperparameters to be tuned and their value ranges. The number of trees determines the number of decision trees in the ensemble, directly affecting the model's stability and computational cost. The search space is: [50, 100, 150, 200]. The feature selection strategy controls the number of candidate features randomly selected when splitting each node. It affects the ensemble effect by balancing the predictive power of a single tree and the correlation between trees. The search space is: [log2, sqrt, 0.5, 1.0]. The tree depth limit restricts the maximum growth depth of the decision tree and is a key parameter to prevent model overfitting. The search space is: [None, 5, 10, 15, 20]. Single-objective optimization sub-step 3: Grid search cross-validation execution. For the three datasets with different feature structures—shift unit consumption, shift loading and unloading volume, and shift working hours—initialize the GridSearchCV object, passing in the base model, hyperparameter search space, and grid cross-validation configuration; call the fit() method to execute the search, filter the optimal hyperparameter combination, and prioritize the combination with the highest average R2; if the average scores are the same, select the combination with the smallest standard deviation. Single-objective optimization sub-step 4: Output the optimal hyperparameter results, outputting the optimal hyperparameter configuration for each dataset; Single-objective optimization sub-step 5: Construct single-objective surrogate models based on the optimal hyperparameter configurations of each dataset. Using the best optimization algorithm, take shift unit consumption, shift loading and unloading volume, and shift working hours as optimization objectives respectively. Perform global optimization within the feasible domain of process parameters. Predict the objective values ​​corresponding to each combination of process parameters based on the single-objective surrogate models, and select the optimal solutions for process parameters under each single objective.

[0012] Furthermore, the process of reconstructing a multi-objective optimization framework centered on shift unit consumption, shift loading and unloading volume, and shift working hours, and achieving multi-indicator collaborative optimization through a trained multi-objective optimal loading and unloading production prediction model and constraint penalty mechanism, is as follows: Multi-objective optimization sub-step 1: Mathematical modeling of the multi-objective optimization problem, constructing a multi-objective comprehensive objective function to achieve a coordinated balance between minimizing shift unit consumption, maximizing shift loading and unloading volume, and minimizing shift working hours. To ensure that the optimization results meet the process feasibility requirements, three corresponding hard constraints are set, and a soft penalty mechanism is used to handle constraint violations. For any parameter combination that violates the hard constraints, a linear penalty term is applied according to the degree of violation. The final optimization objective is the sum of the original objective function and all constraint penalty terms. Multi-objective optimization sub-step 2: Use a hyperparameter-tuned random forest regression model as a surrogate model for multi-objective optimization; Multi-objective optimization sub-step 3: Perform multi-objective optimization and feature importance analysis based on the surrogate model. Execute the global optimization algorithm within the feasible domain of process parameters, traverse all possible parameter combinations, calculate the predicted values ​​of shift unit consumption, shift loading and unloading volume, shift working hours, and the final objective function value for each combination, and select the parameter combination with the smallest objective function value as the optimal solution. Based on the optimal RFR model, calculate the importance score of each process parameter to the comprehensive objective function and identify the key features affecting the optimization objective.

[0013] Furthermore, the expression for the multi-objective integrated objective function is as follows:

[0014] in, 1(x), 2(x), 3(x) represent the predicted values ​​of shift unit consumption, shift loading and unloading volume, and shift working hours, respectively; w1, w2, and w3 are weighting coefficients determined based on port production priorities; ΣP j (x) represents the sum of process constraint penalties; The expression for the constraint penalty term is as follows:

[0015] Among them, g j (x) represents the actual value of the j-th process constraint, g j ,limit is the corresponding constraint threshold, λ j This is the penalty coefficient.

[0016] Furthermore, the objective function of the single-objective optimal surrogate model is as follows:

[0017] Where x is the loading and unloading production process parameter vector, i (x) represents the predicted output of the trained loading and unloading production prediction model for the i-th objective variable, where i ∈ {shift unit consumption, shift loading and unloading volume, shift working hours}; if the objective is to maximize the shift loading and unloading volume, then the objective function is transformed into f(x) = - Shift loading and unloading volume (x).

[0018] This invention provides an optimization method for the loading and unloading production process of dry bulk cargo ports. By introducing a data-driven intelligent optimization algorithm and combining key production data in the loading and unloading process, an accurate prediction model and optimization model are established to achieve intelligent optimization of the loading and unloading production process, thereby improving the efficiency of loading and unloading operations and reducing energy consumption.

[0019] (1) Improve the level of intelligence in the production process This invention utilizes data analysis, machine learning, and optimization algorithms to incorporate the main influencing factors in loading and unloading production into the modeling scope, thereby achieving intelligent management of the production process.

[0020] Through intelligent data analysis and prediction, the system can automatically optimize operating parameters under different working conditions, significantly reducing the need for manual intervention, improving production efficiency, and the optimization suggestions are highly readable, reducing errors and deviations caused by human operation.

[0021] (2) Reduce energy consumption By optimizing loading and unloading operation parameters and production scheduling, this invention can effectively reduce energy consumption and improve energy efficiency. The system can identify and reduce unnecessary energy consumption in production, and through accurate prediction and optimization algorithms, maximize the energy efficiency of the production process without affecting the expected loading and unloading production indicators. In addition, the system's optimization process can further reduce waste, improve the environmental friendliness of the overall production process, reduce carbon emissions, and meet the requirements of green development.

[0022] Compared with the prior art, the background technology involved in this invention has the following improvements and enhancements: (1) Loading and unloading production data processing Current technical characteristics: At present, data processing in the loading and unloading production process mainly relies on threshold and combinational logic filtering methods, but its ability to identify abnormal and invalid data in complex loading and unloading production is limited, and it is difficult to adapt to the changing and complex data patterns in the loading and unloading production process.

[0023] The improvements of this invention are as follows: This invention provides a data cleaning method. First, for negative values ​​in numerical columns, they are replaced with zeros to avoid potential biases in subsequent calculations and analyses. For missing values, the methods are categorized according to the data characteristics of the columns: for core process parameter columns, non-zero value mean imputation is used to ensure data integrity; for general numerical columns, global mean imputation is used to ensure data stability and integrity; and for categorical columns, mode imputation is used to ensure data integrity. For data segments with consecutive missing values ​​exceeding a set threshold, a complete segment deletion strategy is implemented to avoid interference with the analysis results. Through these measures, this invention can effectively identify and process abnormal and invalid data in the loading and unloading production process, significantly improving data quality and providing a more reliable foundation for subsequent data analysis and modeling.

[0024] (2) Aggregation of loading and unloading production data Existing technical characteristics: The data generated during loading and unloading operations not only comes from diverse sources but also has numerous dimensions, often resulting in time alignment issues. Due to the different collection frequencies of different data sources, the timestamps between data are usually misaligned. This makes traditional data aggregation methods inflexible in processing this multi-source, multi-dimensional data, unable to adapt to dynamic changes in the data, thus affecting the efficiency and accuracy of data aggregation.

[0025] The improvement of this invention lies in providing a dynamic data aggregation method based on statistical information. This method uses an intelligent data processing system to adaptively adjust data processing strategies, thereby improving the accuracy and effectiveness of data analysis. The method statistically analyzes a large amount of collected loading and unloading production process data, evaluates the statistical characteristics of the data, and dynamically adjusts the data aggregation method and strategy based on the evaluation results. The system can select appropriate aggregation methods according to the characteristics of different data sources and dimensions, as well as the time alignment of the data, thus optimizing the data processing effect. While ensuring data accuracy, this invention effectively reduces computational overhead, improves the efficiency of data aggregation, and ultimately enhances the quality and accuracy of data processing, providing more efficient support for subsequent data analysis and optimization.

[0026] (3) Loading and unloading production prediction model Existing technical characteristics: Current methods for predicting loading and unloading production data typically rely on traditional machine learning models, such as linear regression, support vector machines, and decision trees. These methods often depend on manual intervention or experience-based selection for feature selection, potentially overlooking important features and thus affecting prediction accuracy. Furthermore, hyperparameter tuning in existing technologies often uses default settings or manual adjustments, lacking a systematic tuning process and failing to fully exploit the model's potential. Regarding model selection, typically only a single model is used, ignoring the potential of multi-model ensembles, which limits the stability and accuracy of prediction results. Moreover, due to the susceptibility to overfitting or underfitting during model training, existing methods have poor generalization ability and struggle to adapt to complex and dynamically changing loading and unloading production data.

[0027] The improvements of this invention are as follows: This invention proposes an optimization method for predicting loading and unloading production data. Through automated feature selection, hyperparameter tuning, and multi-model optimization, it significantly improves prediction accuracy. In the data preprocessing stage, the loading and unloading production data is first standardized to ensure that all features have a uniform scale, thus avoiding the impact of scale differences between features on model training. Subsequently, the Recursive Feature Elimination and Cross-Validation (RFECV) method is used for automated feature selection to determine the optimal feature subset. Next, GridSearchCV is used to tune the hyperparameters of the base model, finding the optimal hyperparameter combination through a systematic search, thereby improving the model's predictive performance. In the model training and evaluation stage, three evaluation metrics—R², RMSE, and MAE—are used to comprehensively measure the model's goodness of fit, error, and prediction bias. By evaluating multiple prediction models, the best model is automatically selected for subsequent training based on its optimal performance, ensuring that the best-performing model is used in different prediction tasks. This invention, addressing the complexity of loading and unloading production data, significantly improves the accuracy and stability of the prediction model in practical applications through the aforementioned automated optimization method. This method provides more accurate data support for decision-making in the loading and unloading production process, which can effectively assist in production optimization, improve production efficiency and reduce energy consumption.

[0028] (4) Loading and unloading production optimization model Existing technological characteristics: Existing optimization models adjust production parameters through optimization algorithms, primarily focusing on single objectives such as maximum storage capacity, minimum time, or minimum energy consumption. However, these traditional methods typically lack a systematic consideration of feature selection, failing to fully utilize predictive models to provide more accurate input to the algorithm during the optimization process. Furthermore, existing technologies often rely on manual intervention or experience-based selection for feature selection, frequently failing to dynamically adjust based on the importance of features in the dataset, resulting in some important features not being fully utilized.

[0029] The improvements of this invention are as follows: This invention employs a dual simulated annealing algorithm as the optimization algorithm in the optimization model. The simulated annealing mechanism effectively avoids getting trapped in local optima and can find the global optimum in complex production environments. Unlike traditional methods, this invention tightly integrates the prediction model with the optimization algorithm, enabling the optimization process to utilize the precise input provided by the prediction model, thus improving the accuracy and stability of the optimization. Furthermore, the dual simulated annealing algorithm, through an adaptive temperature adjustment mechanism and a probabilistic acceptance strategy, can extensively explore the search space and make fine adjustments when a better solution is found. The introduction of this algorithm not only enhances the optimization effect of the loading and unloading production process but also improves production efficiency, reduces energy consumption, and provides more accurate data support for subsequent production decisions. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart of the method. Detailed Implementation

[0032] It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Figure 1 This is a flowchart of the method.

[0035] An optimization method for the loading and unloading process of dry bulk cargo ports includes the following steps: S1: Construct a dry bulk cargo port loading and unloading dataset, classify and summarize loading and unloading operation conditions, loading and unloading equipment parameters and loading and unloading operation indicators, dynamically aggregate hourly loading and unloading equipment parameter data into shift level, and generate a combined data table after data cleaning, missing value processing and time series resampling. S2: Preprocess the combined data table, including distribution analysis, data cleaning, recursive feature elimination and cross-validation, and automatically select the optimal feature subset; S3: Establish multiple loading and unloading production prediction models, train them based on preprocessed data, and use R², RMSE, and MAE evaluation indicators for comprehensive evaluation. Determine the best prediction model based on the evaluation results. S4: Establish multiple optimization algorithm models and determine the best optimization algorithm based on the complexity of the problem, the nonlinearity of the objective function, and the quality of the solution; S5: Taking shift unit consumption, shift loading and unloading volume, and shift working hours as prediction targets, the grid search cross-validation method is used to systematically optimize the core hyperparameters of the best prediction model, obtain the optimal hyperparameter combination corresponding to each target, and establish the optimal surrogate model for each single target. S6: Based on the optimal proxy model for each single objective, global optimization is performed within the feasible domain of process parameters to solve the optimal solution of process parameters under each single objective. S7: Construct a multi-objective optimization framework with shift unit consumption, shift loading and unloading volume, and shift working hours as the core, use the best prediction model as the surrogate model, and adopt the best optimization algorithm and constraint penalty mechanism to achieve multi-index collaborative optimization and obtain the optimal solution for collaborative optimization of shift unit consumption, shift loading and unloading volume, and shift working hours. S8: Based on the optimal solutions for each single objective and multiple objectives, output corresponding process parameter adjustment suggestions.

[0036] The raw data from loading and unloading operations are categorized and summarized, and classification labels are determined for each dataset. The data is divided into three categories based on loading and unloading operation conditions, equipment parameters, and operational indicators. Loading and unloading operation conditions are at the shift level, including cargo type, ship type, tides, and weather conditions. Equipment parameters are at the hour level, including equipment runtime, equipment downtime, power consumption, fuel consumption, and dispatch instructions. Operational indicators are at the shift level, including shift unit consumption, shift loading / unloading volume, and shift man-hours. A data supplementation interface is reserved, using a standard format and unified units. Supplementing data will improve the accuracy of model training.

[0037] Since the loading and unloading equipment parameter data is in the hourly range, it needs to be aligned with other data time periods. Therefore, the hourly loading and unloading equipment parameter data is aggregated into the shift-level loading and unloading equipment parameter data.

[0038] The time column needs to be converted to datetime type, the values ​​less than zero in the numeric column need to be replaced with 0, and the core process parameter column needs to be filled with the average of non-zero values. For general numerical columns, use the global mean to fill; for categorical columns, use the mode to fill.

[0039] For data segments with consecutive missing values ​​exceeding a set threshold, a deletion strategy is implemented to remove invalid data. A function is defined to select the appropriate aggregation method based on the descriptive statistics of the columns: if the standard deviation is less than 10% of the mean, the mean is used; if the skewness is greater than 1, the median is used; otherwise, the mean is used. The data is resampled according to the frequency of the shift cycle, and the previously selected aggregation method is applied to generate new time series data. The resampled data is then merged with the loading and unloading operation condition data and loading and unloading operation indicator data by time column to obtain the final combined data table. This achieves effective integration and processing of the time series data, ensuring data integrity and consistency.

[0040] Furthermore, the process of preprocessing the data collected in the dry bulk cargo port loading and unloading dataset is as follows: S21: First, through data distribution analysis, the mean, median, standard deviation, distribution type, and trend of each characteristic variable in the dry bulk cargo port loading and unloading dataset are evaluated. S22: Secondly, data cleaning is performed. For a small number of missing values, mean imputation is usually used. For data segments with continuous missing values ​​exceeding a set threshold, a whole segment deletion strategy is implemented. For negative values ​​in numerical columns, they are replaced with zeros. For core process parameter columns, mean imputation of non-zero values ​​is used to ensure data integrity. For general numerical columns, global mean imputation is used to ensure data stability and integrity. For categorical columns, mode imputation is used to ensure data integrity. S23: Finally, feature selection is performed using the Recursive Feature Elimination (RFECV) method. RFECV is a method that combines Recursive Feature Elimination (RFE) with cross-validation. It can automatically select the most useful features by gradually removing unimportant features, thereby further optimizing the model performance.

[0041] Furthermore, the core purpose of establishing predictive models is to provide fundamental support for subsequent data optimization. Especially in single-objective and multi-objective optimization processes, the accuracy of the predictive model is crucial. By predicting the target variable, the model can provide a clear direction for the optimization algorithm, making the optimization process more efficient and accurate. In single-objective optimization, the model helps determine a specific objective function, aiming to maximize or minimize a key indicator. Multi-objective optimization is more complex; the model not only needs to balance multiple objectives but also find the optimal solution based on weights and priorities. In these optimization processes, the results of the predictive model serve as input data, providing a basis for the optimization algorithm, thereby making the optimized solution more operable and practical.

[0042] First, the loading and unloading production data is standardized: features with significant outliers are standardized using robust standardization based on median and interquartile range, and bounded features are normalized using Min-Max to ensure that all features have a uniform scale. Furthermore, the process of establishing and training the loading and unloading production prediction model is as follows: S31: Based on the sample size, feature dimensions, and the characteristics of the loading and unloading production dataset—namely, the target variables being continuous process indicators such as shift unit consumption, shift loading and unloading volume, and shift working hours—as well as the degree of nonlinear relationship and feature interaction complexity between equipment parameters and operational indicators during the loading and unloading production process, establish P1, P2, P3, P4…P n A loading and unloading production prediction model; Taking Random Forest Regression (RFR) as the core base model as an example, a bootstrap sampling strategy is used to generate multiple training subsets, and decision tree base learners are trained independently on each subset. The core model configuration includes: the number of trees (n_estimators) controls the ensemble size and prediction stability; the feature selection strategy (max_features) determines the size of the candidate feature subset when splitting nodes, reducing inter-tree correlation by introducing randomness; and the tree depth constraint (max_depth) constrains the growth complexity of a single tree to prevent overfitting. Node splitting uses the mean squared error (MSE) minimization criterion, and the final prediction result is the arithmetic mean of the outputs of all decision trees. The basic parameter configuration is: bootstrap=True, min_samples_split=2, min_samples_leaf=1, random_state=42; S32: Divide the dataset into a training set and a test set according to a preset ratio, and perform fitting training on each loading and unloading production prediction model based on the feature set and target value in the training set. During the model training phase, based on the known feature set and target value in the loading and unloading production prediction model, the model is fitted and trained on dry bulk port loading and unloading data. For the RFR model, the training process generates multiple independent training subsets through bootstrap sampling, and a decision tree is trained on each subset. When splitting nodes, the optimal split point is selected from the feature subset specified by max_features to ensure diversity among trees. The model gradually finds the optimal fitting result through multiple iterations and parameter adjustments. The core of this process is to enable the model to predict the target variable from the feature input and continuously optimize its performance.

[0043] S33: Use preset evaluation indicators to cross-validate and evaluate each loading and unloading production prediction model on the test set, and determine the optimal loading and unloading production prediction model based on the evaluation results. After model training is complete, the prediction phase begins. New, unknown datasets are received, and the loading and unloading production prediction model makes predictions, calculates the arithmetic mean, and provides an estimate of the target value. By predicting new data, the loading and unloading production prediction model provides a basis for subsequent decision-making, and through further evaluation and adjustments, the accuracy and stability of the predictions are improved.

[0044] S34: Receive new unknown datasets, make predictions using the optimal loading and unloading production prediction model, and output an estimate of the target value.

[0045] The training and testing process for the loading and unloading production prediction model mainly consists of three stages: Phase 1: 80% of the dataset is used to train the model. In this phase, the model reduces prediction error by continuously adjusting parameters to ensure that it fully learns the patterns of the input data.

[0046] Phase 2: Use the remaining 20% ​​of the data as a test set. This part of the data is not used for model training and is used to evaluate the model's generalization ability and test its performance on unseen new data.

[0047] Phase 3: Perform 5-fold cross-validation, dividing the dataset into multiple subsets for multiple training and testing sessions. This eliminates the randomness of a single partition, maximizes data utilization efficiency, evaluates model stability, and prevents overfitting.

[0048] After completing model training and testing, it is necessary to evaluate the model's performance. Evaluation not only helps us understand the model's performance, but also detects its generalization ability.

[0049] First, predictions are made on both the training and test sets to ensure that the model can not only perform well on known data, but also make reasonable predictions on unseen data.

[0050] Next, multiple evaluation metrics are used, including R², RMSE, and MAE, to comprehensively evaluate the model's performance, ensuring the model's predictive quality and measuring the accuracy and error of the predictions from different perspectives.

[0051] Finally, by comparing the evaluation results on the training and test sets, we can further determine whether the model is overfitting or underfitting. If the model performs well on the training set but poorly on the test set, it may indicate overfitting and requires further optimization.

[0052] This series of evaluation steps provides a clear direction for subsequent model optimization, ensuring that the model has good predictive and generalization capabilities in practical applications.

[0053] Based on the single dataset of shift-based resource consumption, shift-based loading and unloading volume, and shift-based working hours, as well as the multi-objective dataset of shift-based resource consumption, shift-based loading and unloading volume, and shift-based working hours, the optimization algorithm is used to optimize and train the best loading and unloading production prediction model. The process of obtaining the trained single-objective best prediction model and multi-objective best loading and unloading production prediction model is as follows: First, it is necessary to define the optimization model, select an appropriate optimization algorithm, and establish O1, O2, O3, O4…O n This model clarifies the algorithm selection principles, choosing algorithms based on problem complexity and the nonlinearity of the objective function, determining whether it is single-objective or multi-objective optimization. Available algorithms include: Particle Swarm Optimization (PSO), Differential Evolution (DEE), a hybrid PSO / DEE algorithm, dual simulated annealing, and evolutionary strategy algorithms.

[0054] Next, the fitting model step is carried out. First, the feature data and objective function are determined. Then, multiple candidate solutions are generated through the algorithm. The quality of the solution is adjusted in each iteration. Hyperparameters such as population size, mutation rate and number of iterations also need to be fine-tuned according to the objective function and constraints.

[0055] Finally, the optimized model is used to predict the objective function value, and the model output is evaluated to determine whether it meets the constraints and objective value. The best optimization algorithm is selected based on the quality of the solution. This process ensures the effectiveness and reliability of the model and lays the foundation for subsequent applications.

[0056] Taking the Dual Simulated Annealing algorithm as an example, this algorithm adopts a master-slave two-layer annealing structure: the outer master annealing uses a weighted average of multiple objectives such as production energy consumption and operation time as the fitness function to perform a broad search in the global process parameter space; the inner slave annealing uses the output of the prediction model as a constraint condition to perform local fine-tuning of the loading and unloading sequence timing parameters. This invention uses the prediction model as a fast evaluator for the optimization process and establishes a prediction-optimization closed loop: the current solution is input into the prediction model to obtain the fitness value, and the solution is updated using a probabilistic acceptance strategy through the Metropolis criterion, effectively avoiding getting trapped in local optima, thereby finding the global optimum.

[0057] Specific workflow: S51, Initial settings are performed, and an initial solution is randomly generated within the feasible domain of the process parameters. An objective function is defined. The objective function takes the process parameter vector as input and outputs the predicted value of the target variable through the trained loading and unloading production prediction model. The predicted value is used as the fitness evaluation standard. In single-objective optimization, the objective function is the predicted output of the prediction model for any one of shift unit consumption, shift loading and unloading volume, or shift working hours. In multi-objective optimization, the objective function is a weighted combination of the predicted values ​​of shift unit consumption, shift loading and unloading volume, and shift working hours, and a process constraint penalty term is superimposed. S52 explores the search space. The outer main annealing randomly perturbs the current parameters at each step to generate new candidate solutions and expand the search range. An adaptive temperature scheduling mechanism controls the perturbation amplitude. In the initial stage, when the temperature is greater than the temperature threshold, a large jump is allowed to avoid being confined to local optima. As the iteration progresses, the temperature gradually decreases and the step size shrinks, with the inner annealing performing a fine search. S53, Evaluation and Acceptance Strategy: After each new solution is generated, the current solution is input into the prediction model to obtain a fitness value, which is used as the optimization evaluation criterion. If the new solution is better than the current solution, the new solution is accepted; if the new solution is worse than the current solution, the Metropolis acceptance criterion is applied, with a probability P = exp(-Δf / Tk), where ΔF = F(xnew). F(xcurrent) is the difference between the objective function value of the new solution and the current solution, and Tk is the temperature parameter in the k-th iteration. To avoid getting trapped in local optima, a boundary condition check is then performed to check whether each generated new parameter combination is within a preset boundary range. If it exceeds the boundary, the "np.clip" function is used to limit it to a reasonable range. Then, a stopping condition is set. After the algorithm runs for a period of time, it stops running when the number of iterations reaches a preset value or the change in the objective function value is close to convergence. S54 outputs the optimal solution, including the best combination of parameters found and the target value obtained through model prediction.

[0058] The expression for the Metropolis acceptance criterion is as follows:

[0059] Where ΔF = F(x_new) - F(x_current) is the difference between the objective function values ​​of the new solution and the current solution, and T_k is the temperature parameter of the k-th iteration.

[0060] Single-objective optimization (501) selects one of the following as the optimization objective: shift unit consumption, shift loading and unloading volume, or shift working hours, with the aim of improving the generalization ability of the Random Forest Regression (RFR) model. The Grid Search Cross Validation (GridSearchCV) method is used to systematically tune three core hyperparameters: the number of trees, the feature selection strategy, and the tree depth limit. The implementation process consists of four sub-steps: Single-objective optimization sub-step 1: Initialize the Random Forest Regression (RFR) model, fix the basic parameters, set the evaluation metric as the coefficient of determination R2, use 5-fold cross-validation to evaluate the model performance, generate multiple independent training subsets through bootstrap sampling, train a decision tree on each subset, and finally take the average of the prediction results of all decision trees as the final output; the basic configuration is: bootstrap=True, min_samples_split=2, min_samples_leaf=1, random_state=42.

[0061] Single-objective optimization sub-step 2: Hyperparameter search space definition. Define the hyperparameters to be tuned and their value range. The number of trees determines the number of decision trees in the ensemble (n_estimators), which directly affects the stability of the model and the computational cost. The search space is: [50, 100, 150, 200]. The feature selection strategy (max_features) controls the number of candidate features randomly selected when splitting each node. It affects the ensemble effect by balancing the predictive power of a single tree and the correlation between trees. The search space is: [log2, sqrt, 0.5, 1.0]. The tree depth limit (max_depth) limits the maximum growth depth of the decision tree. It is a key parameter to prevent the model from overfitting. The search space is: [None, 5, 10, 15, 20]. Single-objective optimization sub-step 3: Grid search cross-validation execution. For the three datasets with different feature structures—shift unit consumption, shift loading and unloading volume, and shift working hours—initialize the GridSearchCV object, passing in the base model, hyperparameter search space, and grid cross-validation configuration; call the fit() method to execute the search, filter the optimal hyperparameter combination, and prioritize the combination with the highest average R2; if the average scores are the same, select the combination with the smallest standard deviation. Single-objective optimization sub-step 4: Outputting the optimal hyperparameter results. The optimal number of trees is related to the size and noise level of the dataset. The effectiveness of the feature selection strategy depends on the feature correlation structure of the dataset. The tree depth limit balances fitting ability and generalization ability by controlling the complexity of the model. Based on the step-by-step analysis of the single hyperparameter influence mechanism, the optimal hyperparameter configuration for each dataset is output. For shift unit consumption data, the optimal hyperparameter configuration is: max_depth=None, max_features=log2, n_estimators=50; for shift loading and unloading volume data, the optimal hyperparameter configuration is: max_depth=None, max_features=sqrt, n_estimators=100; for shift working hours data, the optimal hyperparameter configuration is: max_depth=10, max_feature=sqrt, n_estimators=200.

[0062] Furthermore, based on single-objective optimization, a multi-objective optimization framework is constructed with shift unit consumption, shift loading and unloading volume, and shift working hours as the core. This framework achieves multi-indicator collaborative optimization through surrogate model prediction and constraint penalty mechanisms. The implementation process consists of three sub-steps: Multi-objective optimization sub-step 1: Mathematical modeling of the multi-objective optimization problem, constructing a comprehensive objective function to achieve a coordinated balance between minimizing shift unit consumption, maximizing shift loading and unloading volume, and minimizing shift working hours. To ensure that the optimization results meet the process feasibility requirements, three corresponding hard constraints are set, and a soft penalty mechanism is used to handle constraint violations. For any parameter combination that violates the hard constraints, a linear penalty term is applied according to the degree of violation. The final optimization objective is the sum of the original objective function and all constraint penalty terms. The base model is an initialized but not yet hyperparameter-tuned instance of a random forest regression model; that is, an initial model using default or preset base hyperparameters. In grid search cross-validation, this base model serves as the base learner to be tuned, and the optimal hyperparameter combination is obtained by traversing its hyperparameter search space.

[0063] The comprehensive objective function is constructed through a weighted combination method, unifying the three sub-objectives of shift unit consumption, shift loading and unloading volume, and shift working hours into a single optimization index.

[0064] The hard constraints are set according to the feasibility requirements of the loading and unloading process of dry bulk cargo ports, including but not limited to: (1) equipment running time constraints, that is, the continuous running time of the equipment does not exceed the safety limit; (2) energy consumption constraints, that is, the total power consumption / fuel consumption per shift does not exceed the rated threshold; (3) loading and unloading capacity constraints, that is, the loading and unloading volume per shift is not lower than the minimum operating capacity requirement of the port. The above constraints ensure the feasibility and safety of the optimization results in actual production.

[0065] The soft penalty mechanism refers to the process of optimizing a combination of process parameters that violates hard constraints. Instead of directly removing it from the search space, a penalty term is added to its objective function value to reduce the fitness score of the combination. This allows the optimization algorithm to continue exploring near the constraint boundary and avoids missing potential optimal regions by prematurely eliminating candidate solutions.

[0066] The original objective function refers to the comprehensive objective function before the addition of constraint penalty terms, that is, the objective function part after the weighted combination of shift unit consumption, shift loading and unloading volume, and shift working hours.

[0067] All constraint penalty terms refer to the sum of linear penalty terms corresponding to each hard constraint condition.

[0068] Multi-objective optimization sub-step 2: Construction and validation of multi-objective optimization surrogate model. A hyperparameter-tuned random forest regression (RFR) model is used as a surrogate model for multi-objective optimization to replace real physical experiments for rapid performance evaluation. Multi-objective optimization sub-step 3: Perform multi-objective optimization and feature importance analysis based on the surrogate model. Execute the global optimization algorithm within the feasible domain of process parameters, traverse all possible parameter combinations, calculate the predicted values ​​of shift unit consumption, shift loading and unloading volume, shift working hours, and the final objective function value for each combination, and select the parameter combination with the smallest objective function value as the optimal solution. Based on the optimal RFR model, calculate the importance score of each process parameter to the comprehensive objective function and identify the key features affecting the optimization objective.

[0069] The optimal optimization algorithm determined through comparative evaluation includes, but is not limited to, particle swarm optimization, differential evolution, a hybrid particle swarm optimization and differential evolution algorithm, dual simulated annealing, or evolutionary strategy algorithms. The specific algorithm is determined by evaluating and selecting from the above candidate algorithms based on the complexity of the loading and unloading production optimization problem, the nonlinearity of the objective function, and the single / multi-objective nature of the problem.

[0070] In multi-objective optimization, there is no independent "multi-objective random forest regression model." Multi-objective optimization uses a random forest regression model, tuned for hyperparameters, as a surrogate model. This model is used to quickly predict the shift unit consumption, shift loading / unloading volume, and shift working hours corresponding to various combinations of process parameters, thus replacing real physical experiments for fitness evaluation. This surrogate model has already been trained and tuned in the single-objective optimization stage; in the multi-objective optimization stage, its predictive capabilities are directly invoked, eliminating the need for retraining or building a new predictive model.

[0071] The process of outputting the loading and unloading optimization results is as follows: For single-objective optimization, based on the optimal hyperparameter combinations corresponding to each dataset obtained in "Single-objective Optimization Sub-step 4", three single-objective random forest regression surrogate models were constructed for shift unit consumption, shift loading / unloading volume, and shift working hours, respectively. All models were trained on the complete training set of their respective datasets, maintaining consistency between basic parameters and hyperparameters during training, and using a fixed random seed to ensure reproducibility. Using the three trained single-objective RFR models as predictors, a global optimization algorithm was executed within the feasible domain of process parameters to solve for the single-objective optimal solutions for each of the three objectives. During the optimization process, physical boundary constraints and operational safety constraints of the process parameters were satisfied, and the optimal combination of process parameters and the corresponding theoretical optimal index values ​​were output for each single objective. Based on the single-objective global optimal solutions and key feature analysis results, specific process adjustment suggestions for each of the three objectives were output, clarifying the adjustment priority of process parameters such as equipment operation and scheduling instructions, prioritizing the adjustment of key features with the greatest impact on the objective, and providing specific adjustment directions and recommended value ranges for key loading / unloading parameters.

[0072] For multi-objective optimization (502), based on the optimal parameter combination obtained in "multi-objective optimization sub-step 3", the corresponding core process index values ​​are predicted by the RFR model. Under the optimal parameter combination, all indicators meet the process constraints, and the comprehensive objective function value is better than the optimization result of the original data model. Based on the feature importance analysis results and the optimal parameter combination, the recommended adjustment values ​​of equipment operation and scheduling instructions under the current cargo type, ship type, tide, and weather conditions are proposed to achieve the coordinated optimization of shift unit consumption, shift loading and unloading volume, and shift working hours.

[0073] The expression for the multi-objective integrated objective function is as follows:

[0074] in, 1(x), 2(x), 3(x) represent the predicted values ​​of shift unit consumption, shift loading and unloading volume, and shift working hours, respectively; w1, w2, and w3 are weighting coefficients determined based on port production priorities; ΣP j (x) represents the sum of process constraint penalties; The expression for the constraint penalty term is as follows:

[0075] Among them, g j (x) represents the actual value of the j-th process constraint, g j ,limit is the corresponding constraint threshold, λ j This is the penalty coefficient.

[0076] Furthermore, the objective function of the single-objective optimal surrogate model is as follows:

[0077] Where x is the loading and unloading production process parameter vector, i (x) represents the predicted output of the trained loading and unloading production prediction model for the i-th objective variable, where i ∈ {shift unit consumption, shift loading and unloading volume, shift working hours}; if the objective is to maximize the shift loading and unloading volume, then the objective function is transformed into f(x) = - Shift loading and unloading volume (x).

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An optimization method for the loading and unloading process of dry bulk cargo ports, characterized in that: Includes the following steps: S1: Construct a dry bulk cargo port loading and unloading dataset, classify and summarize loading and unloading operation conditions, loading and unloading equipment parameters and loading and unloading operation indicators, dynamically aggregate hourly loading and unloading equipment parameter data into shift level, and generate a combined data table after data cleaning, missing value processing and time series resampling. S2: Preprocess the combined data table, including distribution analysis, data cleaning, recursive feature elimination and cross-validation, and automatically select the optimal feature subset; S3: Establish multiple loading and unloading production prediction models, train them based on preprocessed data, and use R², RMSE, and MAE evaluation indicators for comprehensive evaluation. Determine the best prediction model based on the evaluation results. S4: Establish multiple optimization algorithm models and determine the best optimization algorithm based on the complexity of the problem, the nonlinearity of the objective function, and the quality of the solution; S5: Taking shift unit consumption, shift loading and unloading volume, and shift working hours as prediction targets, the grid search cross-validation method is used to systematically optimize the core hyperparameters of the best prediction model, obtain the optimal hyperparameter combination corresponding to each target, and establish the optimal surrogate model for each single target. S6: Based on the optimal proxy model for each single objective, global optimization is performed within the feasible domain of process parameters to solve the optimal solution of process parameters under each single objective. S7: Construct a multi-objective optimization framework with shift unit consumption, shift loading and unloading volume, and shift working hours as the core, use the best prediction model as the surrogate model, and adopt the best optimization algorithm and constraint penalty mechanism to achieve multi-index collaborative optimization and obtain the optimal solution for collaborative optimization of shift unit consumption, shift loading and unloading volume, and shift working hours. S8: Based on the optimal solutions for each single objective and multiple objectives, output corresponding process parameter adjustment suggestions.

2. The optimization method for the loading and unloading process of dry bulk cargo ports according to claim 1, characterized in that: The process of preprocessing the data in the dry bulk cargo port loading and unloading dataset is as follows: First, through data distribution analysis, the mean, median, standard deviation, distribution type, and trend of each characteristic variable in the dry bulk cargo port loading and unloading dataset are evaluated. Secondly, data cleaning is performed. For a small number of missing values, mean imputation is usually used, while for data segments with consecutive missing values ​​exceeding a set threshold, a whole segment deletion strategy is implemented. For negative values ​​in numerical columns, they are replaced with zeros. For core process parameter columns, mean imputation of non-zero values ​​is used to ensure data integrity. For general numerical columns, global mean imputation is used to ensure data stability and integrity. For categorical columns, mode imputation is used to ensure data integrity. Finally, feature selection is performed using recursive feature elimination, which automatically selects the most useful features by progressively removing unimportant features.

3. The optimization method for the loading and unloading process of dry bulk cargo ports according to claim 1, characterized in that: The process of establishing and training the loading and unloading production prediction model is as follows: Based on the sample size, feature dimensions, and the characteristics of the loading and unloading production dataset—namely, the target variables being continuous process indicators such as shift unit consumption, shift loading and unloading volume, and shift man-hours—as well as the degree of nonlinear relationship and feature interaction complexity between equipment parameters and operational indicators during the loading and unloading production process, P1, P2, P3, P4…P n A loading and unloading production prediction model; The dataset is divided into a training set and a test set according to a preset ratio. Based on the feature set and target value in the training set, the loading and unloading production prediction model is fitted and trained. The loading and unloading production prediction models were cross-validated and evaluated on a test set using preset evaluation indicators. The optimal loading and unloading production prediction model was determined based on the evaluation results. Receive new unknown datasets, make predictions using the optimal loading and unloading production prediction model, and output an estimate of the target value.

4. The optimization method for the loading and unloading process of dry bulk cargo ports according to claim 1, characterized in that: The various optimization algorithm models employ particle swarm optimization, differential evolution, a hybrid particle swarm and differential evolution algorithm, dual simulated annealing, or evolutionary strategy algorithms.

5. The optimization method for the loading and unloading process of dry bulk cargo ports according to claim 4, characterized in that: When the optimization algorithm employs the dual simulated annealing algorithm, the trained loading and unloading production prediction model is used as a fast evaluator to optimize the loading and unloading production process parameters; the optimization process is as follows: First, initial settings are made by randomly generating an initial solution within the feasible domain of process parameters and defining an objective function. The objective function takes the process parameter vector as input and outputs the predicted value of the target variable through a trained loading and unloading production prediction model. The predicted value is used as the fitness evaluation standard. In single-objective optimization, the objective function is the predicted output of the prediction model for any one of shift unit consumption, shift loading and unloading volume, or shift working hours. In multi-objective optimization, the objective function is a weighted combination of the predicted values ​​of shift unit consumption, shift loading and unloading volume, and shift working hours, with process constraint penalty terms added. Next, the search space is explored. In each step of the outer main annealing, the current process parameters are randomly perturbed to generate new candidate solutions to expand the search range. The perturbation amplitude is controlled by an adaptive temperature scheduling mechanism. In the initial stage, when the temperature is higher than the temperature threshold, it is allowed to jump with a step size greater than the threshold. As the iteration progresses, the temperature is gradually reduced and the step size is reduced to perform a fine search in the inner annealing. Then, an evaluation and acceptance strategy is implemented. After each new solution is generated, the current combination of process parameters is input into the trained loading and unloading production prediction model to obtain a fitness value, which serves as the optimization evaluation criterion. If the new solution is better than the current solution, it is accepted; if the new solution is worse than the current solution, the Metropolis acceptance criterion is used to avoid getting trapped in local optima with probability P. Next, a boundary condition check is performed to check whether each generated combination of process parameters is within the preset boundary range. If it exceeds the boundary, it is limited to a reasonable range. When the number of iterations reaches the preset value, or the change in the objective function value is close to convergence, the process stops. Finally, the optimal solution is output, including the best combination of process parameters found and the target value obtained through model prediction.

6. The optimization method for the loading and unloading process of dry bulk cargo ports according to claim 5, characterized in that: The expression for the Metropolis acceptance criterion is as follows: Where ΔF = F(x_new) - F(x_current) is the difference between the objective function values ​​of the new solution and the current solution, and T_k is the temperature parameter of the k-th iteration.

7. The optimization method for the loading and unloading process of dry bulk cargo ports according to claim 3, characterized in that: Using shift unit consumption, shift loading and unloading volume, and shift working hours as prediction targets, respectively, the core hyperparameters of the random forest regression model were systematically tuned using a grid search cross-validation method. The process is as follows: Single-objective optimization sub-step 1: Initialize the random forest regression model, fix the basic parameters, and set the evaluation index as the coefficient of determination R. 2 The model performance is evaluated using grid cross-validation. Multiple independent training subsets are generated through bootstrap sampling, and a decision tree is trained on each subset. Finally, the average of the prediction results of all decision trees is taken as the final output. Single-objective optimization sub-step 2: Hyperparameter search space definition. Define the hyperparameters to be tuned and their value ranges. The number of trees determines the number of decision trees in the ensemble, directly affecting the model's stability and computational cost. The search space is: [50, 100, 150, 200]. The feature selection strategy controls the number of candidate features randomly selected when splitting each node. It affects the ensemble effect by balancing the predictive power of a single tree and the correlation between trees. The search space is: [log2, sqrt, 0.5, 1.0]. The tree depth limit restricts the maximum growth depth of the decision tree and is a key parameter to prevent model overfitting. The search space is: [None, 5, 10, 15, 20]. Single-objective optimization sub-step 3: Grid search cross-validation execution. For the three datasets with different feature structures—shift unit consumption, shift loading and unloading volume, and shift working hours—initialize the GridSearchCV object, passing in the base model, hyperparameter search space, and grid cross-validation configuration; call the fit() method to execute the search, filter the optimal hyperparameter combination, and prioritize the combination with the highest average R2; if the average scores are the same, select the combination with the smallest standard deviation. Single-objective optimization sub-step 4: Output the optimal hyperparameter results, outputting the optimal hyperparameter configuration for each dataset; Single-objective optimization sub-step 5: Construct single-objective surrogate models based on the optimal hyperparameter configurations of each dataset. Using the best optimization algorithm, take shift unit consumption, shift loading and unloading volume, and shift working hours as optimization objectives respectively. Perform global optimization within the feasible domain of process parameters. Predict the objective values ​​corresponding to each combination of process parameters based on the single-objective surrogate models, and select the optimal solutions for process parameters under each single objective.

8. The optimization method for the loading and unloading process of dry bulk cargo ports according to claim 7, characterized in that: The process of reconstructing a multi-objective optimization framework centered on shift unit consumption, shift loading and unloading volume, and shift working hours, and achieving multi-indicator collaborative optimization through a trained multi-objective optimal loading and unloading production prediction model and constraint penalty mechanism, is as follows: Multi-objective optimization sub-step 1: Mathematical modeling of the multi-objective optimization problem, constructing a multi-objective comprehensive objective function to achieve a coordinated balance between minimizing shift unit consumption, maximizing shift loading and unloading volume, and minimizing shift working hours. To ensure that the optimization results meet the process feasibility requirements, three corresponding hard constraints are set, and a soft penalty mechanism is used to handle constraint violations. For any parameter combination that violates the hard constraints, a linear penalty term is applied according to the degree of violation. The final optimization objective is the sum of the original objective function and all constraint penalty terms. Multi-objective optimization sub-step 2: Use a hyperparameter-tuned random forest regression model as a surrogate model for multi-objective optimization; Multi-objective optimization sub-step 3: Perform multi-objective optimization and feature importance analysis based on the surrogate model. Execute the global optimization algorithm within the feasible domain of process parameters, traverse all possible parameter combinations, calculate the predicted values ​​of shift unit consumption, shift loading and unloading volume, shift working hours, and the final objective function value for each combination, and select the parameter combination with the smallest objective function value as the optimal solution. Based on the optimal RFR model, calculate the importance score of each process parameter to the comprehensive objective function and identify the key features affecting the optimization objective.

9. The optimization method for the loading and unloading process of dry bulk cargo ports according to claim 8, characterized in that: The expression for the multi-objective integrated objective function is as follows: in, 1(x), 2(x), 3(x) represent the predicted values ​​of shift unit consumption, shift loading and unloading volume, and shift working hours, respectively; w1, w2, and w3 are weighting coefficients determined based on port production priorities; ΣP j (x) represents the sum of process constraint penalties; The expression for the constraint penalty term is as follows: Among them, g j (x) represents the actual value of the j-th process constraint, g j ,limit is the corresponding constraint threshold, λ j This is the penalty coefficient.

10. The optimization method for the loading and unloading process of dry bulk cargo ports according to claim 7, characterized in that: The objective function of the single-objective optimal surrogate model is as follows: Where x is the loading and unloading production process parameter vector, i (x) represents the predicted output of the trained loading and unloading production prediction model for the i-th objective variable, where i ∈ {shift unit consumption, shift loading and unloading volume, shift working hours}; if the objective is to maximize the shift loading and unloading volume, then the objective function is transformed into f(x) = - Loading and unloading volume per shift (x).

Images