Multi-source data fused ship fuel consumption prediction method

Abstract

The invention provides a ship fuel consumption prediction method fusing multi-source data, and belongs to the technical field of ship fuel consumption. The method comprises the steps of collecting relevant features of ship fuel consumption and inputting the relevant features into a plurality of machine learning models; each machine learning model outputs a respective oil consumption prediction result; fusing the prediction results of the machine learning models based on a preset weight to obtain a final ship fuel consumption prediction result; historical data related to ship fuel consumption are collected to construct a training set; and training the plurality of machine learning models based on a training set, and determining the weight of each machine learning model based on the prediction result of the training set as the preset weight. According to the method, the oil consumption is predicted simultaneously through a plurality of machine learning models, and the prediction results are fused based on the weight of each model to obtain the final oil consumption prediction result, so that the prediction accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ship fuel consumption, in particular to a ship fuel consumption prediction method fusing multi-source data. BACKGROUND

[0002] At present, the global shipping industry is facing unprecedented economic and environmental pressures, and ship fuel consumption, as a key component of total operating costs, has become a core link that shipping companies urgently need to optimize. According to statistics, international shipping fuel consumption accounts for about 13% of global transportation energy consumption, and its carbon dioxide emissions account for about 10% of global transportation total emissions, and it is also one of the main sources of global sulfur oxide emissions. In order to achieve low-carbon transformation of the shipping industry, the International Maritime Organization has introduced mandatory technical standards such as ship energy efficiency index, carbon intensity index, and energy efficiency operational index.

[0003] In recent years, with the rapid development of artificial intelligence technology, machine learning methods have provided a new technical path for ship fuel consumption prediction due to their non-linear modeling capabilities. Existing technologies fuse ship noon reports, automatic identification systems, ECMWF reanalysis weather data (ERA5), and other multi-source information to build a high-temporal and spatial resolution feature space database, which to some extent improves the model prediction accuracy and generalization ability. However, existing methods still face the following challenges in key technical aspects: first, a single machine learning model cannot balance global prediction accuracy and local working condition stability, and the model generalization ability is limited; second, multi-source heterogeneous data have inconsistencies in temporal and spatial scales, structure, and form, and feature redundancy and collinearity problems are easily caused during fusion, affecting model performance; third, there is a lack of feature importance evaluation mechanism from multiple model perspectives, and existing feature selection methods are easily affected by bias, resulting in low modeling efficiency and unstable results; fourth, most high-precision machine learning models are "black box" structures, making it difficult to provide explainability analysis, limiting their application value and trust basis in actual shipping management.

[0004] Therefore, there is an urgent need for a ship fuel consumption prediction method fusing multi-source data. SUMMARY

[0005] Therefore, the present application provides a ship fuel consumption prediction method fusing multi-source data, which integrates multi-source data fusion, weighted feature selection, hyperparameter adaptive optimization, and multi-model integrated prediction into a high-precision, strongly interpretable ship fuel consumption modeling method, providing a decision support tool for compliance supervision, cost control, and energy efficiency management for shipping companies.

[0006] To this end, the present application provides the following technical solutions: A ship fuel consumption prediction method fusing multi-source data, comprising: Collecting historical data related to ship fuel consumption to construct a training set; Training the plurality of machine learning models based on the training set, determining the weight of each machine learning model based on the prediction result of the training set as a preset weight; Collecting features related to ship fuel consumption to input into the plurality of machine learning models; Each machine learning model outputs its own fuel consumption prediction result; based on the preset weight, the prediction results of each machine learning model are fused to obtain the final ship fuel consumption prediction result.

[0007] Further, it also includes: Based on the training result, calculate the SHAP value of each machine learning model as the initial feature importance measure; According to the initial feature importance measure, arrange all features in descending order to obtain the initial feature importance sorting list of each model; According to the feature importance sorting list, sequentially construct a subset containing the first most important features , from 1 to , representing the total number of features; representing the index of the model; Calculate the weighted voting score of all optimal subsets of features based on the weight of weighted voting: For each feature subset , each model is trained on the corresponding training data subset, and the prediction performance of each model is evaluated through 5-fold cross-validation to find the number of features that make the model determination coefficient score reach the peak , and according to the determined number of features, the best feature subset is constructed, and the formula is represented as:

[0008]

[0009] Among them, is the best feature subset of model m, represents the first feature of the model; when , the determination coefficient score of the model is used as the weight of weighted voting, that is ; represents the determination coefficient obtained by training only the first n ranked features of model m and evaluating on the validation set; represents that the determination coefficient reaches the maximum value; Calculate the weighted voting score of all optimal subsets of features based on the weight of weighted voting, and the formula is represented as:

[0010] wherein, is a weighted voting score, is a performance weight of the model, is a set of models; features with all weighted voting scores greater than 1 are taken as the relevant features of the ship fuel consumption.

[0011] Further, the prediction result based on the training set determines the weight of each machine learning model, comprising: based on the feature set, hyperparameter optimization is performed on multiple machine learning models through grid search combined with random search, and prediction results of each machine learning model are obtained; based on the prediction results corresponding to the samples, result features are extracted; based on the result features, the samples are clustered into a preset number of data clusters through K-means, and the model fusion weight is solved with the minimum error between the fused prediction value and the true value as the target.

[0012] Further, the relevant features of the ship fuel consumption, comprising: relative surge direction angle, ship heading, average wave period, average speed throughout the journey, atmospheric pressure, remaining distance to the destination port, main engine speed, 2-meter sea surface temperature, cumulative sailing distance, total sailing time, daily sailing distance and daily sailing time.

[0013] Further, the machine learning model, comprising: SVR, Random Forest, Gradient Boosting, XGBoost, LightGBM and CatBoost.

[0014] Further, the result features, comprising: mean, standard deviation, extreme value, median, 25th and 75th percentile of each model prediction.

[0015] Further, the historical data related to the ship fuel consumption, comprising: ship noon report, AIS data and ERA5 data.

[0016] Further, the hyperparameter optimization on multiple machine learning models based on the feature set through grid search combined with random search, and the prediction results of each machine learning model, comprising: set an initial hyperparameter search space ; randomly sample a group of hyperparameter combinations for each group of parameters , the model performance is evaluated on the training set and validation set by cross-validation or hold-out method, and the evaluation error is obtained ; , The formula for determining the parameter combination with the best performance in the current stage is:

[0017]

[0018]

[0019] wherein, is the number of random sampling times, represents the dimension of hyperparameters, and are the lower bound and upper bound of the i-th hyperparameter, respectively, represents the i-th combination of hyperparameters; represents the independent variable set of the training set; represents the dependent variable set of the training set; represents the feature matrix and label vector of the validation set; represents the feature matrix and label vector of the test set; is the optimal solution of the initial random search; the optimal solution of the initial stage is subjected to local search: taking as the center, the search space is reduced according to the scaling factor , and the local parameter space is constructed; each parameter dimension is divided into equidistant points to form a grid hyperparameter set ; for each combination of hyperparameters , the model performance is evaluated again and the hyperparameter combination with the best comprehensive performance in the global range is determined.

[0020] Further, the result-based features are clustered into a preset number of data clusters by K-means, and the model fusion weight is solved with the minimum error between the fused prediction value and the true value as the target, including: For each sample , its cluster is represented by ; for sample , define the feature vector , which includes statistical features and correlation features , and the formula is:

[0021] ​

[0022] where, is the mean value predicted by each model, is the standard deviation, is the range, is the median, and are the 25th and 75th percentiles, respectively; is the model and model predicting the correlation coefficient for the group the sample belongs to; use K-means algorithm to cluster all samples into clusters, the number of clusters is automatically determined by silhouette coefficient, and clusters with sample number less than a threshold are merged into the nearest large cluster:

[0023]

[0024]

[0025] where, is the center of the th cluster, is the average distance of sample to other samples in the same cluster, is the average distance of sample to the nearest sample not in the same cluster, is the center of cluster ; the largest value of is selected as the optimal number of clusters; is the silhouette coefficient when the number of clusters is K; is the Euclidean distance between the center of cluster and cluster , where c' and c are any two clusters; calculate the optimal fusion weight vector , which is expressed as:

[0026]

[0027] where, is the weight vector to be solved in the optimization process, is the true value of the th sample, is the predicted value of the th model for the th sample, is the weights of the models; the constraints ensure that the calculated weights An effective proportion distribution is formed.

[0028] Further, the method comprises the following steps: for each sample A new and more informative meta-feature vector is constructed as the input of the machine learning model to generate the final prediction output, which is expressed as:

[0029]

[0030]

[0031] wherein, is the meta-feature vector of the i-th sample, which is the input of the meta-model is the base fusion prediction value of the i-th sample; it is a weighted average prediction result calculated according to the optimal weights of the cluster to which the sample belongs; is the standard deviation, and represent the maximum and minimum values of the prediction value respectively, is the final prediction value of the i-th sample. Advantages and positive effects of the present application: The method determines the relevant features of the ship fuel consumption prediction through the weighted voting method, thereby improving the accuracy of the detection. The method simultaneously predicts the fuel consumption through multiple machine learning models, and fuses the prediction results based on the weights of the models to obtain the final fuel consumption prediction result, thereby improving the prediction accuracy.

[0032] The data fusion method of the method not only fully retains the key information and unique advantages of each heterogeneous data source, but also effectively makes up for the inherent limitations of single data source through complementary integration, thereby providing high-quality and multi-dimensional data support for subsequent ship fuel consumption prediction model construction. BRIEF DESCRIPTION OF DRAWINGS

[0033]

[0034] The data fusion method of the method not only fully retains the key information and unique advantages of each heterogeneous data source, but also effectively makes up for the inherent limitations of single data source through complementary integration, thereby providing high-quality and multi-dimensional data support for subsequent ship fuel consumption prediction model construction. BRIEF DESCRIPTION OF DRAWINGS

[0035] ​​​In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without any creative labor.

[0036] Figure 1 is the feature ranking and selection based on SWFS weighted voting score; Figure 2 is the hyperparameter optimization (HPO) performance and computational cost comparison; Figure 3 is the consistency distribution of ACME model predicted value and actual value; Figure 4 is the distribution of samples in each cluster; Figure 5 is the average fusion weight of each cluster; Figure 6 is the SHAP feature importance analysis of ship fuel consumption prediction; Figure 7 is the SHAP feature importance bee swarm scatter plot; Figure 8 is the flowchart of the ship fuel consumption prediction method based on fusion of multiple source data. DETAILED DESCRIPTION

[0037] In order to make the person skilled in the art better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without any creative labor should be within the scope of protection of the present application.

[0038] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily limit to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0039] The application provides a ship fuel consumption prediction method for fusing multi-source data, comprising: S1, taking the noon report of a bulk carrier as the core data source, supplemented by AIS data, ERA5 data and GOPAF data, high-precision fusion of multi-source heterogeneous data is realized by using time-space alignment and interpolation resampling technology.

[0040] S2, based on SHAP-based Weighted Feature Selection (SWFS), through multi-model SHAP value evaluation, recursive feature evaluation and weighted voting mechanism, redundant features are effectively identified and removed, and the prediction performance of the input dimension is significantly improved.

[0041] S3, based on the optimized feature subset, a Hierarchical Adaptive Parameter Space Exploration (HAPSE) method is used to optimize the hyperparameters of multiple advanced machine learning models (Hyperparameter Optimization, HPO); and an Adaptive Cluster-based Multi-Ensemble (ACME) model is constructed; a regionalized weighted strategy is used to fuse multiple supervised learning algorithms, which fully utilizes the advantages of different models and adapts to the internal heterogeneity of the data, and realizes high-precision regression prediction of fuel consumption.

[0042] S4, the SHAP method is used to carry out global and local interpretability analysis, and the contribution and interaction of each feature to the prediction result are quantified, which provides a solid physical and operational mechanism support for the model decision-making process.

[0043] Embodiment In combination Figure 8 As shown in the figure, a ship fuel consumption modeling method based on multi-source data fusion, weighted feature selection, hyperparameter adaptive optimization and adaptive ensemble learning, the method comprises the following steps: S1, collecting multi-source data of the ship and performing data preprocessing to obtain preprocessed multi-source data.

[0044] 1) Ship noon report as a standardized document recording the daily operational status of the ship in the shipping industry, usually completed by the captain or the officer on duty at local noon and submitted to the shore management department, its content includes the ship's navigation status and environmental meteorological parameters. At the same time, AIS data is used as a supplement to ship noon report data. Through high-frequency acquisition of ship position information such as latitude and longitude, AIS data provides strong support for the spatio-temporal matching of ship noon report and external meteorological data, thereby ensuring the accuracy and integrity of data fusion.

[0045] To overcome the low frequency of meteorological data collection in the ship noon report and the possible errors caused by manual recording, ERA5 dataset is introduced. ERA5 is provided by the European Centre for Medium-Range Weather Forecasts, and its dataset contains hourly estimates of global atmosphere, ocean waves and land surface conditions since 1979, providing high spatio-temporal resolution environmental parameter support.

[0046] The GOPAF dataset is used to further improve the marine environmental information. The GOPAF dataset is provided by the European Copernicus Marine Service and is generated based on the Mercator Ocean Analysis and Forecast System. This dataset contains global marine three-dimensional analysis and prediction results since 2020, providing detailed basis for the analysis of ship operation characteristics in dynamic marine environment.

[0047] 2) Adopting outlier detection and missing value filling strategies to ensure data integrity and reliability.

[0048] Isolation Forest algorithm is used for outlier identification: Based on random subspace projection, a decision tree ensemble is constructed, and the anomaly degree of sample points is evaluated by quantifying the length of the isolation path required, and the observation values significantly deviating from the main data distribution are identified.

[0049] For the labeled abnormal data, the tail-cutting processing technique is applied for numerical constraint, and the extreme values exceeding the preset quantile threshold are converted to the numerical value of the appropriate quantile point, so as to reduce the interference of extreme values on subsequent statistical analysis while preserving the data distribution characteristics.

[0050] K-Nearest Neighbors (KNN) is used to fill in the missing values: Based on the inherent topological structure of the data and the similarity measure between samples, the missing items are estimated by calculating the weighted average value of the neighboring samples in the feature space.

[0051] Compared with simple mean or median filling, KNN interpolation method can more accurately capture the inherent correlation between data, significantly improving the interpolation accuracy.

[0052] S2, fuse the preprocessed multi-source data through the multi-dimensional data fusion framework based on spatio-temporal alignment.

[0053] 1) Normalize each data source using a standardized georeferencing system, and establish a unified time base and spatial coordinate system.

[0054] 2) Use inverse distance weighted interpolation to map the data of each data source in spatial resolution, and combine time resampling technology to realize the unification of time series data.

[0055] 3) Take the high-precision positioning information of AIS trajectory data as the main reference to construct a spatio-temporal index structure, accurately match and associate the navigation state parameters in the noon report of the ship with the meteorological elements provided by ERA5 and the ocean environmental data of GOPAF, and then form a multi-dimensional comprehensive feature vector.

[0056] S3, integrate the SHAP value evaluation results of multiple machine learning models, combine recursive feature evaluation and weighted voting mechanism, and comprehensively and objectively quantify the importance of each feature, so as to effectively avoid the selection bias that may be introduced by a single model.

[0057] In the problem of ship fuel consumption prediction, although multi-source data fusion can provide rich feature information, it often accompanies the problems of multiple collinearity between features, information redundancy, and significant differences in the prediction contribution of each feature.

[0058] 1) Determine a set of machine learning models with different structures and learning mechanisms, and train these models to predict the target variable on the complete training feature set. For each trained model, use SHAP to calculate its SHAP value on the training data. Calculate the average absolute SHAP value of each feature as the initial feature importance measure of the model, which is represented by the formula:

[0059] where, represents the initial feature importance measure, is the number of training samples, is the model For the th sample, the SHAP value of the th feature; represents the index of the model.

[0060] 2) According to the initial feature importance measure, all features are arranged in descending order to obtain the initial feature importance ranking list of each model .

[0061] 3) According to its feature importance ranking list , the subset containing the top most important features is constructed in turn, incremented from 1 to , denotes the total number of features.

[0062] For each feature subset , the models are trained on the corresponding training data subset and the predictive performance of each model is evaluated by 5-fold cross-validation to find the number of features that maximizes the model's coefficient of determination score , and the best feature subset is constructed according to the determined number of features, which is denoted as:

[0063]

[0064] where, is the best feature subset of model m, represents the th feature of the model; when , the coefficient of determination score of the model is taken as the weight of the weighted vote, that is, . represents the coefficient of determination obtained by training (on the training set) only using the top n ranked features of model m and evaluating on the validation set; represents the maximum value of the coefficient of determination.

[0065] 3) Calculate the weighted vote score of all features in the optimal subset based on the weight of the weighted vote, which is denoted as:

[0066] where, is the weighted vote score, is the performance weight of the model, is the set of models. This score reflects the comprehensive importance of the feature from the perspective of multiple well-performing models. Each model assigns its entire weight to each feature in its feature subset Sm by directly assigning the feature.

[0067] 4) According to the calculated weighted vote score , all features participating in the vote are ranked in descending order to obtain the final global feature importance ranking list .

[0068] 5) The optimal feature set is composed of all features with a weighted vote score greater than 1.

[0069] Since the performance weight of a single model is usually less than or equal to 1, It means that the feature is selected by at least two well-performing models at the same time, or is selected by multiple models and the cumulative weight of these models exceeds 1, which significantly enhances the confidence and consistency of the importance of the selected features across models, indicating that these features exhibit strong predictive value from different algorithm perspectives.

[0070] S3、In this embodiment, the machine learning model is six supervised learning models, including SVR, Random Forest (RF), Gradient Boosting (GB), XGBoost (XB), LightGBM (LB), and CatBoost (CB).

[0071] The original data set is divided into a training set and a test set in a ratio of 8:2; strict 5-fold cross-validation is performed on the training data set to evaluate the stability of the model performance and minimize the influence of random factors.

[0072] To optimize the model training efficiency and improve the prediction accuracy, the Z-score method is used to perform standardization preprocessing on the input features, and the formula is:

[0073] Among them, denotes the standardized feature value, is the original feature value, and represent the mean and standard deviation of the feature, respectively. Through standardization processing, each input feature is converted to a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0074] The hierarchical adaptive parameter space exploration method is constructed by combining Grid search (GS) and Random search (RS), using a hierarchical search structure and adaptive parameter space reduction strategy, including: Initial random search and refined search space: 1) Set the initial hyperparameter search space ; randomly sample groups of hyperparameter combinations from the initial hyperparameter search space , for each parameter group, evaluate the model performance on the training set and validation set using cross-validation or hold-out method, and obtain its evaluation error . 、 The formula for determining the parameter combination with the best performance in the current stage is:

[0075]

[0076]

[0077] where, is the number of random sampling times, denotes the dimension of hyperparameters, and are the lower and upper bounds of the i-th hyperparameter, respectively, denotes the i-th combination of hyperparameters, denotes the union of independent variables of the training set, denotes the union of dependent variables of the training set, denotes the feature matrix and label vector of the validation set, denotes the feature matrix and label vector of the test set, is the optimal solution of the initial random search. 2) Perform local search on the optimal solution of the initial stage

[0078] : Centered at , reduce the search space according to the scaling factor , construct the local parameter space ; divide each parameter dimension into equidistant points to form the grid hyperparameter set . For each combination of hyperparameters , evaluate the model performance again and determine the combination of hyperparameters with the best overall performance in the global range. The process is represented by the formula:

[0079]

[0080]

[0081] where, is the search space contraction factor, is the number of divisions in each dimension during the refined search, is the new search range for the i-th parameter, denotes the evaluation error after model training and validation for the combination of parameters , and is the finally determined global optimal combination of hyperparameters. Ship fuel consumption prediction is a complex nonlinear regression problem, and a single machine learning model often has difficulty capturing all features and patterns in the data. Through adaptive clustering multi-ensemble model, different model weights are adaptively assigned to different data clusters, realizing regionalized precise fusion.

[0082]

[0083] ​​The distribution characteristics of the model prediction are extracted, the samples are clustered into different groups based on the extracted characteristics, the fusion weight is optimized for each group, and the fusion result is further optimized through meta-learning. The problem to be solved by the adaptive clustering multi-ensemble model is: There are training samples and base models , and each model predicts the value of the sample as . The goal of the adaptive clustering multi-ensemble model is to find the optimal fusion function , so that the error between the fused prediction value and the true value can be minimized, and the process can be represented by the formula:

[0084] where is the feature vector of the th training sample, is the true target value of the th training sample, represents the final fusion prediction value of the th sample obtained by the fusion function .

[0085] 1) Extract features from the model prediction results, and cluster the samples into groups based on these features; for each sample , its cluster is represented by . For sample , define the feature vector , which contains statistical features and correlation features , and the formula is:

[0086]

[0087] where is the mean of the model prediction, is the standard deviation, is the range, is the median, and are the 25th and 75th percentiles, respectively. is the correlation coefficient of model and model for sample in the group.

[0088] Based on the feature vector of each sample, the K-means algorithm is used to cluster all samples into clusters. Clusters, number of clusters The silhouette coefficient is automatically determined. To ensure each cluster has a sufficient number of samples for weight optimization, clusters with fewer than a threshold are excluded. Clusters are merged into the nearest larger cluster. This process is expressed by the formula:

[0089]

[0090]

[0091] in, It is the first The center of each cluster, It is a sample The average distance to other samples in the same cluster, It is a sample The average distance to the nearest non-clustered sample, It is a cluster The center. Select. The largest The value is used as the optimal cluster number; It is the silhouette coefficient when the number of clusters is K; It is the Euclidean distance between the centers of cluster c' and cluster c, where c' and c are any two clusters.

[0092] 2) For each data cluster Find an optimal weight combination to fuse the data from... The predictions or information come from several different sources. By optimizing the weights, a more accurate final result is expected than that from any single source. The optimal fusion weight vector is calculated. The formula is expressed as:

[0093]

[0094] in, Let be the weight vector to be solved during the optimization process. The weights of the first model in the k-th cluster are... For the first The true value of each sample For the first The model for the first The predicted value for each sample, For the first The weights of each model. Constraints ensure that the calculated weights are accurate. This resulted in an effective proportional distribution.

[0095] 3) Instead of using the original base model prediction results directly, the base fusion results calculated in the last step and some statistical characteristics of these base predictions are used to construct a new meta-feature vector for each sample , which is then used as the input of a new machine learning model to produce the final prediction output, which can be represented as:

[0096]

[0097]

[0098] where, is the meta-feature vector for the th sample, which is the input of the meta-model . is the base fusion prediction value for the th sample. It is the weighted average prediction result calculated according to the optimal weights of the cluster it belongs to, which represents the best estimate after fusion. is the standard deviation, and represent the maximum and minimum values of the prediction values, is the final prediction value for the th sample.

[0099] Example 1 S1, collect multi-source data of the ship and perform data preprocessing to obtain preprocessed multi-source data.

[0100] 1) The original data used is from the noon report data of a certain ocean-going bulk carrier provided by an international shipping company, with a data recording time span from January 14, 2021 to July 30, 2024, covering 479 days of sailing records. The main fields recorded in the noon report include speed, draft, main engine power, heading, position, and manually observed weather data.

[0101] 2) High-frequency verification of speed and position is performed using AIS data, and ERA5 reanalysis weather data and GOPAF ocean forecast data are introduced as environmental variables for supplementation. ERA5 data is published by the European Centre for Medium-Range Weather Forecasts, providing hourly resolution global weather reconstruction information; GOPAF data is provided by the European Copernicus Marine Service, covering global ocean three-dimensional parameter analysis and prediction results.

[0102] 3) Isolation Forest algorithm is used to detect and remove outliers in each numerical field, and Winsorization boundary compression operation is performed on abnormal samples to prevent extreme values from interfering with the model. Missing data is estimated by K-Nearest Neighbor Imputation (KNN) to improve data integrity.

[0103] 4) After data cleaning, implement unified spatio-temporal reference processing, and use WGS-84 geographic coordinate system and UTC time format to unify the spatial and temporal information of each data source. For the spatial accuracy mismatch between ERA5 and GOPAF grid data and noon report or AIS data, inverse distance weighted (IDW) interpolation processing is implemented, and time resampling strategy is used to unify the sampling frequency. Finally, the high-precision trajectory record in the AIS data is used as the reference to build a spatio-temporal registration index, and the multi-dimensional accurate fusion of the noon report, weather data and ocean data is realized.

[0104] After the above process, the feature sources and their corresponding feature names of the multi-source data fusion are obtained, as shown in Table 1.

[0105] Table 1

[0106] S2, by SWFS method, integrate SHAP value evaluation results of multiple machine learning models, combine recursive feature evaluation and weighted voting mechanism, comprehensively and objectively quantify the importance of each feature, effectively avoid the selection bias that may be introduced by a single model.

[0107] In this embodiment, the machine learning model is six kinds of supervised learning models, including SVR, Random Forest (RF), Gradient Boosting (GB), XGBoost (XB), LightGBM (LB) and CatBoost (CB).

[0108] SHAP value evaluation on different feature subsets, and weighting according to the corresponding subset prediction performance, can obtain more stable and reliable feature importance ranking. Figure 1 The final feature subset determined according to the standard is shown.

[0109] S3, by HAPSE method, the precision and efficiency optimization balance is realized in the HPO task of ship fuel consumption prediction.

[0110] Compared with the traditional GS, HAPSE saves about 50% of the calculation time under the condition of maintaining the same or even better accuracy; compared with RS and BO, the prediction performance is improved by about 5%. Figure 2The logarithmic coordinate axis in the figure clearly shows that HAPSE reduces the optimization time from the order of thousands of seconds to the order of hundreds of seconds, and such an order of magnitude improvement is of great value to large-scale data analysis and real-time decision-making systems. Experiments also show that HAPSE maintains stable performance advantages under different model architectures, reflecting its robustness and universality. In particular, the excellent performance of the advanced integrated model CB demonstrates the adaptability of HAPSE in handling complex nonlinear relationships, which is of great practical significance for complex prediction tasks such as ship fuel consumption affected by multiple factors, and provides reliable technical support for improving prediction accuracy and model deployment efficiency.

[0111] The total model constructed by integrating multiple models proposed in Example 1 is referred to as the ACME model.

[0112] S3, six mainstream models in ship fuel consumption prediction are used as benchmarks, and three commonly used fusion methods are selected for comparison with the ACME model in this method. All models are based on the test set, and five-fold cross-validation is performed to ensure the stability of the evaluation results. Table 2 summarizes the overall performance of each model.

[0113] Overall performance, the ACME model has achieved significant advantages in various evaluation indicators. Specifically, reaches 0.9088, far exceeding all comparison models, indicating that it can explain more than 90% of the total variation of observed data, which not only reflects the model's ability to capture the main factors affecting fuel consumption, but also shows excellent data fitting effect. At the same time, the RMSE and MAE values of the ACME model are 1.7817 and 0.8322, respectively, further verifying the model's outstanding performance in controlling absolute and relative errors. In addition, its MAPE is only 2.73%, highlighting the model's robustness and precision advantage in handling high-dimensional, nonlinear fuel consumption prediction problems. It is worth noting that the EV index, as an important indicator of measuring the model's explanation of the overall variance of the data, has a value of 0.9096, which further confirms that the ACME model can effectively capture and explain the main information in the data under the background of high-dimensional nonlinear data through clustering partitioning and adaptive weight adjustment.

[0114] Table 2

[0115] Figure 3is the consistency distribution diagram of ACME model prediction value and actual value. It can be observed that most data points are highly distributed near the y=x diagonal line, and the density hot spot area is basically consistent with the ideal straight line, which directly proves the high precision fitting ability of the ACME model under the normal navigation working condition. In order to further reveal the advantages of the ACME model in the task of ship fuel consumption prediction, the internal fusion mechanism of the model is further discussed through multi-angle visualization. Combined with the analysis of clustering results and fusion weights, the following aspects are described in detail.

[0116] Using clustering method to divide data into different subspaces is helpful to capture the local feature difference existing in fuel consumption. Figure 4 For the distribution of samples in each cluster, it can be seen that the number of samples in cluster 1 (labeled 1) is significantly more than that in cluster 0 (labeled 0), containing about 316 and 67 samples respectively. This uneven distribution reflects the inherent characteristics of ship fuel consumption data, that is, under most normal navigation conditions, the fuel consumption mode shows relatively consistent characteristics (cluster 1), while under specific working conditions or abnormal environmental conditions (cluster 0), the fuel consumption mode shows unique feature structure.

[0117] In the ACME model architecture, the contribution weight of each base model to different cluster samples is a key factor to explain the performance of the model. As shown in Figure 5 , the present application analyzes the average fusion weight of six base models in two clusters. From Figure 5 , it can be clearly observed that in cluster 0, GradientBoosting and XGBoost models obtain relatively high weights (about 0.172), while LightGBM has lower weight (about 0.155); in cluster 1, CatBoost and RandomForest models perform more outstandingly, with weights of about 0.178 and 0.171 respectively, while SVR model has the lowest weight (about 0.152). This differentiated weight allocation fully embodies the core advantage of ACME model - it can adaptively adjust the contribution of each base model according to the local characteristics of data, thereby targetedly improving the overall prediction accuracy. It is worth noting that CatBoost model obtains higher weight in both clusters, which is closely related to its innate advantage in dealing with category features and unknown distribution data. Although SVR model has lower weight in cluster 1, it performs well in cluster 0, which reflects its specific advantage in dealing with structural data. This differentiated weight allocation mechanism based on clustering enables ACME model to fully exploit the advantages of each base model in specific data subspace, thereby achieving significant improvement in overall prediction performance.

[0118] In summary, through a comprehensive comparison of various base models and fusion methods across multiple indicators, and a systematic study combining sample clustering distribution characteristics, model fusion weight analysis, and principal component analysis of the feature space, this study demonstrates the superior performance of the ACME model in predicting ship fuel consumption. This model effectively addresses the shortcomings of traditional fusion algorithms in handling data diversity and local anomalies by fully exploring the local features within the data and performing differentiated modeling for different subspaces. The successful implementation of the ACME model not only provides a new technical path for accurate prediction of ship fuel consumption but also offers a methodological framework that can be referenced for handling similar high-dimensional nonlinear regression problems.

[0119] S4. The interpretability of the model is analyzed and discussed using the SHAP method. Figure 6 The results of global feature importance analysis based on the SHAP method are presented, clearly showing significant differences in the contribution of each feature to fuel consumption prediction. Figure 6 It can be observed that among the many factors affecting ship fuel consumption, daily sailing time, total sailing time, and daily sailing distance rank among the top, which is highly consistent with general empirical patterns in the shipping industry. Furthermore, main engine speed also ranks highly in the importance ranking of characteristics, indicating that the operating status of the ship's propulsion system has a direct and crucial impact on fuel demand.

[0120] At the medium importance level, the remaining distance to the destination port and the cumulative sailing distance, with SHAP values ​​of 0.3913 and 0.2649 respectively, indicate their significant impact on the prediction results. Although these route planning features are less important than time and dynamic parameters, they indirectly regulate fuel consumption by influencing the captain's choice and adjustment of the route. On the other hand, environmental factors such as 2-meter sea surface temperature and atmospheric pressure also show some importance in the SHAP analysis. Although these environmental factors are less important than the aforementioned features, their indirect impact on fuel consumption cannot be ignored. These environmental features can significantly alter ship drag and sailing efficiency, thus having a ripple effect on fuel consumption.

[0121] Figure 7 The distribution of SHAP values ​​across various feature dimensions is presented as a scatter plot, with a color gradient (from blue to yellow) visually representing the magnitude of each sample's value on the corresponding feature. It can be observed that daily sailing time, as the most influential key variable, exhibits a clear positive correlation with its SHAP value; high values ​​(yellow dots) generally and steadily increase fuel consumption, while low values ​​(blue dots) significantly reduce the predicted results. This phenomenon not only aligns with the principle of energy conservation and the fundamental laws of ship dynamics but also provides a quantitative basis for optimizing sailing time as a core strategy for fuel management.

[0122] In summary, the SHAP interpretability analysis based on the ACME model not only quantifies the impact of each feature on ship fuel consumption but also reveals complex feature interaction mechanisms and nonlinear influence patterns. This invention provides technical support for a deeper understanding of ship energy efficiency mechanisms, optimization of fuel management strategies, and the design of more precise monitoring systems.

[0123] 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. A method for predicting ship fuel consumption by fusing multi-source data, characterized in that, include: Collect historical data related to ship fuel consumption to construct a training set; The multiple machine learning models are trained based on the training set, and the weights of each machine learning model are determined based on the prediction results of the training set as preset weights. Collect relevant features of ship fuel consumption and input them into multiple machine learning models; Each machine learning model outputs its own fuel consumption prediction results; The final prediction result of ship fuel consumption is obtained by fusing the prediction results of various machine learning models based on preset weights.

2. The method according to claim 1, characterized in that, Also includes: The SHAP value of each machine learning model is calculated based on the training results, serving as an initial measure of feature importance. All features are sorted in descending order based on the initial feature importance metric to obtain the initial feature importance ranking list for each model; Based on the feature importance ranking list, construct sequentially the list containing the top features. A subset of the most important features , Increasing from 1 to , Represents the total number of characteristics; Indicates the index of the model; Calculate the weighted voting scores of features in all optimal subsets based on the weights of the weighted voting: For each feature subset Each model is trained on its corresponding subset of training data, and the predictive performance of each model is evaluated using 5-fold cross-validation to identify the number of features that cause the model's coefficient of determination to peak. And based on the determined number of features, construct the optimal feature subset, expressed by the formula: in, The optimal feature subset for model m The representative model's first One characteristic; when At that time, the decision coefficient score of the model is used as the weight of the weighted vote, that is... ; This represents the determination coefficients obtained by training the model m using only the first n ranked features and evaluating them on the validation set. This indicates that the coefficient of determination has reached its maximum value; The weighted voting score for features in all optimal subsets is calculated based on the weights of the weighted voting, as expressed by the formula: in, To calculate the weighted vote score, For the performance weights of the model, A collection of models; Features with a weighted voting score greater than 1 are used as relevant features of ship fuel consumption.

3. The method according to claim 1, characterized in that, The weights of each machine learning model are determined based on the prediction results from the training set, including: Based on feature sets, hyperparameter optimization of various machine learning models is performed using grid search combined with random search, and prediction results of each machine learning model are obtained. Extract features from the prediction results corresponding to the samples; Based on the results features, the samples are clustered into a preset number of data clusters using K-means, and the model fusion weights are solved with the goal of minimizing the error between the fused predicted values ​​and the true values.

4. The method according to claim 1, characterized in that, The relevant characteristics of ship fuel consumption include: Relative surge direction angle, ship heading, average wave period, average speed over the entire journey, atmospheric pressure, remaining distance to the destination port, main engine speed, sea surface temperature at 2 meters, cumulative sailing distance, total sailing time, daily sailing distance, and daily sailing time.

5. The method according to claim 1, characterized in that, The machine learning model includes: SVR, Random Forest, Gradient Boosting, XGBoost, LightGBM and CatBoost.

6. The method according to claim 1, characterized in that, The resulting features include: The predicted mean, standard deviation, extreme values, median, 25th percentile, and 75th percentile for each model.

7. The method according to claim 1, characterized in that, The historical data related to ship fuel consumption includes: Ship midday report, AIS data, and ERA5 data.

8. The method according to claim 3, characterized in that, The process of optimizing the hyperparameters of multiple machine learning models based on feature sets through grid search combined with random search, and obtaining the prediction results of each machine learning model, includes: Setting the initial hyperparameter search space Random sampling from the initial hyperparameter search space Group of hyperparameter combinations For each set of parameters The model performance is evaluated on the training and validation sets using cross-validation or hold-out methods to obtain its evaluation error. ; , The formula for determining the optimal parameter combination for the current stage is expressed as: in, The number of random samples, Indicates the hyperparameter dimension. and The first Lower and upper bounds of each hyperparameter. This represents the i-th hyperparameter combination; This represents the set of independent variables in the training set. This represents the set of dependent variables in the training set; This represents the feature matrix and label vector of the validation set; This represents the feature matrix and label vector of the test set; This is the optimal solution for the initial random search; For the optimal solution in the initial stage Perform a local search: Centered on the scaling factor Narrowing the search space and constructing a local parameter space Divide each parameter dimension into Equidistant points constitute a gridded hyperparameter set. For each set of hyperparameter combinations We then re-evaluate the model performance and determine the optimal combination of hyperparameters for overall performance across the entire model.

9. The method according to claim 3, characterized in that, The step of clustering the samples into a preset number of data clusters using K-means based on the result features, and solving the model fusion weights with the objective of minimizing the error between the fused predicted values ​​and the true values, includes: For each sample The cluster to which it belongs uses Indicate; for the sample Define the feature vector , including statistical characteristics Correlation characteristics The formula is expressed as: in, It is the mean of the predictions from all models. It is the standard deviation. It's extremely bad. It is the median. and These are the 25th and 75th percentiles, respectively; It is a model and model For the sample The predictive correlation coefficient of the group; The K-means algorithm was used to cluster all samples into groups. Clusters, number of clusters Automatically determined by the silhouette coefficient, samples with fewer than a threshold are excluded. Clusters are merged into the nearest larger cluster: in, It is the first The center of each cluster, It is a sample The average distance to other samples in the same cluster, It is a sample The average distance to the nearest non-clustered sample, It is a cluster The center; selection The largest The value is used as the optimal cluster number; It is the silhouette coefficient when the number of clusters is K; is the Euclidean distance between the centers of cluster c' and cluster c, where c' and c are any two clusters; Calculate the optimal fusion weight vector The formula is expressed as: in, Let be the weight vector to be solved during the optimization process. For the first The true value of each sample For the first The model for the first The predicted value for each sample, For the first The weights of each model; constraints ensure that the calculated weights are... This resulted in an effective proportional allocation.

10. The method according to claim 1, characterized in that, The final ship fuel consumption prediction result is obtained by fusing the prediction results of various machine learning models based on preset weights, including: For each sample A new, more informative meta-feature vector is constructed and used as input to the machine learning model to produce the final prediction output, expressed by the formula: in, For the first The meta-feature vector of each sample is the meta-model. Input; For the first The base fusion prediction value for each sample; this is based on its cluster. Optimal weight The calculated weighted average prediction result; It is the standard deviation. and These represent the maximum and minimum predicted values, respectively. It is the first The final predicted value for each sample.

Images