Spatiotemporal prediction method for ship carbon emissions and carbon emission monitoring platform using the method

Abstract

This invention discloses a spatiotemporal prediction method for ship carbon emissions, comprising: calculating the carbon emissions of each ship in multiple trajectory segments; spatially mapping the carbon emissions of each ship in multiple trajectory segments onto a spatial grid to obtain carbon emission image data with time-series characteristics; dividing the processed carbon emission image data into multiple "input-output" sample pairs to obtain a dataset; training a carbon emission spatiotemporal prediction model based on the dataset to obtain a trained carbon emission spatiotemporal prediction model; inputting historical carbon emission image data before the period to be predicted into the trained carbon emission spatiotemporal prediction model for prediction to obtain a spatial distribution image of ship carbon emissions for the prediction period. The method of this invention can accurately predict the temporal and spatial distribution characteristics of global ship carbon emissions in the future. In addition, this invention also discloses a carbon emission monitoring platform using this method.

Description

Technical Field

[0001] This invention relates to the fields of green shipping and intelligent transportation, and in particular to a method for spatiotemporal prediction of ship carbon emissions and a carbon emission monitoring platform using this method. Background Technology

[0002] Shipping is a vital pillar of global trade, handling approximately 90% of cargo volume. However, pollutants emitted by ships, such as greenhouse gases, nitrogen oxides, and sulfur oxides, have a severe impact on the environment. As global attention to climate change continues to grow, the international community is imposing increasingly stringent environmental requirements on the shipping industry.

[0003] Currently, research on ship carbon emissions mainly focuses on carbon emission measurement or total statistics. The commonly used carbon emission measurement method at home and abroad is mainly the "statistical method." This method uses statistical data on ship activities to determine the ship emission factors based on different ship types, engine types, and navigation states. It combines the average activity time and the average load factor under different navigation states, and estimates the ship's carbon emissions based on the ship's energy consumption factor, power model, or typical operating conditions.

[0004] While such methods have some application value in estimating total emissions, they rely excessively on static data and employ simplistic computational models. They can only estimate ship carbon emissions or output temporal trends, lacking in-depth research into the dynamic changes of emissions over time and space. This leads to problems such as inaccurate calculation results, low spatiotemporal resolution, and weak dynamic perception capabilities. They fail to reflect the spatial distribution characteristics and trends of carbon emissions during actual operation, especially for container ships—ships that navigate frequently globally, have complex routes, and high emission intensity. Furthermore, these methods lack the ability to accurately capture and predict the spatial distribution and future trends of carbon emissions at the regional scale, thus failing to provide future emission trend assessments and predictions of hotspot evolution. Summary of the Invention

[0005] To address the problems in the prior art, this invention proposes a spatiotemporal prediction method for ship carbon emissions. This method can accurately predict the temporal trend and spatial distribution characteristics of global ship carbon emissions over a future period, providing key technical support for building an intelligent and forward-looking shipping carbon emission regulatory system.

[0006] To achieve the above objectives, the present invention provides a method for spatiotemporal prediction of ship carbon emissions, comprising the following steps:

[0007] (1) Based on the static data of the ship, construct a calculation model for the carbon emissions of the ship, and based on the calculation model, calculate the carbon emissions of each ship in multiple trajectory segments, wherein the trajectory segment is the trajectory formed by connecting two adjacent AIS trajectory points;

[0008] (2) The carbon emissions of each ship in multiple trajectory segments are spatially mapped into a spatial grid to obtain the global carbon emissions. The global carbon emissions are obtained according to a set time interval and then visualized after being arranged in time to obtain carbon emission image data with time series characteristics.

[0009] (3) Normalize and grayscale the carbon emission image data, and divide the processed carbon emission image data into multiple "input-output" sample pairs to obtain a dataset;

[0010] (4) Based on the improved ConvLSTM network structure, a carbon emission spatiotemporal prediction model is constructed, and the carbon emission spatiotemporal prediction model is trained and tested based on the dataset to obtain a trained carbon emission spatiotemporal prediction model.

[0011] (5) Obtain historical carbon emission image data with time series characteristics before the period to be predicted, and after normalization and grayscale processing, input it into the trained carbon emission spatiotemporal prediction model for prediction to obtain the spatial distribution image of ship carbon emissions during the prediction period.

[0012] On the other hand, the present invention also provides a carbon emission monitoring platform, including an external database, a carbon emission spatiotemporal prediction module, a retrieval enhancement generation module, a large model interface module, a large model module, a hotspot tracking and early warning module, a carbon emission spatiotemporal visualization module, a green route recommendation module, an intelligent question answering and assistant analysis module, a port low-carbon scheduling module, and a carbon emission reporting and trend analysis module.

[0013] The external database is used to store information related to the shipping industry; the carbon emission spatiotemporal prediction module is used to calculate and predict ship carbon emissions using the aforementioned ship carbon emission spatiotemporal prediction method, and provides carbon emission calculation results and carbon emission prediction results; the retrieval enhancement generation module is connected to the external database and the carbon emission spatiotemporal prediction module, and is used to obtain information related to the shipping industry, carbon emission calculation results, and carbon emission prediction results from the external database and the carbon emission spatiotemporal prediction module respectively according to business requests, and construct prompt information corresponding to the business requests based on the obtained information and results; the large model interface module is connected to the retrieval enhancement generation module, the large model module, and the green route recommendation module. The intelligent question-and-answer and assistant analysis module, the port low-carbon scheduling module, and the carbon emission report and trend analysis module are connected and used to forward the business requests obtained from the green route recommendation module, the intelligent question-and-answer and assistant analysis module, the port low-carbon scheduling module, and the carbon emission report and trend analysis module to the retrieval enhancement generation module. The module encapsulates the prompt information obtained from the retrieval enhancement generation module into text or table information corresponding to the business request that can be directly parsed and used by the large model and sends it to the large model module. The module also sends the answers or analyses corresponding to the business request obtained from the large model module to the green route recommendation module, the intelligent question-and-answer and assistant analysis module, and the port low-carbon scheduling module, respectively. The carbon scheduling module and the carbon emission reporting and trend analysis module are displayed to the user; the large model module is used to generate corresponding answers or analyses from the text or table information corresponding to the business request obtained from the large model interface module, and send the answers or analyses to the large model interface module; the hotspot tracking and early warning module is connected to the carbon emission spatiotemporal prediction module, and is used to obtain the carbon emission prediction results of high emission hotspot areas in real time from the carbon emission spatiotemporal prediction module, compare the carbon emission prediction results with a preset threshold, and issue an early warning notification when the preset threshold is exceeded; the carbon emission spatiotemporal visualization module is connected to the carbon emission spatiotemporal prediction module, and is used to obtain the carbon emission prediction results in real time from the carbon emission spatiotemporal prediction module. The system displays emission prediction results and carbon emission calculation results to users in an interactive and visual manner. The green route recommendation module sends route planning requests to the large model interface module and displays the corresponding answers or analyses obtained from the large model interface module to the user. The intelligent question-and-answer and assistant analysis module sends question requests to the large model interface module and displays the corresponding answers or analyses obtained from the large model interface module to the user. The port low-carbon scheduling module sends low-carbon scheduling requests to the large model interface module and displays the corresponding answers or analyses obtained from the large model interface module to the user.The carbon emission reporting and trend analysis module sends a business request to the large model interface module to generate a carbon emission report and trend analysis, and displays the corresponding response or analysis obtained from the large model interface module to the user.

[0014] The beneficial effects of the technical solution provided by this invention are:

[0015] 1. This invention fully utilizes a large amount of AIS trajectory data of ships and combines it with static data of ships to construct a carbon emission calculation module based on the STEAM model. It can accurately calculate the carbon emissions of ships at different times and in different sea areas, which is conducive to realizing continuous spatial modeling of carbon emissions and prediction of future distribution.

[0016] 2. Based on the spatial distribution characteristics of carbon emissions, this invention constructs visualized spatial gridded carbon emission image data by mapping and visualizing the spatial grid of carbon emissions, thereby visually reflecting the spatial distribution characteristics of carbon emissions and revealing the spatial aggregation characteristics and changing trends of carbon emissions in regional sea areas.

[0017] 3. This invention enhances the stability of carbon emission image data distribution through normalization processing, thereby improving the stability of carbon emission prediction;

[0018] 4. This invention constructs a spatiotemporal carbon emission prediction model based on a deep learning network structure with spatiotemporal feature extraction capabilities. This model can uncover the spatial clustering and temporal evolution patterns of carbon emissions. Using this model for prediction, on the one hand, it can continuously and accurately predict the temporal trend and spatial distribution characteristics of global ship carbon emissions over a future period, reflecting the future spatial distribution characteristics of carbon emissions during actual operation, and providing scientific support for building an intelligent and forward-looking shipping carbon emission regulatory system. On the other hand, based on the prediction results, it can analyze the spatial trend and differences in carbon emissions, effectively identifying the spatial distribution characteristics of carbon emission hotspots, which is beneficial for carbon emission hotspot monitoring and trend assessment. Furthermore, based on this model, using large-scale, multi-time carbon emission data for inference can significantly improve the accuracy, timeliness, spatial resolution, and stability of ship carbon emission prediction. Attached Figure Description

[0019] Figure 1 A flowchart of the spatiotemporal prediction method for ship carbon emissions provided in an embodiment of the present invention;

[0020] Figure 2 for Figure 1 Flowchart of step S3;

[0021] Figure 3 for Figure 1 Flowchart of step S4;

[0022] Figure 4 A schematic diagram of the improved ConvLSTM network structure provided in an embodiment of the present invention;

[0023] Figure 5 This is a schematic diagram comparing the predicted results and the actual results obtained by using the spatiotemporal prediction method for ship carbon emissions provided in this embodiment of the invention.

[0024] Figure 6 for Figure 1 Flowchart of step S7;

[0025] Figure 7 This is a schematic diagram of the carbon emission monitoring platform provided in an embodiment of the present invention. Detailed Implementation

[0026] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0027] like Figure 1 As shown in the figure, this invention provides a method for spatiotemporal prediction of ship carbon emissions, including the following steps:

[0028] S1: For each ship's abnormal AIS trajectory points, interpolation reconstruction is performed on the longitude and latitude. The interpolation reconstruction formula is as follows:

[0029] (1-1)

[0030] (1-2)

[0031] In the formula, The longitude of the reconstructed anomalous AIS trajectory points. The longitude of the preceding normal AIS trajectory point for the abnormal AIS trajectory point. The longitude of the next normal AIS trajectory point after the abnormal AIS trajectory point. The latitude of the reconstructed anomalous AIS trajectory points. This represents the latitude of the preceding normal AIS trajectory point for the abnormal AIS trajectory point. This refers to the latitude of the next normal AIS trajectory point following the abnormal AIS trajectory point. The time when the abnormal AIS trajectory point occurred. The reporting time of the previous normal AIS trajectory point for the abnormal AIS trajectory point. The reporting time of the next normal AIS trajectory point after the abnormal AIS trajectory point. The previous normal AIS trajectory point and the next normal AIS trajectory point are located before and after the abnormal AIS trajectory point, respectively, and are both adjacent to the abnormal AIS trajectory point.

[0032] Understandably, when an anomaly is detected in the latitude and longitude of a point in an AIS trajectory, the latitude and longitude of the adjacent normal AIS trajectory points can be used for interpolation reconstruction to clean up the latitude and longitude of the abnormal AIS trajectory point. This can ensure the accuracy and reliability of AIS trajectory data, thereby improving the accuracy of subsequent calculations of ship carbon emissions and the prediction results of spatiotemporal carbon emission prediction models.

[0033] Preferably, the ship speed of abnormal AIS trajectory points can also be reconstructed by interpolation according to the following formula:

[0034]

[0035] In the formula, The ship speed at the reconstructed abnormal AIS trajectory points. The ship's speed at the preceding normal AIS track point adjacent to the abnormal AIS track point. The ship's speed at the next normal AIS track point adjacent to the abnormal AIS track point. The time when the abnormal AIS trajectory point occurred. The reporting time of the previous normal AIS trajectory point for the abnormal AIS trajectory point. The reporting time of the next normal AIS trajectory point after the abnormal AIS trajectory point;

[0036] Preferably, the ship's course can also be reconstructed by interpolation for abnormal AIS trajectory points, which includes the following steps:

[0037] (i) Select the two closest normal AIS trajectory points before and after the abnormal AIS trajectory point;

[0038] (ii) Calculate the original heading difference between two normal AIS trajectory points using the following formula:

[0039]

[0040] In the formula, The difference between the original headings of two normal AIS trajectory points. The trajectory direction of the normal AIS trajectory point located after the abnormal AIS trajectory point. The trajectory direction of the normal AIS trajectory point located before the abnormal AIS trajectory point;

[0041] (iii) Adjust to -180° to 180° to obtain the net heading change. The adjustment rule is as follows: If ,but ;like ,but ;

[0042] (iv) Calculate the ship's average turning rate using the following formula:

[0043]

[0044] In the formula, The average turning rate of the ship, The time interval between two normal AIS trajectory points. and These are the reporting times of two normal AIS trajectory points located before and after the abnormal AIS trajectory point, respectively.

[0045] (v) Calculate the reconstructed ship track direction according to the following formula:

[0046]

[0047] In the formula, The reconstructed abnormal AIS trajectory points indicate the ship's course direction. The track direction of the normal AIS track point located before the abnormal AIS track point. The average turning rate of the ship, The time when the abnormal AIS trajectory point occurred. The reporting time for normal AIS trajectory points that precede abnormal AIS trajectory points;

[0048] (vi) If the reconstructed ship track exceeds the range of 0° to 360°, then add or subtract 360° for correction.

[0049] Preferably, the ship's bow direction of abnormal AIS trajectory points can also be reconstructed by interpolation, which includes the following steps:

[0050] (i) Select the two most recent valid trajectory points before and after the abnormal AIS trajectory point;

[0051] (ii) Calculate the original heading difference between the two valid trajectory points according to the following formula:

[0052]

[0053] In the formula, The original heading difference between two valid trajectory points. The bow direction of the valid trajectory point located after the abnormal AIS trajectory point. The bow direction of the valid trajectory point located before the abnormal AIS trajectory point;

[0054] (iii) Adjust to -180° to 180° to obtain the net heading change. The adjustment rule is as follows: If ,but ;like ,but ;

[0055] (iv) Calculate the ship's average rotational speed using the following formula:

[0056]

[0057] In the formula, The average rotational speed of the ship, and These are the reporting times of valid trajectory points located before and after the abnormal AIS trajectory point, respectively.

[0058] (v) Calculate the reconstructed ship's heading according to the following formula:

[0059]

[0060] In the formula, The ship's bow direction for the reconstructed abnormal AIS trajectory points. The bow direction of the valid trajectory point located before the abnormal AIS trajectory point. The average rotational speed of the ship, The time when the abnormal AIS trajectory point occurred. The reporting time for valid trajectory points located before abnormal AIS trajectory points;

[0061] (vi) If the bow of the reconstructed ship exceeds the range of 0° to 360°, then add or subtract 360° for adjustment.

[0062] S2: Based on the static data of the ship, construct a calculation model for the ship's carbon emissions, and based on the calculation model, calculate the carbon emissions of each ship in multiple trajectory segments, where each trajectory segment is the trajectory formed by connecting two adjacent AIS trajectory points.

[0063] Specifically, in this embodiment, the ship is a container ship, and its carbon emissions mainly come from the main engine, auxiliary engines, and boiler. In step S2, the ship's static data includes the rated power of the main engine, the load factor of the main engine, the emission factor of the main engine, the rated power of the auxiliary engines, the load factor of the auxiliary engines, the emission factor of the auxiliary engines, the actual power of the boiler, and the emission factor of the boiler.

[0064] First, determine the power of the main unit, auxiliary units, and boiler.

[0065] The maximum container capacity is determined by the ship's beam, measured in Twenty-feet Equivalent Units (TEUs). The type of main engine for the container ship is then determined based on this maximum capacity, which in turn determines the engine's rated power. Specific classification standards for container ships are detailed in Table 1.

[0066]

[0067] The classification criteria for main engine types of container ships are shown in Table 2.

[0068]

[0069] The primary function of auxiliary machinery is to provide electrical support for the ship, and it operates while the ship is at sea or berthed in port. The boiler is mainly used to provide thermal support when the main engine load is low (typically below or equal to 20%) or when the main engine is shut down, and is shut off during normal navigation. The navigation modes of the container ship in this embodiment are as follows: when speed < 1 knot, the ship is moored; when speed < 3 knots, the ship is anchored; when speed < 8 knots, the ship is towed; when speed > 8 knots, the ship is self-propelled. The rated power of the auxiliary machinery and the actual power of the boiler are determined based on the ship's maximum container capacity and navigation status. Table 3 shows the rated power of the auxiliary machinery and the actual power of the boiler for different ship sizes and navigation statuses.

[0070]

[0071] Secondly, determine the load factors for the main engine, auxiliary engines, and boiler. The load factor for the main engine is calculated based on the actual sailing speed and the designed maximum sailing speed. For different sailing conditions, the load factors for the auxiliary engines can be divided into five categories, as shown in Table 4.

[0072]

[0073] Finally, the fuel types and emission factors for the main unit, auxiliary units, and boiler were determined separately, as shown in Table 5.

[0074]

[0075] In detail, after determining the ship's static data, a model for calculating the ship's carbon emissions is constructed based on the static data. The calculation formula for the model is as follows:

[0076]

[0077] In the formula, This represents the carbon emissions of a ship within a given trajectory segment. This refers to the carbon emissions of the main unit. Carbon emissions from auxiliary equipment. Carbon emissions from boilers;

[0078] Specifically, the formula for calculating the carbon emissions of the host computer is as follows:

[0079]

[0080]

[0081] In the formula, This refers to the carbon emissions of the main unit. This refers to the rated power of the host unit. This represents the load factor of the host computer. The sailing time of the ship under different conditions; The emission factor of the main unit, The actual speed of the ship. The maximum speed at which the ship is designed;

[0082] The formula for calculating the carbon emissions of auxiliary equipment is as follows:

[0083]

[0084] In the formula, Carbon emissions from auxiliary equipment; The rated power of the auxiliary machine; The load factor for auxiliary equipment. The sailing time of the ship under different conditions. The emission factor for auxiliary equipment;

[0085] The formula for calculating the carbon emissions of a boiler is as follows:

[0086]

[0087] In the formula, Carbon emissions from boilers. This refers to the actual power of the boiler. The sailing time of the ship under different conditions. The emission factor of the boiler.

[0088] It should be noted that each trajectory segment is the trajectory formed by connecting two adjacent AIS trajectory points, that is, the trajectory within the time interval of two adjacent AIS location reports.

[0089] S3: Spatially map the carbon emissions of each ship in multiple trajectory segments into a spatial grid to obtain the global carbon emissions. Obtain the global carbon emissions at set time intervals, arrange them in time, and visualize them to obtain carbon emission image data with time series characteristics.

[0090] In detail, such as Figure 2 As shown, step S3 includes the following steps:

[0091] S31: The global region is divided using the WGS84 coordinate system, with 360 cells in the longitude direction and 180 cells in the latitude direction, to construct a spatial grid with a resolution of 1°×1°.

[0092] By constructing a spatial grid system and using geographic information spatial mapping methods, the spatial location and attribution of ship carbon emissions can be realized. The carbon emissions of all grid units are combined to form the global carbon emissions, thus realizing the spatial explicit expression of global carbon emission data.

[0093] S32: Assign a number to each grid cell in the spatial grid, and the relationship between the number and the latitude and longitude of any point within the global region is as follows:

[0094] (2-1)

[0095] (2-2)

[0096] (2-3)

[0097] In the formula, For the row index of the grid cell, The latitude coordinates of any point within the global region. For the column index of the grid cell, For any location within the global region, the longitude coordinates are... This indicates a round-down operation. The grid cell number;

[0098] Specifically, the numbering proceeds from bottom to top and from left to right, starting with 1 and incrementing sequentially until all grid cells are numbered. The latitude range is from... 90° to 90°, covering a total of 180 grid rows, with a longitude range from The grid consists of 180° to 180°, 360 grid columns, and a total of 64,800 grid cells. Each grid cell is assigned a unique number, facilitating rapid location and indexing of any latitude and longitude point globally, enabling effective querying of global geographic data, and allowing spatial data to be stored and retrieved according to a specific structure.

[0099] S33: The carbon emissions of each ship in multiple trajectory segments are assigned to the corresponding grid cells in the spatial grid, and the carbon emissions of all ships in the same grid cell are accumulated to obtain the total carbon emissions of each grid cell. The total carbon emissions of all grid cells are combined to form the global carbon emissions. Specifically, based on the latitude and longitude coordinates of the endpoint of each trajectory segment, the grid cell number corresponding to the trajectory segment is calculated according to formulas (2-1), (2-2) and (2-3), and the carbon emissions of the trajectory segment are assigned to the grid cell with the number.

[0100] Specifically, the formula for cumulative carbon emissions from all ships within the same grid cell is as follows:

[0101]

[0102] In the above formula, For grid cells j The carbon emissions of ships, For navigation through grid cells j The number of all ships, For navigation through grid cells j The One ship, For the first Each ship in the grid unit j Carbon emissions.

[0103] S34: Obtain the global carbon emissions according to the set time interval, and arrange them according to time to obtain the global carbon emissions with time series characteristics;

[0104] Specifically, the time interval can be set according to actual needs; in this example, it can be set to 1 day.

[0105] S35: Visualize global carbon emissions with time-series characteristics to obtain carbon emission image data with time-series characteristics.

[0106] S4: Normalize and grayscale the carbon emission image data, and divide the processed carbon emission image data into multiple "input-output" sample pairs to obtain a dataset.

[0107] In detail, such as Figure 3 As shown, step S4 includes the following steps:

[0108] Step S41: Normalize the carbon emission image data using formula (3-1):

[0109] (3-1)

[0110] In the formula, For normalized carbon emission image data, The raw carbon emission image data, and These represent the maximum and minimum values ​​of the original carbon emission image data, respectively. To prevent the logarithmic zero error, a small constant of 10 is chosen. -6 ;

[0111] Step S42: Convert the normalized carbon emission image data into grayscale image data;

[0112] Step S43: Using a sliding window method, the grayscale image data is divided into multiple "input-output" sample pairs that meet the structural requirements of the deep learning model to obtain the dataset.

[0113] Understandably, using logarithmic normalization to process carbon emission image data can utilize the nonlinear characteristics of the logarithmic function to map the original data with large numerical ranges to a relatively smooth and concentrated interval. Moreover, while performing logarithmic transformation on the original data, maximum and minimum normalization processing is also performed, compressing the results to the [0,1] interval, thereby suppressing the influence of large values, enhancing numerical stability, improving the accuracy and efficiency of calculation, and improving the accuracy and convergence speed of the model.

[0114] S5: Based on the improved ConvLSTM network structure, construct a spatiotemporal prediction model for carbon emissions, and train and test the spatiotemporal prediction model for carbon emissions based on the dataset to obtain a trained spatiotemporal prediction model for carbon emissions.

[0115] Specifically, the improved ConvLSTM network structure includes a ConvLSTM module and a CBAM module. The CBAM module includes a channel attention submodule and a spatial attention submodule. The spatiotemporal feature map extracted by the ConvLSTM module is used as the input to the CBAM module, wherein:

[0116] The algorithm expression for the ConvLSTM module is as follows:

[0117] (4-1)

[0118] (4-2)

[0119] (4-3)

[0120] (4-4)

[0121] (4-5)

[0122] In the formula, , , , and These are the input gate, forget gate, output gate, current cell state, and current hidden state, respectively. This represents the Sigmoid activation function. This represents a two-dimensional convolution operation. The convolutional kernel weights are input to the gate. The input feature map at the current time. To hide the convolutional kernel weights to the gate, This is the hidden state from the previous moment. For input gate bias terms; , and These are the convolution kernel weights input to the forget gate, the convolution kernel weights hidden to the forget gate, and the forget gate bias term, respectively. , and These are the convolutional kernel weights from input to output gate, the convolutional kernel weights hidden to output gate, and the output gate bias term, respectively. This represents the cell state at the previous moment. The candidate cell states extracted via convolution. This represents the Hadamard element-wise product. tanh (·) is the hyperbolic tangent activation function.

[0123] More specifically, the input sequence is first subjected to time-by-time feature learning through an encoding module. For example... Figure 4 As shown, the encoding module consists of three cascaded ConvLSTM units, each using a 3×3 convolutional kernel. The number of output channels is set to 32, 32, and 64 respectively, and padding is used to ensure that the spatial size of the output feature map remains 360×180. Multiple ConvLSTM units progressively extract deep spatiotemporal features from the carbon emission data. The bottom ConvLSTM units focus on capturing fine-grained emission patterns in local regions, while the upper ConvLSTM units extract more global temporal evolution trends. Thanks to the convolutional gating mechanism, the ConvLSTM units can retain the spatial structure information of the image at each time step when updating the hidden state, giving the model a strong spatiotemporal feature extraction capability. After encoding by three ConvLSTM units, a high-dimensional feature representation (containing a 360×180 feature map with 64 channels) is obtained at the final time step. The ConvLSTM units use convolution operations in the gating calculation, allowing the model to retain the spatial structure information of the carbon emission image during state updates.

[0124] Furthermore, through convolutional gating, each ConvLSTM unit can fuse the hidden state from the previous time step with the input image features of the current time step, updating both temporal memory and spatial features at each time step, giving the model powerful spatiotemporal modeling capabilities. The stacking of multiple ConvLSTM units further enhances feature extraction; lower-level units can capture fine-grained local emission patterns, while upper-level units extract global temporal evolution features. After processing by the encoding module, the model outputs a high-dimensional spatiotemporal feature representation, providing rich information support for subsequent prediction stages.

[0125] The calculation method for the channel attention submodule is as follows:

[0126] (4-6)

[0127] (4-7)

[0128] In the formula, For channel weight vectors, This represents the Sigmoid activation function. and These are the trainable weight parameters for the multilayer perceptron. The spatiotemporal feature map is extracted by the ConvLSTM module. and These represent spatiotemporal feature maps respectively. Perform global max pooling and average pooling operations. This is the spatiotemporal feature map after channel attention enhancement. This indicates element-wise multiplication;

[0129] The calculation method for the spatial attention submodule is as follows:

[0130] (4-8)

[0131] (4-9)

[0132] In the formula, For spatial weight vectors, This represents the Sigmoid activation function. This indicates splicing along the channel dimension. It is a 7×7 convolution operator. and These represent the spatiotemporal feature maps after channel attention enhancement. Perform global max pooling and average pooling operations along the channel direction. This is a spatiotemporal feature map after spatial attention enhancement. This indicates element-wise multiplication.

[0133] In detail, the CBAM module consists of a channel attention submodule and a spatial attention submodule connected in series, which perform attention enhancement on the channel dimension and spatial dimension of the feature map, respectively. The multidimensional spatiotemporal feature map obtained by the ConvLSTM module is input into the convolutional block attention module (i.e., the CBAM module) for feature weight adjustment, which improves the model's ability to identify key regions and important patterns in the spatial distribution of carbon emissions.

[0134] Specifically, first, channel attention is calculated. This involves analyzing the feature maps extracted from the ConvLSTM module. F Global average pooling and global max pooling are applied in the spatial dimension to obtain two description vectors of length C, which are the number of channels. The channel attention submodule performs global pooling on the 360×180×64 feature map in the spatial dimension. Through average pooling and max pooling, two channel description vectors of length 64 are obtained. These vectors are then transformed using a shared multilayer perceptron and activated using the sigmoid function to generate 64-dimensional channel weight coefficients. These two vectors are then nonlinearly transformed and summed using the shared multilayer perceptron, and activated again using the sigmoid function to generate the channel weight vector. Given a channel weight vector of size 1×1×C, combine it with the original feature map. F Multiply each channel element-wise to perform channel attention enhancement and obtain the weighted feature map. F' Important channels are given higher weights, while less important channels are suppressed. The channel attention mechanism enables the model to amplify those feature channels that are most useful for carbon emission prediction, reflecting specific navigation conditions or emission intensity patterns, while reducing the interference of noisy channels on the prediction.

[0135] Then, the feature map after channel enhancement The input is fed into the spatial attention module. Spatial attention extracts key location information by compressing the channel dimension: for Global average pooling and max pooling are performed along the channel dimension to obtain two single-channel feature maps of size H×W. These maps are then concatenated along the channel dimension to form a two-channel feature combination. A 7×7 convolution is then applied to this combination and activated by the Sigmoid function to obtain the spatial weight vector. The result For a two-dimensional weighted image with the same planar dimensions as the original feature map, compare it with... F' Element-wise multiplication yields the final attention-weighted feature map. F''After spatial attention enhancement, the model can adaptively highlight the most critical emission hotspots in the carbon emission feature map, such as major shipping lanes and ports, while suppressing irrelevant information in the background, thus improving its ability to perceive and represent spatial clustering of carbon emissions. After processing by the CBAM module, the model obtains an enhanced spatiotemporal feature representation, still 360×180 pixels (channels integrated into a 64-channel weighted feature map). Finally, the model uses an output convolutional layer (one kernel, 1 channel) to map the attention-enhanced 64-channel feature map into a single-channel carbon emission intensity prediction map. This output image, with a size of 360×180, matches the input frame size and reflects the carbon emissions of each grid cell globally at the corresponding prediction time.

[0136] Understandably, this invention addresses the domain characteristics of ship carbon emissions (including strong constraints at the land-sea boundary, significant directionality of the main channel, steep nearshore emission gradient, sparse AIS observations, and uneven time steps). While maintaining the overall structure of the ConvLSTM backbone and CBAM attention, it proposes a domain-specific improvement strategy: 1) Introducing land-sea masking and shoreline buffer modulation into the channel attention and spatial attention of CBAM to suppress the weight of land pixels and avoid false hotspots generated by prediction results crossing the shoreline; 2) Increasing the directionality gain of the main channel in the spatial attention, appropriately amplifying the attention weight and lateral attenuation along the historical track direction, thereby maintaining the continuous distribution of hotspots along the channel.

[0137] Specifically, the loss function of the spatiotemporal prediction model for carbon emissions is:

[0138]

[0139] In the formula, The mean absolute error loss after gating. and These are the row and column indices of the grid cell, respectively. For time steps, For missing measurement masks, only if the cell is located in the sea area and at time step The value is 1 if it is a valid observation or a high-confidence interpolation; otherwise, the value is 0. For the spatiotemporal prediction model of carbon emissions at time step For grid cells The predicted carbon emissions, For grid cells At time step The actual carbon emissions.

[0140] Understandably, in the design of the loss function, to adapt to the sparse and irregular characteristics of AIS trajectory data in time and space, a missing measurement mask is introduced for gating the loss function calculation. This mechanism does not change the training samples and network structure, but only suppresses the interference of missing or low-confidence pixels on the loss calculation through the mask, thereby improving the stability and prediction accuracy of the model under sparse observation conditions.

[0141] S6: Obtain historical carbon emission image data with time series characteristics before the period to be predicted, and after normalization and grayscale processing, input it into the trained carbon emission spatiotemporal prediction model for prediction to obtain the spatial distribution image of ship carbon emissions during the prediction period.

[0142] In detail, the model takes normalized carbon emission image data from multiple consecutive historical frames as input and uses a trained spatiotemporal carbon emission prediction model based on ConvLSTM-CBAMNet to progressively predict the carbon emission distribution of container ships over the next n time steps. Specifically, the model first calculates the hidden state representation of the last time step based on the input sequence, and then uses this hidden state as the initial condition to iteratively predict the carbon emission distribution of the next time step. ; then Feedback is fed into the model to make predictions. This process continues until the desired result is obtained. , , ..., Prediction results for multiple future time points. Each frame of the predicted image output by the model has the same size as the original input image and is represented in a two-dimensional grid format. Table 6 compares the prediction performance of the spatiotemporal prediction model for carbon emissions built using the improved ConvLSTM network structure provided in this embodiment of the invention with that using a regular ConvLSTM model.

[0143]

[0144] As shown in Table 6, the spatiotemporal carbon emission prediction model provided in this embodiment of the invention achieves lower RMSE and MAE and higher SSIM in the spatiotemporal prediction task of ship carbon emissions, thus verifying its effectiveness in the task. Furthermore, Figure 5 This also demonstrates that the spatiotemporal prediction method for ship carbon emissions provided in this embodiment of the invention has a high degree of accuracy in prediction.

[0145] S7: Based on the total carbon emissions of each grid cell obtained in step S33, predict the ship carbon emissions of key grid cells.

[0146] Understandably, to compensate for the shortcomings of macro-level global carbon emission spatiotemporal prediction models in terms of accuracy across single-grid time series, this embodiment further introduces a Mamba-based model on top of the global prediction framework. NDA's single-grid emissions prediction model is used to finely characterize the carbon emission trends of key grid units over time. This allows for understanding not only the spatial evolution of global carbon emissions but also precise identification of carbon emissions in key areas. By combining global and key grid unit predictions, it provides technical support for subsequent carbon emission management. The key grid units are those in high-emission areas, such as those with average emissions in the top 10% of the entire region, typically located in straits, waterways, and ports. This model combines state-space modeling with graph-structured spatial propagation mechanisms, enabling it to dynamically integrate neighborhood spatial information while preserving the temporal evolution patterns, thus improving the accuracy and stability of single-point predictions.

[0147] In detail, such as Figure 6 As shown, step S7 includes the following steps:

[0148] Step S71: Select the grid cell where the high emission area is located as the key grid cell, and determine the number of the key grid cell. The high emission area is a strait, waterway or port.

[0149] Understandably, the latitude and longitude of a strait, waterway, or port can be used to determine the corresponding grid cell number.

[0150] Step S72: According to the set time interval, obtain the total carbon emissions of each grid cell obtained in step S33, and arrange the total carbon emissions of the same grid cell according to time to obtain the carbon emission time series of each grid cell;

[0151] Understandably, the carbon emissions time series for each grid cell are stored in their respective grid cells.

[0152] Step S73: Based on the key grid cell number, determine the neighboring grid cell numbers of the key grid cell, and calculate the spatial weight between the key grid cell and each neighboring grid cell respectively. The neighboring grid cells share vertices with the key grid cell.

[0153] Specifically, the formula for calculating the spatial weight between the key grid cell and the neighboring grid cells is as follows:

[0154] (5-1)

[0155] (5-2)

[0156] In the formula, As key grid units i With neighboring grid cells j Distance decay weights between (based solely on distance) and attenuation coefficient Sure), As key grid units i Center point and neighboring grid cells j The geographical distance between the center points, This is the distance attenuation coefficient; As key grid units i With neighboring grid cells j Spatial weights between them For the channel enhancement coefficient, As a channel enhancement factor and , Indicates key grid cells i The set of neighboring grid cells, which is a set composed of neighboring grid cells. j Indicates the first j neighborhood grid cells and , i Indicates the first i One key grid unit;

[0157] Step S74: Based on the number of the neighboring grid cell, extract the carbon emission time series of the neighboring grid cell, and based on the spatial weight between the key grid cell and each neighboring grid cell, perform a weighted summation of the carbon emission time series of each neighboring grid cell at each time step to obtain a neighborhood diffusion feature time series that corresponds one-to-one with the carbon emission time series of the key grid cell in the time dimension.

[0158] Specifically, the NDA model calculates the neighborhood diffusion characteristics at one time step according to formula (5-3):

[0159] (5-3)

[0160] In the formula, As key grid units i At any moment t The neighborhood diffusion characteristics, As key grid units i With neighboring grid cells j Spatial weights between them Neighborhood grid cell j At any moment t carbon emissions, Indicates key grid cells i The set of neighboring grid cells, which is a set composed of neighboring grid cells.j Indicates the first j neighborhood grid cells and ;

[0161] Step S75: Extract the carbon emission time series of the key grid unit according to the key grid unit number;

[0162] Step S76: Construct a single grid cell emission prediction model based on the Mamba model, and train and test the single grid cell emission prediction model based on the carbon emission time series of the key grid cell and its corresponding neighborhood diffusion characteristic time series to obtain a trained single grid cell emission prediction model.

[0163] Specifically, the Mamba model concatenates the carbon emission time series of key grid cells with their corresponding neighborhood diffusion feature time series to obtain a feature vector for a single time step, as shown in the following formula:

[0164]

[0165] In the formula, For feature vectors, As key grid units i In the moment t carbon emissions, As key grid units i At any moment t The neighborhood diffusion characteristics, Indicates the focus on grid cells i In the moment t Carbon emissions and key grid units i At any moment t The neighborhood diffusion features are concatenated into a feature vector.

[0166] Understandably, this embodiment introduces a Neighborhood Diffusion Adapter (NDA) model at the input end to incorporate neighborhood diffusion features to characterize emission changes in the area surrounding key grid cells. This model is independent of the time-series backbone structure and is responsible for extracting weighted emission features from the spatial neighborhood of key grid cells. These features reflect the overall level of dynamic carbon emission changes over time around the key grid cells and are used as supplementary input along with the historical emission data of the key grid cells for joint modeling in the subsequent Mamba state-space network. This enhances the ability of the single grid cell emission prediction model to perceive the local spatial environment. The neighborhood diffusion features output by the NDA model are aligned with the historical carbon emission sequence of the key grid cells in the time dimension and are input as an independent channel to the encoding end of the Mamba model, achieving a deep fusion of spatial diffusion information and time-dependent modeling. The specific architectures of the NDA and Mamba models are existing technologies and will not be described further here.

[0167] Specifically, the algorithmic expression for the single-grid cell emission prediction model based on the Mamba model is as follows:

[0168] (6-1)

[0169] (6-2)

[0170] In the formula, The current hidden state. Let A be the hidden state at the previous time step, and let A, B, C, and D be trainable parameter matrices. The input to the model is a feature vector time series obtained by concatenating the carbon emission time series of key grid cells with their corresponding neighborhood diffusion feature time series, and then transforming it through a linear mapping layer. This feature vector time series meets the input requirements of the Mamba model. The output of the model is the predicted carbon emissions of ships in the key grid cells.

[0171] The algorithmic expression for the selective gating mechanism of the single-grid cell emission prediction model based on the Mamba model is as follows:

[0172] (6-3)

[0173] (6-4)

[0174] (6-5)

[0175] In the formula, For selective gating vectors, It is the Sigmoid activation function. , , and For trainable parameters, As input to the model, This is the hidden state from the previous moment. The input feature vector time series after linear transformation. This indicates element-wise multiplication. This is the valid input after gating.

[0176] Understandably, the Mamba model, based on a selective state-space network, is the core structure of a single grid cell emission prediction model. The Mamba model can efficiently capture temporal dependencies over long time spans and dynamically fuse spatial background information provided by neighborhood diffusion features, achieving in-depth modeling of carbon emission trends in key grid cells. Furthermore, in this embodiment, the Mamba model introduces a selective gating mechanism on top of the quasi-state-space network structure. This mechanism can dynamically adjust the information inflow ratio based on the current input and historical states, effectively preserving key information and suppressing irrelevant noise in long sequences, thereby enhancing the model's adaptability.

[0177] Specifically, the single-grid cell emission prediction model selects the mean square error function and the mean absolute error function as loss functions, wherein:

[0178] The mean square error function has the following form:

[0179]

[0180] The form of the mean absolute error function is:

[0181]

[0182] In the formula, H For the predicted future time step, For the first Future time step by step , t For the current moment, For prediction of key grid cells i exist Carbon emissions at any given moment. For prediction of key grid cells i exist Carbon emissions at any given moment. As key grid units i exist Real-time carbon emissions; These are the smoothing regularization coefficients.

[0183] Understandably, to optimize the model, the loss function uses a combination of point prediction error and a regularization term. The point prediction error is chosen as mean squared error (MSE). Mean absolute error (MAE) introduces a smoothing regularization term to enhance the smoothness of the prediction curve and suppress non-physical abrupt changes, thereby achieving more robust performance under different data distributions.

[0184] Step S77: Obtain the historical carbon emission time series of the key grid cell before the period to be predicted and its corresponding neighborhood diffusion characteristic time series, and input them into the trained single grid cell emission prediction model to predict the ship carbon emissions of the key grid cell.

[0185] Understandably, in the single-grid-cell emission prediction model of this embodiment, the model input is a feature vector time series composed of two parts: one part is the carbon emission time series of the key grid cell, used to capture its own temporal evolution pattern; the other part is the neighborhood diffusion feature time series, used to incorporate carbon emission change information around the key grid cell to enhance the model's ability to perceive the local spatial background. The model output is the predicted carbon emission sequence of the key grid cell for multiple future steps.

[0186] This invention also provides a carbon emission monitoring platform, such as... Figure 7 As shown, it includes: an external database 1, a carbon emission spatiotemporal prediction module 2, a retrieval enhancement and generation module 3, a large model interface module 4, a large model module 5, a hotspot tracking and early warning module 6, a carbon emission spatiotemporal visualization module 7, a green route recommendation module 8, an intelligent question-and-answer and assistant analysis module 9, a port low-carbon scheduling module 10, and a carbon emission report and trend analysis module 11. Specifically, the retrieval enhancement and generation module 3 is connected to the external database 1 and the carbon emission spatiotemporal prediction module 2; the large model interface module 4 is connected to the retrieval enhancement and generation module 3, the large model module 5, the green route recommendation module 8, the intelligent question-and-answer and assistant analysis module 9, the port low-carbon scheduling module 10, and the carbon emission report and trend analysis module 11; and the hotspot tracking and early warning module 6 and the carbon emission spatiotemporal visualization module 7 are both connected to the carbon emission spatiotemporal prediction module 2.

[0187] Specifically, external database 1 is used to store information related to the shipping industry. This information includes, for example, shipping business knowledge, emission regulations, historical cases, and policy documents.

[0188] The carbon emission spatiotemporal prediction module 2 is used to calculate and predict ship carbon emissions using the aforementioned ship carbon emission spatiotemporal prediction method, and provides carbon emission calculation results and carbon emission prediction results. The carbon emission calculation results represent the carbon emissions of different sea areas at past times, and the carbon emission prediction results represent the predicted carbon emissions of different sea areas at future times.

[0189] The retrieval-augmented generation module 3 is used to obtain information related to the shipping industry, carbon emission calculation results, and carbon emission prediction results from the external database and the carbon emission spatiotemporal prediction module, respectively, based on the business request. It then constructs a prompt message corresponding to the business request based on the obtained information and results. The business request originates from the green route recommendation module 8, the intelligent question-and-answer and assistant analysis module 9, the port low-carbon scheduling module 10, and the carbon emission report and trend analysis module 11. Specifically, the retrieval-augmented generation module 3 incorporates Retrieval-Augmented Generation (RAG) technology, retrieving relevant information from the database in real time and integrating it into the answer when a user's question needs to be answered.

[0190] The large model interface module 4 forwards business requests obtained from the green route recommendation module 8, intelligent question-and-answer and assistant analysis module 9, port low-carbon scheduling module 10, and carbon emission report and trend analysis module 11 to the retrieval enhancement generation module 3. It encapsulates the prompts obtained from the retrieval enhancement generation module 3 into text or table information corresponding to the business requests that the large model can directly parse and use, and sends this information to the large model module 5. It also sends the answers or analyses corresponding to the business requests obtained from the large model module 5 to the green route recommendation module 8, intelligent question-and-answer and assistant analysis module 9, port low-carbon scheduling module 10, and carbon emission report and trend analysis module 11, respectively, for display to the user. Specifically, this module enables the large model module 5 to interface with other parts of the platform, integrating large model services through an application programming interface (API) or local deployment, and building a RAGFlow workflow. The text or table information includes, for example, hot topic summaries and numerical conclusions.

[0191] The large model module 5 is used to generate corresponding answers or analyses from the text or table information corresponding to the business request obtained from the large model interface module 4, and then send the answers or analyses to the large model interface module 4. The answers or analyses can be text descriptions or data tables. It should be noted that the large model module 5 can be a DeepSeek large model or other large models.

[0192] The hotspot tracking and early warning module 6 obtains real-time carbon emission prediction results for high-emission hotspot areas from the carbon emission spatiotemporal prediction module 2, compares these predictions with preset thresholds, and issues an early warning notification when the thresholds are exceeded. Specifically, this module automatically identifies high-emission hotspot areas and their intensity thresholds based on the real-time carbon emission prediction results, and monitors and tracks the formation and future trajectory of carbon emission hotspots (e.g., their movement from the high seas to ports). When carbon emissions in a key area are detected to be about to exceed a preset threshold or experience a sharp increase, the module issues an early warning notification, prompting relevant departments to take proactive measures. This helps regulators quickly locate future carbon emission risk areas and conduct forward-looking management.

[0193] The carbon emission spatiotemporal visualization module 7 obtains real-time carbon emission prediction and calculation results from the carbon emission spatiotemporal prediction module 2 and presents them to users in an interactive visualization manner. In detail, users can intuitively view heatmaps of emissions from different sea areas at various past and future times and observe the evolution of carbon emissions over time by playing them on a timeline. This module also supports area zooming and numerical hover display, helping users quickly identify high-emission areas and trends.

[0194] The Green Route Recommendation Module 8 sends route planning business requests to the Large Model Interface Module 4 and displays the corresponding answers or analyses obtained from the Large Model Module 5 via the Large Model Interface Module 4 to the user. Specifically, the user inputs the port of origin, port of destination, and time requirements (i.e., route planning requirements) through this module. Based on carbon emission prediction results, the platform introduces a carbon emission cost dimension to the route planning, assessing the total emissions of different alternative routes over the expected voyage. The Large Model Module 5 comprehensively considers factors such as distance, speed limits, and emission predictions to provide the optimal green route solution (e.g., how to pass through a certain sea area / at what speed to use to minimize carbon emissions), and provides corresponding emission reduction estimates and justifications. This module helps shipping companies reduce their carbon footprint while meeting operational requirements.

[0195] The intelligent question-answering and assistant analysis module 9 sends business requests for questions to the large model interface module 4, and displays the corresponding answers or analyses obtained from the large model module 5 via the large model interface module 4 to the user. Specifically, users can ask natural language questions about carbon emission data, trend causes, and governance measures. The platform will generate and provide professional answers through the retrieval enhancement generation module 3 and the large model module 5. For example, the question: Why did carbon emissions at a certain port increase abnormally last month? The answer: The platform will query data such as the number of ships berthing and berthing duration at the port last month, and may find that the surge in emissions is due to the concentrated arrival of large container ships. The large model module 5 will then generate an explanation, pointing out the specific reasons and supporting data. This module also supports multi-turn dialogues, allowing users to ask follow-up questions for further analysis.

[0196] The port low-carbon scheduling module 10 sends low-carbon scheduling business requests to the large model interface module 4 and displays the corresponding answers or analyses obtained from the large model module 5 via the large model interface module 4 to the user. Specifically, based on ship arrival forecasts and carbon emission forecasts, the platform can identify potential future emission peaks in the port area. The large model module 5 then generates optimized scheduling plans, such as adjusting berth operation sequences, optimizing tugboat allocation, and improving loading and unloading efficiency, to reduce emissions by smoothing out peak flows. The platform interface lists suggested measures and their expected effects, such as "postponing non-urgent operations can reduce emissions by X%," providing intelligent scheduling suggestions for port managers' decision-making.

[0197] The carbon emission reporting and trend analysis module 11 sends business requests to the large model interface module 4 to generate carbon emission reports and trend analyses. It then displays the corresponding responses or analyses obtained from the large model module 5 via the large model interface module 4 to the user. Specifically, users can periodically or as needed view carbon emission analysis reports and trend analysis reports generated by the large model module 5 based on the latest forecast results. The reports include historical data statistics, current status overview, future trend forecasts, emission trends of key shipping routes / ports, ranking of hotspot areas, and explanations of abnormal situations. Managers can directly reference these reports for internal reporting or policy formulation. Furthermore, this module supports one-click export of report documents and can provide explanations of the report content based on user inquiries.

[0198] Specifically, the large-scale model linkage and invocation process of the carbon emission monitoring platform implemented in this instance is as follows:

[0199] If a user's question directly relates to prediction data (such as "What will be the carbon emission intensity of a certain region tomorrow?" or "Which day will have the highest emissions in the coming week?"), relevant data fragments are retrieved from the carbon emission spatiotemporal prediction module 2 before calling the large model module 5 (e.g., extracting the predicted value for the region tomorrow, or extracting the daily total emission peak information for the coming week). This retrieved structured data is converted into text descriptions or tables, which are then embedded into the prompt context of the large model module 5 by the large model interface module 4. Based on this, the large model module 5 provides a precise and well-supported answer, including specific numerical values ​​and explanations of conclusions.

[0200] If a user requests higher-level analysis or suggestions (such as "Please help me develop a minimum carbon emission shipping route based on the forecast" or "How to reduce port emissions next month?"), the platform will trigger a series of operations: First, it will summarize relevant forecast results (such as emission estimates for multiple candidate shipping routes, emission changes under different port scheduling schemes, etc.) and retrieve relevant background knowledge (such as energy conservation and emission reduction measures, historical optimization schemes). Then, it will organize this material into prompts and provide them to the large model module 5. The large model module 5 will then generate comprehensive analysis results, such as recommending a specific shipping route and explaining its carbon emission reduction effect, or proposing several port operation adjustment measures and giving the expected emission reduction of each measure.

[0201] If a user needs to inquire about further details after receiving the results, the platform will re-enter the cycle of searching and calling the larger model, continuously providing deeper information.

[0202] During the collaborative invocation process, the large model module 5 can also act as an interpreter. For example, based on the results given by the prediction model, users may further ask causal analysis questions such as "Why will emissions in a certain region increase in the future?" At this time, the platform will retrieve possible reasons (such as the expected increase in ship traffic in the region, or the impact of weather, etc.) and provide them to the large model module 5. The large model module 5 then combines shipping knowledge with prediction data to provide a reasonable explanation and answer.

[0203] The carbon emission monitoring platform provided in this embodiment features a user-friendly interface that presents acquired data, results, and generated responses to users in an intuitive manner, enabling data visualization and human-computer interaction. The user interface includes spatiotemporal visualization components for carbon emissions (such as maps and timelines), a hotspot alert panel, an intelligent Q&A chat window, and route optimization interactive tools. Users can view prediction results (such as map heatmaps and trend charts), receive system alerts, and ask questions or receive suggestions in natural language through a web interface or client.

[0204] The carbon emission monitoring platform provided in this invention aims to organically integrate the spatiotemporal prediction method for ship carbon emissions provided by this invention with a large-scale model, creating an AI platform for green shipping and intelligent regulation, and realizing the application of carbon emission prediction results. In this platform, the large-scale model can understand the semantics of queries and provide accurate answers and explanations based on platform data. It will play multiple roles as an "AI engine," providing decision support and human-computer interaction capabilities for intelligent carbon emission regulation. For example, it supports natural language query interfaces, utilizing the powerful language understanding and generation capabilities of the large-scale model to answer various user questions about carbon emissions. It converts complex spatiotemporal prediction data into easily understandable language descriptions, thereby assisting decision-makers in understanding the prediction results. Based on the future emission distribution and hotspot information provided by the prediction model, it acts as an intelligent advisor, integrating shipping business knowledge and prediction data to automatically generate low-carbon operation suggestions, such as suggesting staggered port entry and exit to avoid high-emission periods. It provides actionable solutions and justifications by integrating historical data and real-time predictions.

[0205] It should be noted that the carbon emission monitoring platform provided in this embodiment of the invention is not limited to realizing real-time visualization of carbon emissions, hotspot tracking and early warning, intelligent response and analysis, green route planning, low-carbon port scheduling, and carbon emission reporting and trend analysis. It can also achieve other application goals according to user needs.

[0206] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for spatio-temporal prediction of carbon emissions of a ship, characterized in that, The method comprises the following steps: (1) constructing a ship carbon emission calculation model according to static data of the ship, and calculating carbon emissions of each ship in multiple trajectory segments based on the calculation model, wherein the trajectory segment is a trajectory formed by connecting two adjacent AIS trajectory points; (2) mapping the carbon emissions of each ship in multiple trajectory segments to a spatial grid to obtain global carbon emissions, obtaining the global carbon emissions at a set time interval, arranging the global carbon emissions according to time, and visualizing the global carbon emissions to obtain carbon emission image data with time sequence characteristics; (3) normalizing and greying the carbon emission image data, dividing the processed carbon emission image data into multiple "input-output" sample pairs to obtain a data set; (4) constructing a carbon emission spatio-temporal prediction model based on an improved ConvLSTM network structure, training and testing the carbon emission spatio-temporal prediction model based on the data set, and obtaining a trained carbon emission spatio-temporal prediction model; (5) obtaining historical carbon emission image data with time sequence characteristics before a to-be-predicted period, normalizing and greying the historical carbon emission image data, and inputting the historical carbon emission image data into the trained carbon emission spatio-temporal prediction model for prediction to obtain a spatial distribution image of ship carbon emissions in the to-be-predicted period.

2. The ship carbon emission space-time prediction method of claim 1, wherein, Before step (1), the following steps are further included: The longitude and latitude of each abnormal AIS trajectory point of each ship are respectively interpolated and reconstructed according to formulas (1-1) and (1-2): （1-1） （1-2） wherein, is the longitude of the abnormal AIS track point after reconstruction, is the longitude of the previous normal AIS track point of the abnormal AIS track point, is the longitude of the next normal AIS track point of the abnormal AIS track point, is the latitude of the abnormal AIS track point after reconstruction, is the latitude of the previous normal AIS track point of the abnormal AIS track point, is the latitude of the next normal AIS track point of the abnormal AIS track point, is the time at which the abnormal AIS track point occurs, is the reporting time of the previous normal AIS track point of the abnormal AIS track point, is the reporting time of the next normal AIS track point of the abnormal AIS track point, the previous normal AIS track point and the next normal AIS track point are adjacent to the abnormal AIS track point.

3. The ship carbon emission space-time prediction method of claim 1, wherein, Step (2) specifically comprises: (21) dividing the global region by using a WGS84 coordinate system, dividing 360 cells in the longitude direction and 180 cells in the latitude direction to construct a spatial grid with a resolution of 1°×1°; (22) assigning a number to each grid cell in the spatial grid, and the number has the following relationship with the longitude and latitude of any position point in the global region: （2-1） （2-2） （2-3） wherein, is a row index of the grid cell, is a latitude coordinate of an arbitrary location point within the global area, is a column index of the grid cell, is a longitude coordinate of an arbitrary location point within the global area, denotes a floor operation, is a number of the grid cell; (23) attributing the carbon emissions of each ship in multiple trajectory segments to corresponding grid cells in the spatial grid, and accumulating the carbon emissions of all ships in the same grid cell to obtain the total carbon emissions of each grid cell, wherein the number of the grid cell corresponding to each trajectory segment is calculated according to the longitude and latitude coordinates of the end point of the trajectory segment according to formulas (2-1), (2-2) and (2-3), and the carbon emissions of the trajectory segment are attributed to the grid cell with the number; (24) obtaining the global carbon emissions at a set time interval, and arranging the global carbon emissions according to time to obtain global carbon emissions with time sequence characteristics; (25) visualizing the global carbon emissions with time sequence characteristics to obtain carbon emission image data with time sequence characteristics.

4. The ship carbon emission space-time prediction method of claim 1, wherein, Step (3) specifically comprises: (31) normalizing the carbon emission image data by using formula (3-1): （3-1） wherein is the normalized carbon emission image data, is the original carbon emission image data, and are the maximum and minimum values of the original carbon emission image data, respectively, is a small constant to prevent log zero error, taken as 10 -6 ; (32) converting the normalized carbon emission image data into gray image data; (33) The gray image data is divided into a plurality of "input-output" sample pairs satisfying the structure requirements of the deep learning model by the method of the sliding window, and a data set is obtained.

5. The ship carbon emission space-time prediction method of claim 1, wherein, In step (4), the improved ConvLSTM network structure includes a ConvLSTM module and a CBAM module, the CBAM module includes a channel attention submodule and a spatial attention submodule, and the spatiotemporal feature map extracted by the ConvLSTM module is input into the CBAM module, wherein: The algorithm expression of the ConvLSTM module is: （4-1） （4-2） （4-3） （4-4） （4-5） wherein, , , , and are the input gate, the forget gate, the output gate, the cell state at the current time step, and the hidden state at the current time step, respectively, denotes a Sigmoid activation function, denotes a two-dimensional convolution operation, is a convolution kernel weight input to the gate, is the input feature map at the current time step, is a convolution kernel weight hidden to the gate, is the hidden state at the previous time step, is an input gate bias term; , and are a convolution kernel weight input to the forget gate, a convolution kernel weight hidden to the forget gate, and a forget gate bias term, respectively; , and are a convolution kernel weight input to the output gate, a convolution kernel weight hidden to the output gate, and an output gate bias term, respectively; is the cell state at the previous time step, is a candidate cell state extracted via convolution, denotes a Hadamard element-wise product; The calculation method of the channel attention submodule is: （4-6） （4-7） wherein, is a channel weight vector, denotes a Sigmoid activation function, and are trainable weight parameters of the multi-layer perceptron, is a spatio-temporal feature map extracted by the ConvLSTM module, and denote global max-pooling and average-pooling operations on the spatio-temporal feature map , respectively, is a spatio-temporal feature map enhanced by channel attention, denotes element-wise multiplication; The calculation method of the spatial attention submodule is: （4-8） （4-9） In the formula, is a spatial weight vector, denotes a Sigmoid activation function, denotes concatenation in the channel dimension, is a 7x7 convolution operator, and respectively denote the spatio-temporal feature map after channel attention enhancement global max-pooling and average-pooling operations are performed along the channel direction, is a spatio-temporal feature map after spatial attention enhancement, denotes element-wise multiplication.

6. The ship carbon emission space-time prediction method of claim 3, wherein, After step (5), the following steps are further included: (6) According to the total carbon emissions of each grid cell obtained in step (23), the ship carbon emissions of the key grid cell are predicted.

7. The ship carbon emission space-time prediction method of claim 6, wherein, Step (6) specifically includes: (61) Select the grid cell where the high-emission area is located as the key grid cell, and determine the number of the key grid cell, wherein the high-emission area is a strait, a channel or a port; (62) According to the set time interval, the total carbon emissions of each grid cell obtained in step (23) are obtained, and the total carbon emissions of the same grid cell are arranged according to time to obtain the carbon emission time series of each grid cell; (63) According to the number of the key grid cell, the number of the neighborhood grid cell of the key grid cell is determined, and the spatial weight between the key grid cell and each neighborhood grid cell is calculated, wherein the neighborhood grid cell shares a vertex with the key grid cell; (64) According to the number of the neighborhood grid cell, the carbon emission time series of the neighborhood grid cell are extracted, and the carbon emissions in the carbon emission time series of each neighborhood grid cell at each time step are weighted and summed according to the spatial weight between the key grid cell and each neighborhood grid cell, to obtain the neighborhood diffusion feature time series corresponding to the carbon emission time series of the key grid cell in the time dimension; (65) According to the number of the key grid cell, the carbon emission time series of the key grid cell are extracted; (66) A single grid cell emission prediction model based on the Mamba model is constructed, and the single grid cell emission prediction model is trained and tested based on the carbon emission time series of the key grid cell and the neighborhood diffusion feature time series corresponding thereto, to obtain a trained single grid cell emission prediction model; (67) The historical carbon emission time series of the key grid cell before the to-be-predicted period and the neighborhood diffusion feature time series corresponding thereto are obtained, which are input into the trained single grid cell emission prediction model to predict the ship carbon emissions of the key grid cell.

8. The ship carbon emission spatiotemporal prediction method of claim 7, wherein: In step (63), the spatial weight between the key grid cell and the neighborhood grid cell is calculated according to formula (5-1) and formula (5-2): （5-1） （5-2） wherein is a distance decay weight between the focal grid cell i and the neighboring grid cell j , is a geographic distance between the center point of the focal grid cell i and the center point of the neighboring grid cell j , is a distance decay coefficient; is a spatial weight between the focal grid cell i and the neighboring grid cell j , is a channel enhancement coefficient, is a channel enhancement factor and , denotes a set of neighboring grid cells of the focal grid cell i , the set of neighboring grid cells being a set consisting of neighboring grid cells, j denotes the j th neighboring grid cell and , i denotes the i th focal grid cell; In step (64), the neighborhood diffusion feature of a time step is calculated according to formula (5-3): （5-3） wherein is the focal grid cell i at time t is the neighborhood diffusion feature, is the focal grid cell i is the spatial weight between the focal grid cell j and the neighborhood grid cell is the neighborhood grid cell j at time t is the carbon emission, denotes the set of neighborhood grid cells of the focal grid cell i , the set of neighborhood grid cells being a set composed of neighborhood grid cells, j denotes the j th neighborhood grid cell and .

9. The ship carbon emission space-time prediction method of claim 8, wherein, The algorithm expression of the single grid cell emission prediction model based on the Mamba model is: （6-1） （6-2） In the formula, is the hidden state at the current time, is the hidden state at the previous time, A, B, C and D are trainable parameter matrices, is the input of the model, which is the feature vector time series obtained after the Mamba model splices the carbon emission time series of the key grid cell and the corresponding neighborhood diffusion feature time series, and converts it through a linear mapping layer. The feature vector time series meets the input requirements of the Mamba model, is the output of the model, that is, the predicted ship carbon emission of the key grid cell; The algorithm expression of the selective gating mechanism of the single grid cell emission prediction model based on the Mamba model is: （6-3） （6-4） （6-5） wherein, is a selective gating vector, is a Sigmoid activation function, , , and are trainable parameters, is an input of the model, is a hidden state at the previous time step, is an input feature vector time series after linear transformation, denotes element-wise multiplication, is an effective input after gating.

10. A carbon emission regulation platform, characterized in that, The system comprises an external database, a carbon emission space-time prediction module, a retrieval enhancement generation module, a large model interface module, a large model module, a hotspot tracking and early warning module, a carbon emission space-time visualization module, a green route recommendation module, an intelligent question answering and assistant analysis module, a port low-carbon scheduling module, and a carbon emission report and trend analysis module, wherein: The external database is used to store information related to the shipping field; The carbon emission space-time prediction module is used to calculate and predict ship carbon emissions using the ship carbon emission space-time prediction method of any one of claims 1-9, and provide carbon emission calculation results and carbon emission prediction results; The retrieval enhancement generation module is connected with the external database and the carbon emission space-time prediction module, and is used to obtain information related to the shipping field, carbon emission calculation results, and carbon emission prediction results from the external database and the carbon emission space-time prediction module, respectively, according to a business request, and construct prompt information corresponding to the business request according to the obtained information and results; The large model interface module is connected with the retrieval enhancement generation module, the large model module, the green route recommendation module, the intelligent question answering and assistant analysis module, the port low-carbon scheduling module, and the carbon emission report and trend analysis module, and is used to forward the business request obtained from the green route recommendation module, the intelligent question answering and assistant analysis module, the port low-carbon scheduling module, and the carbon emission report and trend analysis module to the retrieval enhancement generation module, encapsulate the prompt information obtained from the retrieval enhancement generation module into text or table information corresponding to the business request that can be directly parsed and used by the large model, and send it to the large model module, and send the answers or analyses corresponding to the business request obtained from the large model module to the green route recommendation module, the intelligent question answering and assistant analysis module, the port low-carbon scheduling module, and the carbon emission report and trend analysis module, respectively, for display to the user; The large model module is used to generate corresponding answers or analyses from the text or table information corresponding to the business request obtained from the large model interface module, and send the answers or analyses to the large model interface module; The hotspot tracking and early warning module is connected with the carbon emission space-time prediction module, and is used to obtain carbon emission prediction results of high emission hotspot areas from the carbon emission space-time prediction module in real time, and compare the carbon emission prediction results with a preset threshold, and issue a warning notification when the preset threshold is exceeded; The carbon emission space-time visualization module is connected with the carbon emission space-time prediction module, and is used to obtain carbon emission prediction results and carbon emission calculation results from the carbon emission space-time prediction module in real time, and display them to the user in an interactive and visual manner. The green route recommendation module is configured to send a route planning service request to the large model interface module and display a response or analysis corresponding to the route planning service request obtained from the large model interface module to a user; The intelligent question answering and assistant analysis module is configured to send a question service request to the large model interface module and display a response or analysis corresponding to the question service request obtained from the large model interface module to a user; The port low-carbon scheduling module is configured to send a low-carbon scheduling service request to the large model interface module and display a response or analysis corresponding to the low-carbon scheduling service request obtained from the large model interface module to a user; The carbon emission report and trend analysis module is configured to send a service request for generating a carbon emission report and trend analysis to the large model interface module and display a response or analysis corresponding to the service request for generating a carbon emission report and trend analysis obtained from the large model interface module to a user.

Images