A real-time optimization method for power of marine main engine based on hybrid network model

Abstract

The application provides a ship main engine power real-time optimization method based on a hybrid network model, belongs to the technical field of ship energy saving, and constructs a hybrid neural network model. The hybrid neural network model is based on a physical information neural network, introduces a KAN network as a front-end network structure, is used for establishing a high-dimensional nonlinear mapping from a speed, a ship type parameter matrix, a propulsion efficiency, a fuel conversion efficiency and an environmental factor to a minimum power of a main engine, trains the hybrid neural network model by using a composite loss function, and the composite loss function is composed of a mean square error loss, a dynamics constraint loss, a propulsion efficiency constraint loss and a fuel consumption constraint loss. By using the trained hybrid neural network model, real-time collected ship running parameters and environmental parameters are received, a minimum power prediction value of the ship main engine is output through one-time forward propagation calculation, and real-time optimization of the main engine power is realized.

Description

Technical Field

[0001] This invention relates to the field of marine energy conservation technology, specifically to a method for real-time optimization of ship main engine power based on a hybrid network model. Background Technology

[0002] Against the backdrop of the rapid development of green ship and intelligent shipping technologies, real-time optimization of ship main engine power has become an important technological approach to improve navigation energy efficiency and reduce carbon emissions. Currently, the following mainstream technical solutions have emerged regarding the prediction and assessment of ship main engine power:

[0003] The first category is traditional evaluation methods based on physical models. These methods mainly include empirical or semi-empirical formula estimation methods, ship model tank tests, and computational fluid dynamics (CFD) numerical simulations. Empirical formula methods establish mathematical relationships between ship parameters, speed, and power through statistical analysis of a large amount of ship test data; ship model tests directly measure the resistance and propulsion performance under different operating conditions by towing or self-propelled tests on scaled-down ship models in specialized tanks; and CFD simulation technology uses computers to solve the fluid dynamics control equations and perform refined numerical simulations of the flow field around the ship, thereby obtaining highly accurate resistance and power prediction values.

[0004] The second category is data-driven machine learning and deep learning methods, such as artificial neural networks (ANN), support vector regression (SVR), and long short-term memory networks (LSTM). These methods can automatically learn and construct complex nonlinear mapping relationships between input parameters and main engine power or fuel consumption directly from historical data containing multi-dimensional information such as speed, draft, wind speed, and wave height.

[0005] The third category is the integration of physical information and data-driven approaches. This method treats the physical laws describing ship motion and energy conversion as prior knowledge, encoding their mathematical expressions, such as partial differential equations or algebraic equations, into the loss function of a neural network. The optimization objective includes not only minimizing the difference between model predictions and actual observations, but also minimizing the residuals of the model output to the physical control equations, thereby enhancing the model's generalization ability and the physical interpretability of the prediction results in cases of sparse or missing data.

[0006] While existing technologies offer multiple implementation paths for predicting ship main engine power, when applied to high-precision, real-time dynamic optimization scenarios during ship navigation, they still exhibit limitations and defects determined by their technical principles and which are interconnected.

[0007] First, the fundamental flaw of traditional physical model evaluation methods lies in their lack of real-time capability. These methods rely on offline analysis. Ship model testing is constrained by complex physical facilities and lengthy testing cycles, while CFD simulation requires enormous computational resources and extremely high computational time costs. Therefore, they cannot provide instantaneous responses and calculations for dynamically changing operating parameters such as speed and sea state during navigation, nor can they provide real-time power prediction data for online energy efficiency management of ships, thus failing to meet the dynamic optimization needs of intelligent shipping.

[0008] Secondly, the main drawback of purely data-driven models is the lack of physical consistency in their predictions, leading to insufficient generalization ability and reliability. These models rely entirely on statistical fitting of historical data, failing to incorporate physical mechanisms such as ship hydrodynamics into the modeling process. Consequently, when faced with operating conditions not covered by the training dataset or with sparse data, their output may contradict fundamental physical laws.

[0009] Finally, although the PINN model theoretically improves physical consistency, when dealing with complex engineering problems with high-dimensional inputs and strong nonlinear mappings, such as ship main engine power, the PINN model has a slow convergence speed during training and limited final prediction accuracy, thus limiting the learning efficiency and final performance of the entire network model. Summary of the Invention

[0010] This invention proposes a real-time optimization method for ship main engine power based on a hybrid network structure. By constructing a high-dimensional mapping model from speed, hull parameter matrix, propulsion efficiency, fuel conversion efficiency, and environmental factors to the minimum power requirement of the main engine, it directly predicts the minimum power demand of the main engine under any operating condition. The hybrid neural network used is based on a physical information neural network, incorporating the Kolmogorov-Arnold Network (KAN) as the front-end network structure. The KAN network uses trainable spline functions instead of fixed activation functions, further improving the model's expressive power and significantly enhancing network iteration efficiency and interpretability. Simultaneously, the model integrates multiple constraint mechanisms, including propulsion efficiency-speed relationship constraints, fuel consumption-speed constraints, and ship dynamics energy efficiency indicators, ensuring that the network output possesses physical consistency and engineering feasibility.

[0011] A real-time optimization method for ship main engine power based on a hybrid network model includes the following steps:

[0012] S1. Construct a hybrid neural network model, in which a KAN network is introduced as the front-end network structure to establish a high-dimensional nonlinear mapping from speed, ship type parameter matrix, propulsion efficiency, fuel conversion efficiency and environmental factors to the minimum power of the main engine.

[0013] S2. The hybrid neural network model is trained using a composite loss function, which is composed of a weighted sum of mean square error loss, dynamic constraint loss, propulsion efficiency constraint loss and fuel consumption constraint loss.

[0014] S3. Using the trained hybrid neural network model, the ship's operating parameters and environmental parameters are collected in real time. Through one forward propagation calculation, the minimum power prediction value of the ship's main engine is output, thereby realizing the real-time optimization of the ship's main engine power.

[0015] Preferably, in step S2, the mean squared error loss function By calculating the deviation between the predicted main engine power and the actual main engine power, we ensure that the predicted main engine power accurately fits the historical dataset collected from the ship's actual voyages.

[0016] ;

[0017] In the formula, For hybrid neural network models targeting the first The host power predicted for each input sample; The actual main engine power recorded during the ship's actual navigation; The first of the ship's speeds One input sample; The total number of samples.

[0018] Preferably, a dynamic constraint loss function is constructed. :

[0019] ;

[0020] Dynamic constraint loss function By calculating the predicted host power Compared with the theoretical power of the physical model At the corresponding speed The sum of squared differences quantifies the degree to which the predicted behavior deviates from fundamental physical laws. This is achieved in conjunction with the mean squared error loss function during training. minimize.

[0021] Preferably, a propulsion efficiency constraint loss function is constructed. :

[0022] ;

[0023] In the formula, For speed The propulsion efficiency of the ship. Represents the predicted host power and known propulsion efficiency The calculated theoretical effective power, It is the actual reference effective power.

[0024] Preferably, a fuel consumption constraint loss function is constructed. :

[0025] ;

[0026] In the formula, at the th The actual fuel consumption under each sample operating condition is: The known fuel conversion efficiency is .

[0027] Preferably, by introducing weighting coefficients , , and Perform a linear weighted summation to construct the final total loss function. :

[0028] ;

[0029] By minimizing the total loss function Find a set that can make The internal parameters of the KAN network that reach the minimum value.

[0030] Preferably, in step S1, a three-layer KAN network structure is constructed. The input layer is used to receive a set of preprocessed and normalized multidimensional feature vectors, and the multidimensional feature vectors at least include the ship's real-time speed. Ship type parameter matrix , Promotion efficiency Fuel conversion efficiency and comprehensive environmental factors ;

[0031] The learnable activation function on each edge of the KAN network structure for:

[0032] ;

[0033] in These are the coefficients that the network needs to learn. Let be the i-th basis function.

[0034] Preferably, the dynamic constraint loss function is based on a semi-empirical cubic polynomial model describing the power-speed relationship of a ship in still water, embedding prior knowledge of ship hydrodynamics into the network model, and forcing the predicted main engine power output by the KAN network. In terms of function form, it follows the physical relationship between the ship's main engine power and speed:

[0035] ;

[0036] In the formula, The calculated theoretical main engine power; For ship speed; Captain; As the coefficient of the cube term of speed, As the coefficient of the square term of the speed, For the linear term coefficient of speed, For constant terms, The drag term is a term that characterizes the resistance related to the ship's geometry.

[0037] Preferably, effective power for:

[0038] ;

[0039] In the formula, The total power requirement of the ship; the propulsion efficiency of the ship. for:

[0040] ;

[0041] In the formula, The optimal speed for the ship; This represents the maximum ship propulsion efficiency. To advance the system characteristic-related constants.

[0042] Preferably, the ship's total power requirement Fuel consumption per unit time The relationship between them is:

[0043] ;

[0044] In the formula, This refers to the fuel conversion efficiency of the main unit.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] First, this invention innovatively employs a KAN network as the main network architecture. Compared to the fixed activation functions commonly used in existing traditional PINN networks, the KAN network of this invention, by setting learnable activation functions parameterized by spline functions on the network edges, endows the model with unprecedented functional expressive power. This enables the model to capture and characterize the highly complex and strongly nonlinear mapping relationship between ship main engine power and multi-dimensional input variables such as speed and environment with higher fidelity, thereby achieving higher accuracy and precision in the final prediction results compared to existing technologies.

[0047] Secondly, compared to purely data-driven black-box models, this invention systematically integrates four key constraints—mean squared error constraint, dynamics constraint, propulsion efficiency constraint, and fuel consumption constraint—into the total loss function. This forces the network to adhere to the fundamental laws of ship hydrodynamics, energy conservation, and fuel economy during the learning process. This fundamentally solves the problem of traditional machine learning models potentially producing predictions that violate physical principles. It ensures that even under conditions where training data is sparse or unavailable, the model's output remains reasonable, reliable, and engineering-interpretable, significantly improving the model's generalization ability and credibility in critical decision-making scenarios.

[0048] Third, traditional PINN networks, due to their fixed activation functions, have limited ability to approximate complex functions, often resulting in slow optimization processes and a tendency to get trapped in local optima. The KAN network architecture used in this invention, with its adaptive adjustment capability of activation function shape, can find the optimal solution path that satisfies multiple constraints more directly and efficiently. Furthermore, the preferred use of orthogonal basis functions such as Chebyshev polynomials to construct splines enhances the numerical stability of the training process. Therefore, while achieving the same or even higher accuracy, the training process of this invention is more efficient, reducing the time and computational costs required for model development.

[0049] Fourth, compared to traditional offline and static analysis methods such as ship model testing or CFD simulation, this invention, after completing one offline training cycle, requires only one network forward propagation calculation for its online prediction process, which can be completed within milliseconds. This high-frequency response real-time prediction capability allows it to be seamlessly integrated into shipboard energy efficiency management systems or shore-based route optimization platforms, providing core technical support for advanced applications such as dynamic speed optimization and real-time energy efficiency monitoring during ship navigation, and possessing clear and significant engineering practical value. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the KAN network structure of the present invention;

[0051] Figure 2 Design of the PINN loss function for this invention;

[0052] Figure 3 The KAN network spline function of this invention;

[0053] Figure 4 This is the real-time prediction data of the ship's main engine power for this invention. Detailed Implementation

[0054] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0055] To ensure the rigor and reliability of the model, this invention constructs a composite loss function consisting of multiple parts. This function systematically integrates data-driven fitting errors and several key physical constraints, thereby performing global optimization of the entire network model. The loss function under physical constraints consists of the following four parts:

[0056] I. Mean Squared Error Loss Function

[0057] The mean squared error loss function calculates the deviation between the model's predicted values ​​and the actual values, ensuring that the predicted engine power output by the KAN network accurately fits the historical dataset collected from actual ship navigation. The specific calculation formula is as follows:

[0058] ;

[0059] In the formula, For hybrid neural network models targeting the first The host power predicted for each input sample; The actual main engine power recorded during the ship's actual navigation; The first of the ship's speeds One input sample; The total number of samples.

[0060] The mean squared error loss function By quantifying and minimizing the deviation between the model's predicted values ​​and the actual observed values, the learnable spline function coefficients within the network are systematically adjusted. This allows the high-dimensional nonlinear mapping constructed by the network to statistically approximate the empirical relationships contained in the training data, thus endowing the model with basic prediction accuracy and data fitting ability.

[0061] However, its scope is limited to ensuring the fidelity of existing data, and it cannot actively ensure that the model's predictive behavior strictly follows known physical laws under all conditions, especially in regions where training data is sparse or missing, where its prediction results may exhibit physical inconsistencies. Therefore, to endow the model with stronger generalization ability and ensure the physical consistency of its output results, it is necessary to introduce subsequent physical constraint terms to collaboratively correct and standardize the results of the data-driven terms.

[0062] II. Dynamic Constraint Loss Function

[0063] The dynamic constraint loss function used in this invention is based on a semi-empirical cubic polynomial model that describes the power-speed relationship of a ship in still water. Its purpose is to embed prior knowledge of ship hydrodynamics into the network model, overcoming the fundamental flaw of purely data-driven models that may produce physically inconsistent predictions in sparse training data regions. This constraint forces the predicted main engine power output by the KAN network to... The function follows the physical relationship between ship engine power and speed, thus significantly improving the model's generalization ability and engineering reliability. The specific calculation formula is as follows:

[0064] ;

[0065] In the formula, The calculated theoretical main engine power; For ship speed; Captain; As the cubic term coefficient of speed, it mainly characterizes the power demand dominated by wave-making resistance when a ship is sailing at high speed, reflecting the nonlinear characteristic of power increasing sharply with speed. As the coefficient of the square term of speed, it is mainly related to the frictional resistance between the hull and the water; This is the coefficient for the linear term of speed, which is related to additional drag. This is a constant term, representing the ship's basic power requirements at zero or low speeds; This represents the resistance term related to the ship's geometry.

[0066] Currently, the vast majority of industry professionals use traditional methods to calculate these coefficients, such as... This was obtained through model experiments or computational fluid dynamics simulations. Calculated using the standard method ITTC-57 formula recommended by the International Technical Committee on Marine Engineering (ITTC). The actual performance curve of the ship is obtained through multiple tests or simulations. This is obtained through ship no-load testing. Wave resistance is obtained through empirical formulas or simulation experiments. However, this invention uses the aforementioned physical information neural network to predict ship resistance, achieving a faster and more accurate result. , , and .

[0067] Based on this physical model, the present invention constructs the following dynamic constraint loss function:

[0068] ;

[0069] Should The function calculates the network's predicted power. The theoretical power of the above physical model At the corresponding speed The sum of squared differences is used to quantify the degree to which the network's predictions deviate from fundamental physical laws. During training, this is achieved by co-minimizing the mean squared error loss function. Instead of simply requiring the network to fit discrete data points, it is forced to learn a smooth mapping relationship that conforms to hydrodynamic trends across the entire function domain. Even in operating conditions where no actual data points are available, this constraint can still guide the network's output to follow the correct physical trends, thereby effectively avoiding overfitting and ensuring the physical consistency of the model's predictions and its reliable inference ability under unknown operating conditions.

[0070] III. Propulsion Efficiency Constraint Loss Function

[0071] The purpose of the propulsion efficiency constraint loss design is to treat the energy conversion relationship in the ship propulsion system as a strong constraint to ensure the predicted main engine power output by the KAN network. Not only does it closely approximate the true value numerically, but it also conforms to basic engineering principles in terms of energy transfer chain, thereby further enhancing the physical realism and accuracy of the model.

[0072] The physical basis of this constraint stems from the definition of effective power in ship propulsion theory. The total power requirement of the ship's main engine. After being transmitted through the shafting and propeller, the power ultimately used to propel the ship against water resistance is called effective power. The conversion efficiency between total power and effective power is called propulsion efficiency. This fundamental energy transformation relationship can be expressed by the following formula:

[0073] ;

[0074] In the formula, The effective power of the ship; The total power requirement of the ship; The propulsion efficiency of a ship can be expressed as:

[0075] ;

[0076] In the formula, The optimal speed for the ship; This represents the maximum ship propulsion efficiency. The constants related to the characteristics of the propulsion system are determined by the parameters of the thruster.

[0077] Based on this physical model, this invention constructs the following propulsion efficiency constraint loss function. :

[0078] ;

[0079] In the formula, for The propulsion efficiency of a ship at its speed. This term, in its physical sense, is based on the total power currently predicted by the network. and the known propulsion efficiency under this operating condition The calculated theoretical effective power, while This is the true benchmark effective power under that operating condition, derived from ship navigation test data or authoritative empirical formulas. Therefore, What is being quantified is the difference between the energy conversion results implied by network predictions and the generally accepted energy conversion benchmark.

[0080] During model training, by minimizing It can force the network to adjust its internal spline function to fit the total power data while taking into account the rationality of its prediction results in terms of energy conversion. This effectively avoids the model learning solutions that violate the law of conservation of energy and are meaningless in engineering simply to pursue a numerical match in total power. This greatly enhances the credibility and engineering value of the model's prediction results in actual ship energy efficiency assessment applications.

[0081] IV. Fuel Consumption Constraint Loss Function

[0082] The fuel consumption constraint loss function integrates the economic indicators of ship operation, namely fuel consumption patterns, as a key physical constraint into the model, ensuring that the predicted main engine power output by the KAN network is accurate. It must not only meet the requirements of hydrodynamics and energy conversion efficiency, but also conform to the fuel economy law of the engine, so that the prediction results of the model can meet the power requirements while also having reasonable economic considerations.

[0083] The physical basis of this constraint is the direct relationship between main engine power and fuel consumption rate. This refers to the amount of fuel consumed per unit time. The effective power generated depends on the fuel conversion efficiency of the main engine. This conversion efficiency characterizes the mechanical work that can be converted from a unit mass of fuel. Therefore, it determines the total power requirement of a ship. The relationship between fuel consumption and fuel consumption can be determined by the following formula:

[0084] ;

[0085] In the formula, The amount of fuel consumed per unit of time; Fuel conversion efficiency is the efficiency of the main engine, representing the effective power generated per unit of fuel. It is determined by the inherent parameters of the engine and the physical properties of the fuel. The value is deliberately smaller than that calculated by empirical formulas because the relationship between fuel consumption and speed and power is non-linear during actual navigation. As the ship's speed increases, the rate of increase in fuel consumption is usually faster than the rate of increase in power.

[0086] Based on this relationship between energy and consumption, this invention constructs the following fuel consumption constraint loss function. :

[0087] ;

[0088] In the formula, It is based on the first Actual fuel consumption under sample operating conditions and known fuel conversion efficiency The theoretical power requirement is calculated in reverse. Therefore, What is quantified is the predicted host power of the network. The discrepancy between the actual fuel consumption and the theoretical power demand calculated based on actual fuel consumption.

[0089] During model training, by By incorporating and minimizing the total loss function, the network is forced to ensure that its final output power prediction matches actual fuel consumption data when learning the complex relationship between power and factors such as speed and environment. This not only further improves the accuracy of power prediction but also directly links the model's prediction results to the ship's operating cost indicators (fuel consumption). This ensures that the main engine power optimization results based on this model are not only technically feasible but also economically reasonable and well-founded, thereby greatly enhancing the value of this invention in practical applications such as green ship development and intelligent navigation energy efficiency management.

[0090] like Figure 2 As shown, the aforementioned four independent loss function components are combined by introducing corresponding weight coefficients. , , and Then, a linearly weighted sum is performed to construct the final total loss function. :

[0091] ;

[0092] In this function, the weight coefficients for each item are hyperparameters that can be determined through methods such as cross-validation, based on different ship types and application scenarios, to balance the relative importance between data fidelity and various physical constraints. The ultimate optimization goal of this invention in the model training process is to find a set of parameters that enable this... The internal parameters of the KAN network reach their minimum value. By minimizing this total loss function, this invention ultimately obtains a fully trained and optimized hybrid network model that accurately represents the nonlinear mapping relationship from multidimensional inputs to main engine power. This directly solves the core technical challenge that existing technologies cannot simultaneously satisfy real-time performance, accuracy, and physical consistency, resulting in a high-frequency response, high-precision real-time prediction tool for ship main engine power that can be directly applied to engineering practice. The real-time prediction data for ship main engine power is as follows: Figure 4 As shown.

[0093] Example 2

[0094] A preferred embodiment of the present invention is described in detail, which realizes real-time optimization of ship main engine power based on a hybrid network model.

[0095] Figure 3 (a) uses the Basis Function to show the graphs of the first few orders of Chebyshev Polynomials of the First Kind and their corresponding formulas. Figure 3 (b) Orthogonalization shows the graphs and formulas of the first few orders of Legendre polynomials. Both sets of polynomials are complete orthogonal polynomial systems, representing a specific and effective technical choice for implementing this invention.

[0096] In a specific embodiment, it is first necessary to construct, as follows: Figure 1 The three-layer KAN network model is shown. The input layer of this model receives a set of preprocessed and normalized multidimensional feature vectors, which at least contain the ship's real-time speed. Ship type parameter matrix , Promotion efficiency Fuel conversion efficiency and comprehensive environmental factors The learnable activation function for each edge in the network. All are represented as these basis functions A linear combination, i.e. ,in These are the coefficients that the network needs to learn. For example, when using Chebyshev polynomials, the network optimizes... coefficient, will , By linearly superimposing the basis functions, any desired activation function shape can be constructed with extremely high accuracy and flexibility. To effectively train the above network structure, a composite loss function is then needed. For four loss functions with clear physical or statistical significance ( , , and This is achieved by weighted summation. During the model training phase, optimization algorithms such as gradient descent are used to minimize this summation. With the sole objective of finding the coefficients of the spline function on all edges of the network... Iterative updates will be performed.

[0097] Through the complete process of this implementation method, a fully optimized host power prediction model can be obtained. The feasibility and superiority of this scheme are reflected in its inherent design logic: First, the entire process, from data input and network construction to loss function construction and optimization, is based on a defined mathematical model and standard deep learning paradigm, making it technically feasible; second, its superior technical effect stems from the synergistic effect of structure and constraints. By integrating four loss functions, the model is forced to find the optimal solution within a feasible domain jointly defined by data patterns and physical laws, thereby fundamentally solving the problem of the lack of physical consistency in purely data-driven models.

[0098] Compared to the traditional PINN network that uses a fixed activation function, the KAN network architecture employed in this embodiment, with its learnable activation function composed of orthogonal basis functions and possessing infinite approximation capabilities, provides the necessary model capacity and flexibility for finding high-quality solutions that simultaneously satisfy four complex constraints. This enables the invention to achieve higher prediction accuracy while ensuring physical reliability, thus providing a more advanced, accurate, and reliable technical solution for real-time energy efficiency management of ships.

[0099] In a preferred embodiment, the basis functions used to construct the learnable activation function can be of other types. Chebyshev polynomials and Legendre polynomials are listed in the foregoing preferred embodiments, but any other function system with good function approximation capabilities is equally applicable. For example, B-spline basis functions, due to their local support properties and good numerical stability, are widely used in computer-aided design and data fitting, and can also construct arbitrarily smooth function shapes by learning their control points. Using Fourier series as basis functions, that is, representing the activation function as a linear combination of a set of sine and cosine functions, is particularly suitable for processing input data with periodic or oscillatory characteristics.

[0100] In the preferred embodiment, although the KAN network is the preferred architecture of the present invention, other advanced network structures that can effectively handle high-dimensional nonlinear mapping relationships and can be combined with physical constraints can also be used as alternatives. For example, a Transformer network architecture based on an attention mechanism can be used. Through its self-attention module, the Transformer model can effectively capture the complex dependencies between input features. The various input parameters of the ship are input as a sequence into the Transformer encoder, which outputs a high-dimensional feature representation, and then passes it through a fully connected layer to obtain the final power prediction value. The aforementioned composite loss function It can still be fully applied to the final output of the model, thereby guiding the learning process of the Transformer network and achieving a prediction model that combines high accuracy and physical consistency.

[0101] In a preferred embodiment, the aforementioned dynamic constraints The model used here is a cubic polynomial, which is an effective simplification. In alternatives requiring higher accuracy, this physical model can be replaced with a more refined theoretical model of ship resistance and propulsion, such as theoretical power curves calculated using the Holtrop-Mennen or Taylor methods, or a high-dimensional interpolation table constructed from pre-calculated power data under a series of standard operating conditions obtained through high-fidelity CFD simulations. The deviation between the network's predicted values ​​and the interpolation table results is then calculated. This allows for the injection of more precise physical prior knowledge into the model, potentially leading to higher prediction accuracy. Similarly, in calculating the loss, in addition to the mean squared error loss (MSE), the mean absolute error (MAE), which is insensitive to outliers, or Huber loss, which combines the advantages of both, can be used as alternatives.

[0102] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for real-time optimization of ship main engine power based on a hybrid network model, characterized in that, Includes the following steps: S1. Construct a hybrid neural network model, in which a KAN network is introduced as the front-end network structure to establish a high-dimensional nonlinear mapping from speed, ship type parameter matrix, propulsion efficiency, fuel conversion efficiency and environmental factors to the minimum power of the main engine. A three-layer KAN network structure is constructed. The input layer receives a set of preprocessed and normalized multidimensional feature vectors. The multidimensional feature vectors contain at least the ship's real-time speed v, the ship's hull parameter matrix Q, and the propulsion efficiency η. p Fuel conversion efficiency η f And the comprehensive environmental factor R; The learnable activation function Φ(x) for each edge in the KAN network structure is: Φ(x)=∑c i ·y i (x), Where c i y is the coefficient that the network needs to learn. i (x) is the i-th basis function; S2. The hybrid neural network model is trained using a composite loss function, which is composed of a weighted sum of mean square error loss, dynamic constraint loss, propulsion efficiency constraint loss and fuel consumption constraint loss. S3. Using the trained hybrid neural network model, the ship's operating parameters and environmental parameters are collected in real time. Through one forward propagation calculation, the minimum power prediction value of the ship's main engine is output, thereby realizing the real-time optimization of the ship's main engine power.

2. The method for real-time optimization of ship main engine power based on a hybrid network model according to claim 1, characterized in that, In step S2, the mean squared error loss function Loss data By calculating the deviation between the predicted main engine power and the actual main engine power, we ensure that the predicted main engine power accurately fits the historical dataset collected from the ship's actual voyages. In the formula, P Pred (v i P represents the host power predicted by the hybrid neural network model for the i-th input sample; true (v i ) represents the actual main engine power recorded during the ship's actual navigation; v i Let be the i-th input sample representing the ship's speed; N is the total number of samples.

3. The real-time optimization method for ship main engine power based on a hybrid network model according to claim 2, characterized in that, Constructing the dynamic constraint loss function Loss dyn : Dynamic constraint loss function Loss dyn By calculating the predicted host power P Pred (v i ) and the theoretical power P of the physical model dyn At the corresponding speed v i The sum of squared differences quantifies the degree to which the predicted behavior deviates from fundamental physical laws. During training, it is used in conjunction with the mean squared error loss function to achieve loss reduction. dyn minimize.

4. The real-time optimization method for ship main engine power based on a hybrid network model according to claim 3, characterized in that, Constructing a propulsion efficiency constraint loss function Loss η : In the formula, η p (v i ( ) represents the speed v i The propulsion efficiency of a ship, P Pred (v i )·η p (v i ) represents the predicted host power P Pred and the known propulsion efficiency η p (v i The theoretical effective power, P, calculated from this. useful (v i ) is the true reference effective power.

5. The method for real-time optimization of ship main engine power based on a hybrid network model according to claim 4, characterized in that, Constructing a fuel consumption-constrained loss function Loss fuel : In the formula, the actual fuel consumption under the i-th sample operating condition is: The known fuel conversion efficiency is η f .

6. The method for real-time optimization of ship main engine power based on a hybrid network model according to claim 5, characterized in that, By introducing weight coefficients W1, W2, W3, and W4 and performing a linear weighted sum, the final total loss function Loss is constructed. total : Loss total =W1·Loss data +W2·Loss dyn +W3·Loss η +W4·Loss fuel ; By minimizing the total loss function Loss total Find a set that can make Loss total The internal parameters of the KAN network that reach the minimum value.

7. The method for real-time optimization of ship main engine power based on a hybrid network model according to claim 2, characterized in that, The dynamic constraint loss function is based on a semi-empirical cubic polynomial model describing the power-speed relationship of a ship in still water. It embeds prior knowledge of ship hydrodynamics into the network model, forcing the predicted main engine power P output by the KAN network to be determined. pred In terms of function form, it follows the physical relationship between the ship's main engine power and speed: The dyn =α1·v 3 +α2·v 2 +α3·v+α4+α5L 2 In the formula, P dyn α1 is the calculated theoretical main engine power; v is the ship speed; L is the ship length; α1 is the coefficient of the cubic term of speed, α2 is the coefficient of the square term of speed, α3 is the coefficient of the linear term of speed, α4 is the constant term, and α5 is the coefficient of the linear term of speed. 2 The drag term is a term that characterizes the resistance related to the ship's geometry.

8. The method for real-time optimization of ship main engine power based on a hybrid network model according to claim 4, characterized in that, Effective power P useful for: P.S useful Hη p ·P In the formula, P represents the ship's total power requirement; η represents the ship's propulsion efficiency. p for: In the formula, v opt η represents the ship's optimal speed. max denoted as the maximum ship propulsion efficiency; a is a constant related to the propulsion system characteristics.

9. The method for real-time optimization of ship main engine power based on a hybrid network model according to claim 5, characterized in that, The ship's total power demand P and the amount of fuel consumed per unit time The relationship between them is: In the formula, η f This refers to the fuel conversion efficiency of the main unit.

Images