A ship main engine fuel oil consumption rate prediction method based on a twin network

Abstract

The application relates to a ship main engine fuel oil consumption rate prediction method based on a twin network, which comprises the following steps: acquiring actual normal ship cruising working condition real ship monitoring data set, inputting the actual data set into a trained oil consumption rate prediction model to obtain a predicted oil consumption rate, wherein the training process of the oil consumption rate prediction model comprises the following steps: S1, feature composition data set; S2, obtaining an LSTM network with an initial value; S3, training a combined network; S4, copying the combined network of S3 to form a twin network; S5, labeling a new training set according to the similarities and differences between ships and routes; S6, training the distance between the two outputs of the combined network and the similarity between the labeled training set by using a stochastic gradient descent method to obtain a trained oil consumption rate prediction model. Compared with the prior art, the application has the advantages of adapting to complex sea conditions and weather conditions in actual ocean ship navigation.

Description

Technical Field

[0001] This invention relates to the technical field of ship fuel consumption prediction, and in particular to a method for predicting ship main engine fuel consumption rate based on twin networks. Background Technology

[0002] Ships are considered the world's primary mode of cargo transportation, handling over 90% of global cargo. During transport, ships consume fuel and emit air pollutants, including carbon dioxide and other greenhouse gases. To address global climate change, both the International Maritime Organization (IMO) and domestic maritime authorities have placed great emphasis on reducing greenhouse gas emissions from ships, and have formulated corresponding requirements and documents. Reducing greenhouse gas emissions from ships is not only a matter of global climate change response, but also a matter of national welfare and people's livelihoods, and is crucial to the future development of a strong transportation nation, a strong maritime nation, and the maritime industry.

[0003] Greenhouse gas emissions from shipping are closely related to fuel consumption, making research on ship fuel consumption data extremely important. In recent years, establishing ship fuel consumption prediction models based on ship fuel consumption data has become a common approach for accurately assessing greenhouse gas emissions from shipping. Research methods for ship fuel consumption models can be broadly categorized into three types: the first is a white-box model based on analyzing the relationships between various physical quantities in the ship's main and auxiliary engines; however, this model is only effective for analyzing internal physical quantity relationships during the ship design phase. The second is a black-box model that directly models data relationships, but requires setting numerous structural parameters for the ANN network. The third is a gray-box model that combines both approaches, primarily employing regression analysis methods for ship fuel consumption-related factors under experimental conditions, and lacks universality. These existing ship fuel consumption modeling methods generally suffer from structural risks in model representation learning, failing to effectively incorporate various features into the model's representation structure. Such models lack generalization ability and are at risk of collapse during prediction. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method for predicting the fuel consumption rate of ship main engines based on twin networks.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A method for predicting ship main engine fuel consumption rate based on twin networks includes the following steps:

[0007] Obtain actual ship monitoring datasets under normal ship cruising conditions, input the actual datasets into a trained fuel consumption rate prediction model to obtain the predicted fuel consumption rate. The training process of the fuel consumption rate prediction model includes:

[0008] S1. Obtain data from actual ship monitoring under normal ship patrol conditions, extract features from the data, and form a dataset from the features;

[0009] S2. Construct a convolutional neural network by sampling the dataset to obtain sequences, which form a training set. Pre-train the convolutional neural network based on the training set to obtain an LSTM network with initial values.

[0010] S3. After the LSTM network with initial values, add a multi-layer neural network. Input the output of the LSTM network with initial values ​​into the multi-layer neural network for training to obtain a trained combined network. The combined network includes the LSTM network and the multi-layer neural network.

[0011] S4. Copy the combined network of S3, connect the two combined networks in parallel to form a twin network;

[0012] S5. Shuffle the sequence of the training set in S2, and form a new training set by extracting two sequences from the training set each time. Label the new training set according to the similarities and differences between ships and routes.

[0013] S6. Input the labeled training set from S5 into the Siamese network. The labeled training set is divided into two parts and input into the combined network respectively. Use stochastic gradient descent to train the distance between the two outputs of the combined network and the similarity between the labeled training set. Optimize the parameters of the combined network to obtain the trained fuel consumption rate prediction model.

[0014] Furthermore, the labeled training set is divided into two parts and input into the combined network. The combined network outputs y1 and y2 respectively. The distance between the outputs y1 and y2 of the combined network is calculated using a contrastive loss function. The expression for the distance between the outputs y1 and y2 of the combined network is as follows:

[0015] L w (y1, y2) = ||y1 - y2|| = d × L + (y1,y2)+(1-d)×L - (y1, y2)

[0016] Where d is a flag indicating whether the two inputs belong to the same category; d = 1 when they belong to the same category, and d = 0 when they belong to different categories. + Let L be the loss function when the two inputs are of the same class. - The loss function when the two inputs are of different categories;

[0017]

[0018]

[0019] Ew is the distance between the two inputs, and m is the threshold.

[0020] Furthermore, the rules for labeling the training set are as follows: two sequences taken from the same ship and the same route are labeled as 1, two sequences taken from the same ship but different routes are labeled as t, t∈(0,1), and two sequences taken from different ships are labeled as 0.

[0021] Furthermore, t takes the value of 0.7.

[0022] Furthermore, the specific process of S2 pre-training is as follows:

[0023] The training set is input into the constructed convolutional neural network. The training set passes through a fully connected hidden layer in sequence. The mean squared error loss function is calculated. The parameters of the convolutional neural network are optimized based on the mean squared error loss function to obtain an LSTM network with initial values.

[0024] Furthermore, the activation function used in the pre-training is the Sigmoid activation function.

[0025] Furthermore, the output of the LSTM network is input into a multi-layer neural network for training. The multi-layer neural network includes two hidden layers, and the data passes through the two hidden layers sequentially. The mean squared error loss function is calculated, and the parameters of the multi-layer neural network are optimized based on the mean squared error loss function to obtain the trained combined network.

[0026] Furthermore, when training a multi-layer neural network, the activation function used is the Sigmoid activation function.

[0027] Furthermore, the features extracted in S1 specifically include fuel consumption per nautical mile of main engine, main engine shaft power, ship speed, standard deviation of ship speed, slip ratio, wind speed, wave height, ship trim, ship list, and ship loading status.

[0028] Furthermore, S2 samples at 30-second intervals, resulting in a sequence step size of 10.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] (1) Based on the judgment of the similarities and differences between ships and routes, the present invention makes judgment on the data monitored by the actual ship, and transforms the prediction problem of real-time fuel consumption of ocean-going ships into the measurement problem of fuel consumption difference between different ships and different routes. This makes the representation structure of the model diverse, avoids collapse, and can adapt to the complex sea conditions and weather conditions in the actual navigation of ocean-going ships, as well as human factors such as ship transshipment and driving.

[0031] (2) In the analysis of ship fuel consumption data, the present invention samples some data for pre-training and then constructs a twin network to realize an end-to-end representation method for real-time energy consumption prediction model of ocean-going ships. Attached Figure Description

[0032] Figure 1 This is a flowchart of the training model of the present invention;

[0033] Figure 2 This is a schematic diagram of the structure of a twin network. Detailed Implementation

[0034] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0035] This invention proposes a method for predicting the fuel consumption rate of ship main engines based on twin networks, which includes the following steps: obtaining a real ship monitoring dataset under actual normal ship cruising conditions, inputting the actual dataset into a trained fuel consumption rate prediction model, and obtaining the predicted fuel consumption rate.

[0036] The flowchart for training the fuel consumption rate prediction model is as follows: Figure 1 As shown. The training process includes the following steps:

[0037] S1. Obtain data from actual ship monitoring under normal ship patrol conditions, extract features from the data, and form a dataset from the features;

[0038] S2. Construct a convolutional neural network by sampling the dataset to obtain sequences, which form a training set. Pre-train the convolutional neural network based on the training set to obtain an LSTM network with initial values.

[0039] S3. Add a multi-layer neural network after the LSTM network with initial values. Input the output of the LSTM network with initial values ​​into the multi-layer neural network for training to obtain a trained combined network. The combined network includes the LSTM network and the multi-layer neural network.

[0040] S4. Copy the combined network of S3, connect the two combined networks in parallel to form a twin network;

[0041] S5. Shuffle the sequence of the training set in S2, and form a new training set by extracting two sequences from the training set each time. Label the new training set according to the similarities and differences between ships and routes.

[0042] S6. Input the labeled training set from S5 into the Siamese network. The labeled training set is divided into two parts and input into the combined network respectively. Use stochastic gradient descent to train the distance between the two outputs of the combined network and the similarity between the labeled training set. Optimize the parameters of the combined network to obtain the trained fuel consumption rate prediction model.

[0043] The data from the actual ship monitoring in S1 includes data collected from the ship's liquid level system (LLS), flow meters, GPS, anemometers, Doppler velocimeters, shaft power monitors, etc. Correlation analysis was performed on the ship's fuel consumption and the measured data. Through feature extraction, 10 features were identified as feature terms for the ship's main engine fuel consumption rate (FOC) prediction model: main engine fuel consumption per nautical mile, main engine shaft power, ship speed (speed relative to water), standard deviation of ship speed, slip ratio, wind speed, wave height, ship trim (difference between bow and stern draft), ship heel (difference between port and starboard draft), and ship loading status (ballast or full load).

[0044] In the pre-training of S2, a convolutional neural network is constructed. The dataset is sampled at 30-second intervals, and sequences with a stride of 10 (i.e., a sequence period of 5 minutes) are generated as the training dataset to train an LSTM network, giving the LSTM network initial values. The input data x∈R of the convolutional neural network... B×l×r Where B is the batch size, l is the selected fixed sequence step size (here, 10), and r is the feature vector length (here, 10). After the input data enters the LSTM network, it passes through a fully connected hidden layer. A fully connected layer means that each input neural network node is connected to every output neural network node. Its function is to preprocess the input data, allowing the LSTM network to obtain more comprehensive and holistic information.

[0045] The output of the fully connected layer is h∈R B×l×1 Where B is the batch size, l is the selected fixed sequence step size (here it is 10), and the output h is... i (i∈[1,10]) is the predicted value of the ship's main engine fuel consumption rate (FOC) at step i.

[0046] In the pre-training of S2, the activation function used is Sigmoid:

[0047]

[0048] Its derivative is:

[0049] σ′(x)=σ(x)1-σ(x))

[0050] The loss function used is the Mean Square Error Loss (MSE Loss Function):

[0051] loss(x,y)=1 / n∑(x i -y i ) 2

[0052] This loss function penalizes larger errors more heavily and smaller errors less heavily. It is quite sensitive to outliers and is significantly affected by them.

[0053] In S3, the pre-trained LSTM network is adjusted by adding a multilayer neural network (MLP) after the LSTM network. The input of the MLP layer is the output of the LSTM network, h∈R. l×1 Where l is the selected fixed sequence step size (here, 10). The output of the MLP layer is y∈R 1×1 This refers to the predicted fuel consumption rate (FOC) of the ship's main engine. The MLP layer consists of multiple hidden layers; typically, the number of hidden layers is greater than one, but here it is two. The number of neurons in each hidden layer decreases from largest to smallest; here, the first hidden layer has 20 neurons, and the second hidden layer has 5 neurons. The MLP uses the Sigmoid activation function and the MSE Loss function. This forms an LSTM+MLP network.

[0054] In S4, a twin network is constructed: the LSTM+MLP network obtained in S3 is copied, and then the two LSTM+MLP networks are connected in parallel to form a twin network.

[0055] In S5, the sequences of the training set from S2 are shuffled, and two sequences are extracted each time to form a new training set sample. The samples in the new training set are labeled according to the differences between ships and routes. For example, as shown in Table 1:

[0056] Table 1 shows the annotations for the new training set based on differences in ships and routes.

[0057]

[0058]

[0059] In S6, the Siamese network from S4 is trained using the labeled training set from S5, and the training samples x∈R are input into the Siamese network. B ×2×l×r The training samples are divided into two parts and fed separately into a Siamese network, which is the trained combined network, to obtain outputs y1 and y2. A contrastive loss function is used to measure the distance between outputs y1 and y2. The loss function is solved using stochastic gradient descent.

[0060] The input to the contrastive loss function is<y1,y2,d> Then the function is divided into y1 and y2 similar (d≥t, L + ), y1 and y2 are not similar (d = 0, L - There are two scenarios.

[0061] but:

[0062] L w (y1, y2) = ||y1 - y2|| = d × L + (y1,y2)+(1d)×L - (y1, y2)

[0063] in,

[0064]

[0065]

[0066] Here, the criterion for determining whether two input datasets come from the same ship is used, assuming 1 for same ship and voyage, and 0 for different ships and voyages. After processing, if the value is close to 1, the two input samples are determined to be from the same ship; if the value is close to 0, the two input samples are determined to be from different ships. Comparing with the new training dataset labeled S5, the similarity is trained using stochastic gradient descent to optimize the parameters in the LSTM+MLP network, resulting in the trained fuel consumption rate prediction model. The structure diagram of the trained fuel consumption rate prediction model (LSTM+MLP network) is shown below. Figure 2 As shown.

[0067] Using the trained model described above, a real-world monitoring dataset under normal ship cruising conditions is obtained. This dataset is then input into the trained fuel consumption rate prediction model to obtain the predicted fuel consumption rate. The input data x∈R of the fuel consumption rate prediction model. l×r Where l is the selected fixed sequence step size (here, 10). h is the output of the MLP layer, which passes through two hidden layers, and the output is y∈R. 1×1 This refers to the predicted value of the ship's main engine fuel consumption rate (FOC).

[0068] This invention provides a method for analyzing ship fuel consumption data based on Siamese networks. From a metric learning perspective, this invention transforms the problem of predicting real-time fuel consumption of ocean-going vessels into a problem of measuring the fuel consumption differences between different vessels on different routes. This results in a diverse model representation structure, avoids collapse, and can adapt to complex sea conditions and weather conditions, as well as human factors such as reloading and navigation, encountered during actual ocean-going vessel navigation. In analyzing ship fuel consumption data, this invention allows for pre-training of models based on some key influencing factors, and then constructs a Siamese network to achieve an end-to-end representation method for real-time energy consumption prediction models of ocean-going vessels. The Siamese network enhances the generalization ability of the prediction model.

[0069] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A twin network-based ship main engine fuel oil consumption rate prediction method, characterized in that, The method comprises the following steps: An actual normal ship cruising condition monitoring data set is obtained, and the actual data set is input into a trained oil consumption rate prediction model to obtain a predicted oil consumption rate, wherein the training process of the oil consumption rate prediction model comprises: S1, obtaining ship monitoring data under normal ship cruising conditions, extracting features from the data, and forming a data set from the features; S2, constructing a convolutional neural network, sampling the data set to obtain a sequence, forming a training set from the sequence, pre-training the convolutional neural network based on the training set, and obtaining an LSTM network with initial values; S3, adding a multi-layer neural network to the LSTM network with initial values, inputting the output of the LSTM network with initial values into the multi-layer neural network for training, and obtaining a trained combined network, wherein the combined network comprises the LSTM network and the multi-layer neural network; S4, copying the combined network of S3, connecting the two combined networks in parallel to form a twin network; S5, shuffling the sequence of the training set of S2, extracting two sequences from the training set each time to form a new training set, and labeling the new training set according to the similarities and differences between ships and routes; S6, inputting the labeled training set of S5 into the twin network, dividing the labeled training set into two parts, inputting them into the combined network respectively, training the distance between the two outputs of the combined network and the similarity between the labeled training set, optimizing the parameters of the combined network, and obtaining a trained oil consumption rate prediction model; After the labeled training set is divided into two parts and input into the combined network, the combined network outputs y1 and y2 respectively, and the distance between the outputs y1 and y2 of the combined network is measured and calculated using a contrast loss function, wherein the expression of the distance between the outputs y1 and y2 of the combined network is: wherein, d is a flag for whether the two inputs are of the same class, when the same d = 1, and different classes d = 0, is a loss function when the two inputs are of the same class, is a loss function when the two inputs are of different classes; is a distance between two inputs, is a threshold value.

2. The twin network based marine main engine fuel oil consumption rate prediction method according to claim 1, characterized in that, The labeling rule for the training set is that two sequences taken from the same ship and the same route are labeled as 1, two sequences taken from the same ship and different routes are labeled as t, t∈(0,1), and two sequences taken from different ships are labeled as 0.

3. The twin network based marine main engine fuel oil consumption rate prediction method according to claim 2, characterized in that, The value of t is 0.

7.

4. The twin network based marine main engine fuel oil consumption rate prediction method of claim 1, wherein, The specific process of pre-training in S2 is: The training set is input into the constructed convolutional neural network, the training set passes through a fully connected hidden layer one by one, the mean square error loss function is calculated, the parameters of the convolutional neural network are optimized based on the mean square error loss function, and an LSTM network with initial values is obtained.

5. The twin network based marine main engine fuel oil consumption rate prediction method according to claim 4, characterized in that, The activation function used in pre-training is a Sigmoid activation function.

6. The twin network based marine main engine fuel oil consumption rate prediction method of claim 1, wherein, The output of the LSTM network is input into the multi-layer neural network for training, the multi-layer neural network comprises two hidden layers, data passes through the two hidden layers one by one, the mean square error loss function is calculated, the parameters of the multi-layer neural network are optimized based on the mean square error loss function, and a trained combined network is obtained.

7. The twin network based marine main engine fuel oil consumption rate prediction method according to claim 6, characterized in that, The activation function used in training the multi-layer neural network is a Sigmoid activation function.

8. The twin network based marine main engine fuel oil consumption rate prediction method according to claim 1, characterized in that, The features extracted in S1 specifically include main engine fuel consumption per nautical mile, main engine shaft power, ship speed, standard deviation of ship speed, slip ratio, wind speed, wave height, ship trim, ship heeling, and ship loading condition.

9. The twin network based marine main engine fuel oil consumption rate prediction method according to claim 1, characterized in that, The sampling in S2 is at an interval of 30 seconds, and the obtained sequence has a step length of 10.

Images