Liquefied natural gas transportation scheduling method and system based on security maximization

Abstract

The application provides a liquefied natural gas transportation scheduling method and system based on safety maximization, relates to the technical field of safety management, and comprises the following steps: feature extraction is performed on multi-dimensional information of a transportation vehicle through deep learning, a safety feature vector is generated by using a time sequence convolution network and an attention mechanism; a road network dynamic graph structure is established, and a candidate route is calculated by using a graph neural network; and multi-objective optimization is performed on the route by using a genetic algorithm, and an optimal scheduling instruction is generated by comprehensively considering a safety situation score and route attributes. The application improves the safety and efficiency of liquefied natural gas transportation.

Description

Technical Field

[0001] This invention relates to safety management technology, and more particularly to a method and system for liquefied natural gas transportation scheduling based on maximizing safety. Background Technology

[0002] Liquefied natural gas (LNG) is an important component of clean energy, and its land transportation volume continues to grow. However, LNG's flammability, explosiveness, and low-temperature characteristics during transportation pose significant safety risks. Accidents could result in severe casualties and property damage. Therefore, ensuring the safety of LNG road transportation has become a crucial issue that urgently needs to be addressed in the energy logistics sector.

[0003] Current liquefied natural gas (LNG) transportation scheduling primarily relies on route planning based on the shortest path or lowest cost principle, typically employing static path planning algorithms combined with human experience for scheduling decisions. With the development of information technology, some companies have introduced geographic information systems (GIS) and electronic mapping technology, improving the accuracy of transportation route planning; however, significant shortcomings remain in safety management.

[0004] The existing dispatch system lacks the ability to comprehensively analyze multi-dimensional information about vehicles and the environment, and cannot effectively utilize multi-source heterogeneous data such as vehicle status, driving behavior, and weather conditions to conduct safety risk assessments, resulting in insufficient safety situation awareness and difficulty in predicting potential risks.

[0005] Existing scheduling optimization algorithms often focus on economic indicators, such as transportation costs and time efficiency, while insufficiently considering the quantitative aspects of safety factors. They lack a multi-objective optimization mechanism that takes safety situation scoring as the core optimization objective, making it difficult to achieve a balance between transportation efficiency and cost while ensuring safety, and failing to meet the fundamental requirement of "safety first" in the transportation of dangerous goods. Summary of the Invention

[0006] The embodiments of the present invention provide a method and system for liquefied natural gas transportation scheduling based on maximizing safety, which can solve the problems in the prior art.

[0007] A first aspect of the present invention provides a liquefied natural gas (LNG) transportation scheduling method based on maximizing safety, comprising:

[0008] Multidimensional information of liquefied natural gas transport vehicles is obtained, and features are extracted from the multidimensional information using a deep learning algorithm. Temporal features are extracted through a temporal convolutional network, and feature weights from different data sources are fused through an attention mechanism to generate a feature vector reflecting the safety status. A safety status score is calculated based on the feature vector.

[0009] A dynamic graph structure for a road transport network is established, comprising nodes and road segments, each road segment having weight information related to traffic attributes; based on the dynamic graph structure of the road transport network, a graph neural network is used to calculate multiple candidate transport routes;

[0010] Multi-objective optimization is performed on the multiple candidate transportation routes, taking the safety situation score and route attributes as optimization objectives. A genetic algorithm is used to solve the problem, resulting in multiple sets of optimized solutions. The optimal transportation route is selected according to preset weights. Transportation scheduling instructions are generated based on the optimal transportation route.

[0011] In one alternative implementation,

[0012] The steps of using deep learning algorithms to extract features from the multidimensional information, extracting temporal features through a temporal convolutional network, fusing feature weights from different data sources through an attention mechanism to generate a feature vector reflecting the security status, and calculating a security situation score based on the feature vector include:

[0013] The multidimensional information is segmented temporally using a sliding window, and preprocessed data is obtained through normalization. Temporal convolutional networks are then used to extract features from the preprocessed data, and temporal features are obtained through causal convolution operations and multi-layer dilated convolutions.

[0014] A modality correlation matrix is ​​constructed based on the temporal features, which represents the correlation strength between different modal features. The modality correlation matrix is ​​input into multiple parallel feature extraction channels with different receptive field sizes to obtain a multi-scale feature representation. A temporal state memory unit is constructed to dynamically fuse the multi-scale feature representation with historical features and calculate the modality anomaly score. Based on the modality anomaly score, the features are adaptively adjusted, and a dynamic weight matrix is ​​obtained by using a softmax function with a temperature coefficient adjustment.

[0015] The dynamic weight matrix and the attention weight matrix are multiplied by Hadamard to obtain the adaptive dynamic weight attention value.

[0016] Multi-head attention calculation is performed based on the adaptive dynamic weight attention value. The calculation result is then connected with the original feature through residual connection and layer normalization to obtain the fused feature.

[0017] The fused features are used to calculate feature weights using a multilayer perceptron. The feature weights are then multiplied by the hidden layer features using a Hadamard product, and a security situation score is obtained after a nonlinear transformation.

[0018] In one alternative implementation,

[0019] The steps for establishing a dynamic graph structure for a road transport network, comprising nodes and road segments, with each road segment having weight information related to its traffic attributes, include:

[0020] A dynamic graph structure for the road transport network is established, mapping road nodes to graph nodes and road segments to graph edges. The feature vectors of the graph nodes include the node's geographical coordinates, node type, and node capacity. The initial weights of the graph edges are calculated using road segment length, average vehicle speed, and traffic flow.

[0021] Real-time traffic flow data and historical statistical data of the road transport network are obtained, and spatiotemporal alignment and standardization are performed to obtain standardized data;

[0022] The standardized data is combined with the feature vectors of the graph nodes to calculate the dynamic association strength between the graph nodes. The dynamic association strength is calculated using a learnable parameter matrix and the LeakyReLU activation function to obtain a temporally enhanced node representation.

[0023] The spatial correlation degree between the graph nodes is calculated based on the spatial query matrix and the spatial key value matrix. The spatial correlation degree is then fused with the temporally enhanced node representation to obtain the spatiotemporal fusion feature.

[0024] The initial weights of the graph edges are updated based on the spatiotemporal fusion features to obtain updated graph edge weights. The deviation score between the updated graph edge weights and the historical graph edge weights is calculated. When the deviation score exceeds a preset deviation threshold, a weight decay factor is calculated based on the deviation score. The updated graph edge weights are then adjusted using the weight decay factor to obtain adjusted graph edge weights.

[0025] In one alternative implementation,

[0026] Based on the dynamic graph structure of the road transport network, the steps of calculating multiple candidate transport routes using a graph neural network include:

[0027] Based on the feature vectors of the graph nodes and the adjusted graph edge weights, feature aggregation and transmission are performed to obtain global features;

[0028] The dynamic association weights between nodes are calculated by the similarity between the query vector and the key vector of the node. The global features are then enhanced based on the dynamic association weights to obtain enhanced features.

[0029] Construct a gated graph convolutional unit including update gates and reset gates, calculate the parameters of update gates and reset gates based on the enhanced features and historical features, and iteratively update the enhanced features using the parameters of update gates and reset gates to obtain a road network representation with temporal memory capability;

[0030] An initial transportation route is generated based on the road network representation. The product of the temporal memory feature score and the position weight of each graph edge in the initial transportation route is calculated to obtain the graph edge weighted score. The weighted scores of all graph edges in the initial transportation route are accumulated to obtain the comprehensive score. The transportation route with the highest comprehensive score is selected as the benchmark transportation route.

[0031] The baseline transportation route is branched and expanded. Based on a preset expansion depth parameter, the graph edges in the baseline transportation route are replaced and searched to generate multiple candidate transportation routes.

[0032] In one alternative implementation,

[0033] Multi-objective optimization is performed on the multiple candidate transportation routes, using the safety situation score and route attributes as optimization objectives. A genetic algorithm is used to solve the problem, obtaining multiple sets of optimized solutions. The step of selecting the optimal transportation route according to preset weights includes:

[0034] The multiple candidate transportation routes are mapped to gene coding sequences, where the graph edges in each route are mapped to gene positions in the gene sequence, and the security weight of each gene position is calculated based on the security situation score.

[0035] A multi-objective optimization function is constructed, which combines the safety situation score and the graph edge weights in the dynamic graph structure of the road transport network to obtain the safety objective value, and combines the time consumption and distance consumption of the route to obtain the efficiency objective value.

[0036] An initial coding population of size N is generated. The safety objective value and efficiency objective value of each individual in the population are calculated as fitness values. The initial coding population is then sorted non-dominated based on the multi-objective optimization function.

[0037] Based on the initial encoded population, calculate the safety weight similarity of adjacent gene loci, determine the crossover interval according to the safety weight similarity, and adjust the crossover probability based on the non-dominated level; calculate the safety weight deviation value and fluctuation value of gene loci to determine the mutation position, adjust the mutation probability according to the number of iterations to perform mutation operation; recalculate the safety target value and efficiency target value for the mutated individuals, perform non-dominated sorting and elite retention, and obtain the offspring population;

[0038] For individuals in the offspring population, a comprehensive score is calculated by weighting the safety target value and the efficiency target value based on preset weights, and the transportation route corresponding to the individual with the highest comprehensive score is selected as the optimal transportation route.

[0039] In one alternative implementation,

[0040] The steps include: calculating the safety weight similarity of adjacent gene loci based on the initial coding population; determining the crossover interval based on the safety weight similarity and adjusting the crossover probability based on the non-dominance level; calculating the safety weight deviation and fluctuation value of gene loci to determine the mutation location; and adjusting the mutation probability according to the number of iterations to perform the mutation operation.

[0041] Calculate the safety weight difference between adjacent gene loci in the gene sequence, divide the safety weight difference by the larger safety weight value among the adjacent gene loci to obtain the safety weight similarity, and determine the gene loci with the safety weight similarity greater than the dynamic similarity threshold and the gene locus distance less than the preset interval length constraint as the cross interval;

[0042] The crossover probability is obtained by adding the lower limit of the crossover probability and the crossover probability adjustment amount. The crossover probability adjustment amount is calculated by multiplying the difference between the upper limit of the crossover probability and the lower limit of the crossover probability by the non-dominated level adjustment coefficient. The non-dominated level adjustment coefficient is the ratio of the difference between the maximum non-dominated level and the current individual's non-dominated level to the maximum non-dominated level.

[0043] The ratio of the difference between the population average safety weight and the safety weight at the current position to the population average safety weight is used as the safety weight deviation value. The ratio of the fluctuation range of the safety weight at the current position to the maximum safety weight is used as the safety weight fluctuation value. The weighted sum of the safety weight deviation value and the safety weight fluctuation value is used as the gene locus evaluation score. Based on the gene locus evaluation score, the mutation position is selected, the mutation probability is adjusted according to the current iteration number, and the mutation operation is performed on the selected position.

[0044] In one alternative implementation,

[0045] The steps for calculating a comprehensive score based on a weighted average of the safety target value and the efficiency target value using preset weights include:

[0046] The route risk coefficient is calculated based on the liquefied natural gas loading capacity, container protection level, and distribution characteristics of emergency resources along the route. The product of the route risk coefficient and the safety target value is used as the corrected safety target value.

[0047] The corrected safety and efficiency target values ​​are normalized, the Euclidean distance between the target values ​​of individuals on the Pareto front is calculated, clustering is performed based on the distance threshold, and the cluster center is determined as the representative individual.

[0048] The main score is calculated for each representative individual based on the weights of the preset safety target value and efficiency target value. At the same time, candidate individuals are selected based on the road segment overlap rate and candidate scores are calculated. The weighted sum of the main score and the candidate scores is used as the comprehensive score of the individual.

[0049] The optimal representative individual and its corresponding alternative individuals are selected based on the comprehensive score, serving as the primary recommended route and alternative routes, respectively, and a dynamic switching threshold based on the safety target value is set.

[0050] A second aspect of the present invention provides a liquefied natural gas transportation dispatching system based on maximizing safety, comprising:

[0051] The first unit is used to acquire multi-dimensional information of liquefied natural gas transport vehicles, use deep learning algorithms to extract features from the multi-dimensional information, extract temporal features through a temporal convolutional network, fuse feature weights from different data sources through an attention mechanism, generate a feature vector reflecting the safety status, and calculate a safety status score based on the feature vector.

[0052] The second unit is used to establish a dynamic graph structure of a road transport network, which includes nodes and road segments, each of which has weight information of traffic attributes; based on the dynamic graph structure of the road transport network, a graph neural network is used to calculate multiple candidate transport routes;

[0053] The third unit is used to perform multi-objective optimization on the multiple candidate transportation routes. The safety situation score and route attributes are used as optimization objectives. A genetic algorithm is used to solve the problem and obtain multiple sets of optimized solutions. The optimal transportation route is selected according to a preset weight. Transportation scheduling instructions are generated based on the optimal transportation route.

[0054] A third aspect of the present invention provides an electronic device, comprising:

[0055] processor;

[0056] Memory used to store processor-executable instructions;

[0057] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0058] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0059] This invention extracts features from multidimensional information of liquefied natural gas transport vehicles using deep learning algorithms, and combines features from different data sources with temporal convolutional networks and attention mechanisms to generate a safety status feature vector and calculate a safety situation score. This enables a comprehensive assessment of safety risks during transportation and improves transportation safety.

[0060] This invention establishes a dynamic graph structure for road transport networks and uses graph neural networks to calculate multiple candidate transport routes, enabling intelligent analysis of complex road networks. It can adapt to dynamic changes in road conditions and improve the accuracy and adaptability of route planning.

[0061] This invention performs multi-objective optimization on candidate transportation routes, taking safety situation scores and route attributes as optimization objectives, and uses a genetic algorithm to solve for the optimal transportation route. This achieves a balance between safety and efficiency, reduces the risks of dangerous goods transportation, improves transportation efficiency and economic benefits, and provides a scientific basis for liquefied natural gas transportation scheduling decisions. Attached Figure Description

[0062] Figure 1 This is a flowchart illustrating the liquefied natural gas transportation scheduling method based on maximizing safety, as described in an embodiment of the present invention.

[0063] Figure 2 A flowchart for multi-objective optimization of liquefied natural gas transportation routes based on safety situation scoring. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0066] Figure 1 This is a flowchart illustrating the liquefied natural gas transportation scheduling method based on maximizing safety, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0067] Multidimensional information of liquefied natural gas transport vehicles is obtained, and features are extracted from the multidimensional information using a deep learning algorithm. Temporal features are extracted through a temporal convolutional network, and feature weights from different data sources are fused through an attention mechanism to generate a feature vector reflecting the safety status. A safety status score is calculated based on the feature vector.

[0068] A dynamic graph structure for a road transport network is established, comprising nodes and road segments, each road segment having weight information related to traffic attributes; based on the dynamic graph structure of the road transport network, a graph neural network is used to calculate multiple candidate transport routes;

[0069] Multi-objective optimization is performed on the multiple candidate transportation routes, taking the safety situation score and route attributes as optimization objectives. A genetic algorithm is used to solve the problem, resulting in multiple sets of optimized solutions. The optimal transportation route is selected according to preset weights. Transportation scheduling instructions are generated based on the optimal transportation route.

[0070] In one optional implementation, the steps of using a deep learning algorithm to extract features from the multidimensional information, extracting temporal features through a temporal convolutional network, fusing feature weights from different data sources through an attention mechanism to generate a feature vector reflecting the security status, and calculating a security situation score based on the feature vector include:

[0071] The multidimensional information is segmented temporally using a sliding window, and preprocessed data is obtained through normalization. Temporal convolutional networks are then used to extract features from the preprocessed data, and temporal features are obtained through causal convolution operations and multi-layer dilated convolutions.

[0072] A modality correlation matrix is ​​constructed based on the temporal features, which represents the correlation strength between different modal features. The modality correlation matrix is ​​input into multiple parallel feature extraction channels with different receptive field sizes to obtain a multi-scale feature representation. A temporal state memory unit is constructed to dynamically fuse the multi-scale feature representation with historical features and calculate the modality anomaly score. Based on the modality anomaly score, the features are adaptively adjusted, and a dynamic weight matrix is ​​obtained by using a softmax function with a temperature coefficient adjustment.

[0073] The dynamic weight matrix and the attention weight matrix are multiplied by Hadamard to obtain the adaptive dynamic weight attention value.

[0074] Multi-head attention calculation is performed based on the adaptive dynamic weight attention value. The calculation result is then connected with the original feature through residual connection and layer normalization to obtain the fused feature.

[0075] The fused features are used to calculate feature weights using a multilayer perceptron. The feature weights are then multiplied by the hidden layer features using a Hadamard product, and a security situation score is obtained after a nonlinear transformation.

[0076] For example, multidimensional information includes vehicle operating parameters, road condition information, weather conditions, and other data. A sliding window technique is used to segment this multidimensional information into time-series segments, with a window size of 120 time points and a sliding step size of 30 time points. The segmented data is then normalized to unify data with different dimensions into the [-1, 1] interval, resulting in preprocessed data.

[0077] The temporal convolutional network comprises causal convolutional layers and multiple dilated convolutional layers. The causal convolutional layers use one-dimensional convolutions with a kernel size of 3, ensuring that the output at the current time step depends only on the input at the current and past time steps, avoiding information leakage. The dilated convolutional layers employ four layers with dilation rates of 1, 2, 4, and 8, achieving a receptive field of 31 time points, capable of capturing long-term dependencies. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function, with 64 output channels for each layer. This network structure extracts temporal features from preprocessed data.

[0078] For the extracted temporal features, assuming there are M different modalities, each with dimension d, an M×M modality correlation matrix is ​​constructed. Each element in the matrix represents the correlation strength between two modal features, obtained by calculating the dot product between the feature vectors of different modalities and then normalizing it. For example, the correlation strength between vehicle speed and fuel consumption is 0.73, indicating a high correlation between the two. The modality correlation matrix is ​​input into parallel feature extraction channels with different receptive field sizes. Three parallel channels are designed with receptive field sizes of 3×3, 5×5, and 7×7, respectively. Each channel contains two convolutional layers: the first layer outputs 32 channels, and the second layer outputs 16 channels. The output features of the three channels are concatenated to obtain a multi-scale feature representation with a dimension of M×48.

[0079] A temporal state memory unit is constructed, employing an LSTM-like structure and comprising three gating components: an update gate, a forget gate, and an output gate. The update gate determines the update ratio by calculating the cosine similarity between the current multi-scale features and historical features; the update ratio increases when the similarity is below 0.65. The forget gate controls the retention of historical features, setting a decay coefficient of 0.05 based on feature timeliness. During fusion, the current and historical features are weighted and summed according to the weights determined by the update gate to obtain the fused features. The modality anomaly score is obtained by calculating the Mahalanobis distance between the current feature and the average of the features over the past 10 time windows, and then exponentially smoothed using a smoothing factor of 0.3. Under normal operating conditions, the anomaly score remains between 0.05 and 0.2, but can rapidly rise above 0.5 under sudden conditions, such as a braking modality anomaly score reaching 0.78 during sudden braking. The weights of modal features with high anomaly scores are increased to enhance the ability to perceive abnormal states. The adjusted weights are normalized using a softmax function with a temperature coefficient of 0.5 to obtain a dynamic weight matrix. The value of each element in the matrix is ​​between [0,1], and the sum of the elements in each row is 1.

[0080] The dynamic weight matrix is ​​multiplied by a pre-trained attention weight matrix using a Hadamard product. The attention weight matrix is ​​calculated via a self-attention mechanism, transforming temporal features into three vectors: query, key, and value, each with a dimension of 64. The dot product of the query and key is calculated and divided by 8 as a scaling factor. After softmax normalization, these vectors are weighted and summed with the value vectors to form an attention weight matrix of the same dimension as the dynamic weight matrix, where each element represents the importance of different features. Corresponding element-wise multiplication of the two matrices yields the adaptive dynamic weight attention value, which considers both feature importance and anomaly severity.

[0081] Eight attention heads are set up, each with an output dimension of 64. The outputs of the eight heads are concatenated and dimensionality reduced through a linear layer, then residually connected to the original features, and finally normalized to obtain the fused features.

[0082] The fused features are weighted using a multilayer perceptron (MLP). The MLP consists of three fully connected layers with 256, 128, and M neurons, respectively, where M is the number of feature modalities. The first and second layers are followed by ReLU activation functions, and the last layer outputs the feature weights. These weights are then multiplied by the hidden layer features using a Hadamard product (element-wise multiplication) to obtain the weighted features.

[0083] Finally, the weighted features are nonlinearly transformed through two fully connected layers. The first layer has 64 neurons and uses the ReLU activation function, while the second layer has 1 neuron and uses the Sigmoid activation function to restrict the output to the [0,1] interval, thus obtaining a safety status score. The closer the score is to 1, the better the safety status; the closer it is to 0, the higher the safety risk. For example, when the vehicle is driving smoothly in good road conditions, the safety status score can reach 0.92; while when driving at high speed in adverse weather conditions, the score drops to 0.41.

[0084] This invention employs a deep learning algorithm to extract features from multi-dimensional information of liquefied natural gas (LNG) transport vehicles. By effectively fusing feature weights from different data sources through temporal convolutional networks and attention mechanisms, it generates feature vectors that accurately reflect the safety status. This method can capture long-term dependencies in time-series data, identify abnormal patterns, and adaptively weight features of different modalities, significantly improving the accuracy and real-time performance of safety status scoring and providing a reliable safety assessment basis for LNG transport scheduling.

[0085] In one optional implementation, the step of establishing a dynamic graph structure of a road transport network, the road transport network including nodes and road segments, and each road segment having weight information of traffic attributes, includes:

[0086] A dynamic graph structure for the road transport network is established, mapping road nodes to graph nodes and road segments to graph edges. The feature vectors of the graph nodes include the node's geographical coordinates, node type, and node capacity. The initial weights of the graph edges are calculated using road segment length, average vehicle speed, and traffic flow.

[0087] Real-time traffic flow data and historical statistical data of the road transport network are obtained, and spatiotemporal alignment and standardization are performed to obtain standardized data;

[0088] The standardized data is combined with the feature vectors of the graph nodes to calculate the dynamic association strength between the graph nodes. The dynamic association strength is calculated using a learnable parameter matrix and the LeakyReLU activation function to obtain a temporally enhanced node representation.

[0089] The spatial correlation degree between the graph nodes is calculated based on the spatial query matrix and the spatial key value matrix. The spatial correlation degree is then fused with the temporally enhanced node representation to obtain the spatiotemporal fusion feature.

[0090] The initial weights of the graph edges are updated based on the spatiotemporal fusion features to obtain updated graph edge weights. The deviation score between the updated graph edge weights and the historical graph edge weights is calculated. When the deviation score exceeds a preset deviation threshold, a weight decay factor is calculated based on the deviation score. The updated graph edge weights are then adjusted using the weight decay factor to obtain adjusted graph edge weights.

[0091] For example, a dynamic graph structure for the road transport network is established, mapping key locations such as road intersections, transport origins, and destinations as graph nodes, and road segments connecting these nodes as graph edges. The feature vectors of the graph nodes are designed as multi-dimensional vectors containing three main types of information: the node's geographical coordinates are represented by latitude and longitude, accurate to six decimal places; the node type includes ordinary intersections, highway entrances / exits, service areas, and hazardous materials transfer stations, represented by integer codes; and the node's capacity is represented by the number of vehicles that can pass through per hour. The initial weights of the graph edges are calculated based on the road segment attributes, using the formula: road segment length divided by average vehicle speed multiplied by a congestion coefficient, where the congestion coefficient is determined by traffic flow data. For example, a road segment with a length of 5 kilometers, an average vehicle speed of 60 kilometers per hour, and a congestion coefficient of 1.2 has an initial weight value of 0.1.

[0092] Real-time traffic flow data is obtained from traffic monitoring, including the number of vehicles, average speed, and congestion status for each road segment, sampled every 5 minutes. Historical statistics include the average traffic flow and standard deviation for the same period over the past 30 days. This data undergoes spatiotemporal alignment and standardization. Spatiotemporal alignment maps data from different sources to a unified time window and spatial unit, filling in missing values ​​with the average of nearby time periods. Standardization transforms various data types to a distribution with a mean of 0 and a standard deviation of 1; specifically, this involves subtracting the mean and then dividing by the standard deviation. For example, if a road segment has a real-time traffic flow of 120 vehicles / hour, a historical mean of 100 vehicles / hour, and a standard deviation of 30 vehicles / hour, the standardized value is 0.67.

[0093] For each node, an input feature vector is constructed, containing the node's own features and standardized traffic data from adjacent road segments. A two-layer fully connected network is used to process the input features: the first layer contains 64 neurons, and the second layer contains 32 neurons. A LeakyReLU activation function is used between the two layers, with a negative slope parameter set to 0.2. The network output vector serves as the hidden representation of the node. The dot product of the hidden representations of any two adjacent nodes is calculated, then transformed using a learnable parameter matrix, and processed using the LeakyReLU activation function to obtain the dynamic association strength between the nodes. The association strength ranges from [0,1], with larger values ​​indicating a stronger association between the two nodes. For example, the association strength between two traffic hub nodes can reach 0.85 during peak hours, but is only 0.42 during off-peak hours. The dynamic association strength is used as a weight to perform weighted aggregation of the node hidden representations, resulting in a time-enhanced node representation.

[0094] Construct a spatial query matrix, with a size equal to the number of nodes multiplied by the feature dimension. Each row represents the query vector of a node. Similarly, construct a spatial key-value matrix, with each row representing the key vector of a node. Both the query vector and key vector have a dimension of 32. Calculate the dot product of the query vector for each node and the key vectors for all nodes to obtain the original association score. Adjust the association score using the geographical distance between nodes, with greater distance resulting in lower weights. The adjustment formula is the original score divided by the square of the distance, multiplied by a distance decay coefficient set to 0.1. Perform softmax normalization on the adjusted association score to obtain the spatial association degree between nodes. For example, two nodes 2 kilometers apart have an original association score of 4.2 and an adjusted spatial association degree of 0.63. Fuse the spatial association degree with the temporally enhanced node representation using a weighted summation method with weights of 0.6 and 0.4, respectively, to obtain the spatiotemporal fusion feature.

[0095] For graph edges connecting nodes i and j, the spatiotemporal fusion features of nodes i and j are extracted and processed through a three-layer fully connected network with 64, 32, and 1 neurons in each layer, using the ReLU activation function. The network output value is used as a weight adjustment factor, and the initial weights are multiplied by this adjustment factor to obtain the updated graph edge weights. The deviation score between the updated graph edge weights and the historical graph edge weights is calculated, defined as the absolute value of the difference between the two divided by the historical graph edge weights. A preset deviation threshold of 0.3 is set; when the deviation score exceeds this threshold, the weight change is considered abnormal and further adjustment is required. A weight decay factor is calculated based on the deviation score, calculated as 1 minus the ratio of the deviation score to the threshold, then multiplied by 0.5. For example, if the deviation score is 0.45, the weight decay factor is 0.75. The updated graph edge weights are adjusted using the weight decay factor, calculated by multiplying the updated weights by the weight decay factor and then adding the historical weights multiplied by 1 minus the weight decay factor.

[0096] For a road network with 200 nodes and 350 edges, establishing the initial graph structure requires collecting information on the geographical location, type, and capacity of each node. For example, the feature vector of a highway exit node is [longitude 118.123456, latitude 39.654321, type code 3, capacity 1200]. Real-time data is updated every 5 minutes. Each update calculates the dynamic correlation strength and spatial correlation degree, and updates the weights of all graph edges. In one update, a road segment with an original weight of 0.08 experienced a traffic accident, resulting in a weight of 0.25 and a deviation score of 2.125, exceeding the preset threshold of 0.3. A weight decay factor of 0.15 was calculated, and the final adjusted weight was 0.173, avoiding drastic weight fluctuations.

[0097] This invention presents a method for establishing a dynamic graph structure for road transport networks. By mapping road nodes and road segments to a graph structure and combining real-time traffic data and historical statistical data, the method calculates the dynamic correlation strength and spatial correlation degree between nodes, thereby achieving an accurate representation of the road network state. This method introduces a weight decay factor mechanism to handle abnormal traffic conditions, effectively suppressing drastic weight fluctuations and improving the accuracy and stability of the road network representation.

[0098] In one optional implementation, the step of calculating multiple candidate transportation routes using a graph neural network based on the dynamic graph structure of the road transportation network includes:

[0099] Based on the feature vectors of the graph nodes and the adjusted graph edge weights, feature aggregation and transmission are performed to obtain global features;

[0100] The dynamic association weights between nodes are calculated by the similarity between the query vector and the key vector of the node. The global features are then enhanced based on the dynamic association weights to obtain enhanced features.

[0101] Construct a gated graph convolutional unit including update gates and reset gates, calculate the parameters of update gates and reset gates based on the enhanced features and historical features, and iteratively update the enhanced features using the parameters of update gates and reset gates to obtain a road network representation with temporal memory capability;

[0102] An initial transportation route is generated based on the road network representation. The product of the temporal memory feature score and the position weight of each graph edge in the initial transportation route is calculated to obtain the graph edge weighted score. The weighted scores of all graph edges in the initial transportation route are accumulated to obtain the comprehensive score. The transportation route with the highest comprehensive score is selected as the benchmark transportation route.

[0103] The baseline transportation route is branched and expanded. Based on a preset expansion depth parameter, the graph edges in the baseline transportation route are replaced and searched to generate multiple candidate transportation routes.

[0104] For example, after establishing the dynamic graph structure of the road transport network, a graph neural network is needed to calculate multiple candidate transport routes. In the specific implementation, feature aggregation and propagation are performed based on the feature vectors of the graph nodes and the adjusted graph edge weights to obtain global features. The specific operation of feature aggregation is as follows: for each graph node, the feature vectors of its directly adjacent nodes are collected, and these feature vectors are weighted and summed using the adjusted graph edge weights to generate the node's neighborhood representation. The larger the graph edge weight, the smaller the feature contribution of the corresponding adjacent nodes. For example, for a graph node with 3 adjacent nodes, its feature vector is [0.5, 0.3, 0.2], and the feature vectors of the 3 adjacent nodes are [0.6, 0.4, 0.1], [0.2, 0.7, 0.3], and [0.3, 0.2, 0.8], with corresponding graph edge weights of 0.08, 0.12, and 0.05, then the neighborhood representation of this node is [0.35, 0.42, 0.46]. Feature propagation combines neighborhood representations with node-specific features, processed through a two-layer fully connected network. The first layer contains 128 neurons, and the second layer contains 64 neurons, using the ReLU activation function. This process is repeated three times for all nodes, completing three layers of graph convolution operations to obtain global features. The global feature dimension for each node is 64.

[0105] For each node's global features, two independent fully connected layers are used to generate query vectors and key vectors, each with a dimension of 32. The dot product of each node's query vector and the key vectors of all nodes is calculated to obtain the original similarity score. The original score is scaled using a scaling factor of 8 and then normalized using a softmax function to obtain the dynamic association weights between nodes. The association weights represent the strength of information interaction between nodes, with a value range of [0,1]. For example, the dynamic association weight distribution of a node with 200 other nodes is as follows: the average weight with directly connected nodes is 0.15, the average weight with second-order connected nodes is 0.05, and the average weight with other nodes is less than 0.01. Based on the dynamic association weights, the global features of all nodes are weighted and summed to obtain the association context features of each node. The association context features are concatenated with the node's original global features and processed through a fully connected layer to obtain an enhanced feature with a dimension of 64.

[0106] The design of the gated graph convolutional unit borrows from the idea of ​​the gated recurrent unit, comprising two key components: the update gate and the reset gate. The update gate controls the fusion ratio of the new input enhanced features and historical features, while the reset gate controls which information from the historical features needs to be retained. The update gate parameter is calculated by concatenating the enhanced features and historical features, processing them through a fully connected layer using the sigmoid activation function, with an output value range of [0,1]. The reset gate parameter is calculated similarly to the update gate, but uses independent fully connected layer parameters. For example, if the enhanced features and historical features of a node at the current time step have a high similarity, the calculated update gate parameter is 0.3, indicating that 70% of the historical information is retained; while the reset gate parameter is 0.85, indicating that 85% of the historical features are retained for the calculation of candidate update states. Based on the update gate and reset gate parameters, candidate update states are calculated and fused with historical features according to the ratio determined by the update gate to obtain a new node representation. This process is repeated for all nodes, iterating 5 times in the time dimension, finally obtaining a road network representation with temporal memory capability, with the same dimension as the enhanced features, 64.

[0107] The initial transportation route generation employs an improved Dijkstra algorithm, but replaces traditional edge weights with a comprehensive edge score calculated based on road network representation. For a graph edge connecting node i and node j, the road network representation vectors of the two nodes are extracted, their cosine similarity is calculated as a compatibility score, and then multiplied by the edge's safety factor to obtain the temporal memory feature score. The safety factor is defined based on road grade, historical accident rate, and distance to hazard sources; the safety factor is 0.9 for highways, 0.7 for national highways, and 0.5 for provincial highways. The position weight considers the influence of the graph edge's position within the entire path, with higher position weights for edges near the start and end points, and lower weights for edges in the middle. Specifically, it is calculated as 1 minus the normalized distance to the midpoint of the path, multiplied by 0.5, and then added back to 0.5. For example, in a path containing 10 graph edges, the position weight of the 1st edge is 0.95, the 5th edge is 0.5, and the 10th edge is 0.95. The temporal memory feature score is multiplied by the position weight to obtain the weighted graph edge score. The weighted scores of all edges in the path are summed to obtain the overall score of the path. Dijkstra's algorithm is used to generate 10 paths with the highest overall scores, and the path with the highest overall score is selected as the baseline transportation route.

[0108] Branch expansion is performed on the baseline transportation route to explore more high-quality routes and increase the diversity of route selection. The expansion depth parameter is set to 3, indicating that a maximum of 3 graph edges in the baseline route can be replaced. In the specific implementation, for each graph edge in the baseline transportation route, alternative parallel road segments are found, and the impact of the replacement on the overall score is evaluated. Parallel road segments are defined as different roads connecting the same start and end areas, identified by geographic coordinates and road network topology. For example, if the baseline route includes highway segment AB, and there is a national highway segment A'-B' as a parallel alternative road segment, the evaluation process is as follows: temporarily replace one graph edge in the baseline route with its parallel segment, recalculate the overall score of the path, and retain the replacement scheme if the score decreases by no more than 10%. By individually or in combination replacing graph edges at different locations in the baseline route, up to 20 candidate transportation routes are generated. To improve computational efficiency, a greedy strategy is used to prioritize replacing graph edges with less impact on the overall score. The final generated candidate route set includes the baseline route and multiple variant routes, providing diverse options for subsequent transportation route optimization.

[0109] For example, liquefied natural gas (LNG) transportation needs to travel from location A to location B. The baseline route passes through 5 main nodes and contains 8 graph edges. Through branch expansion, two parallel alternative routes are found at the positions of the 2nd and 6th graph edges, generating a total of 9 candidate transportation routes. For example, the baseline route has a comprehensive score of 8.75, the variant route after replacing the 2nd graph edge has a score of 8.42, the variant route after replacing the 6th graph edge has a score of 8.63, and the variant route after replacing two graph edges has a score of 8.31.

[0110] This invention employs a graph neural network to compute multiple candidate transportation routes. It captures the topological features of the road network through feature aggregation and dynamic weighting enhancement, and combines gated graph convolutional units to memorize temporal traffic states, generating high-quality baseline routes that consider safety factors. An innovative branch expansion mechanism is introduced to explore diverse candidate routes, ensuring both the overall quality of the routes and providing local optimization potential. This significantly improves the flexibility and adaptability of liquefied natural gas (LNG) transportation scheduling, laying a solid foundation for subsequent multi-objective optimization.

[0111] In one optional implementation, multi-objective optimization is performed on the multiple candidate transportation routes, using the safety situation score and route attributes as optimization objectives, and a genetic algorithm is used to solve the problem to obtain multiple sets of optimized solutions. The step of selecting the optimal transportation route according to preset weights includes:

[0112] The multiple candidate transportation routes are mapped to gene coding sequences, where the graph edges in each route are mapped to gene positions in the gene sequence, and the security weight of each gene position is calculated based on the security situation score.

[0113] A multi-objective optimization function is constructed, which combines the safety situation score and the graph edge weights in the dynamic graph structure of the road transport network to obtain the safety objective value, and combines the time consumption and distance consumption of the route to obtain the efficiency objective value.

[0114] An initial coding population of size N is generated. The safety objective value and efficiency objective value of each individual in the population are calculated as fitness values. The initial coding population is then sorted non-dominated based on the multi-objective optimization function.

[0115] Based on the initial encoded population, calculate the safety weight similarity of adjacent gene loci, determine the crossover interval according to the safety weight similarity, and adjust the crossover probability based on the non-dominated level; calculate the safety weight deviation value and fluctuation value of gene loci to determine the mutation position, adjust the mutation probability according to the number of iterations to perform mutation operation; recalculate the safety target value and efficiency target value for the mutated individuals, perform non-dominated sorting and elite retention, and obtain the offspring population;

[0116] For individuals in the offspring population, a comprehensive score is calculated by weighting the safety target value and the efficiency target value based on preset weights, and the transportation route corresponding to the individual with the highest comprehensive score is selected as the optimal transportation route.

[0117] Combination Figure 2 The flowchart illustrating the multi-objective optimization of liquefied natural gas (LNG) transportation routes based on safety situation scores is as follows: An integer encoding method is used, where each candidate route is represented as an integer sequence, the length of which equals the number of graph edges contained in the route. Each integer in the sequence represents a unique identifier for a graph edge. For example, a candidate route containing 8 graph edges can be encoded as [105,221,367,429,512,675,723,890]. For each node, its safety weight is calculated, defined as the safety situation score of that graph edge divided by the average safety situation score of the route containing that edge. The safety situation score comprehensively considers factors such as road grade, traffic flow, weather conditions, and historical accident rate, with a score range of [0,1], where a higher value indicates greater safety. For example, if a candidate route contains 8 graph edges, and the security status scores of each edge are 0.82, 0.75, 0.90, 0.63, 0.86, 0.78, 0.92, and 0.71, with an average security status score of 0.80, then the security weight of the first gene position is 1.03, the security weight of the third gene position is 1.13, and the security weight of the fourth gene position is 0.79.

[0118] A multi-objective optimization function is constructed, including a safety objective value and an efficiency objective value. The safety objective value is calculated as the weighted average of the safety situation scores of all graph edges along the route. The weights are the reciprocals of the edge weights, indicating that the greater the difficulty of traversing an edge, the more important its safety. Specifically, the safety situation score of each edge is first multiplied by the reciprocal of its weight, then all products are summed, and finally divided by the sum of the reciprocals of all edge weights. The efficiency objective value is calculated as the weighted sum of the route's time consumption and distance consumption, with weights of 0.7 and 0.3, respectively. Time consumption is calculated based on the route length and average travel speed, while distance consumption is directly calculated using the total route length. For example, if a candidate route has a weighted average safety situation score of 0.85, a time consumption of 3.5 hours, and a distance consumption of 210 kilometers, then the safety objective value is 0.85, and the efficiency objective value is 0.7 × 3.5 + 0.3 × 210 = 65.45. Since the safety target value and the efficiency target value have different dimensions, they need to be normalized. The normalized safety target value ranges from [0,1] to [0,1], and the efficiency target value ranges from [0,1]. The smaller the value, the better the performance.

[0119] An initial coded population of size N is generated, with N set to 100. Each individual in the population represents a transportation route. The initial population is generated by directly encoding some routes from the candidate transportation routes, while simultaneously generating new individuals through local modifications to these routes to ensure the diversity of the initial population. Local modifications include replacing some graph edges with parallel segments and adjusting the node order in the routes. The safety objective value and efficiency objective value of each individual are calculated, and these two objective values ​​constitute the individual's fitness vector. The initial population is then sorted into different non-dominated levels. Specifically, for any two individuals a and b, if a's safety objective value and efficiency objective value are not inferior to b's, and at least one objective value is superior to b's, then a is said to dominate b. Individuals not dominated by any other individual form the first non-dominated level. From the remaining individuals, non-dominated individuals are then selected to form the second non-dominated level, and so on, until all individuals are assigned a level. For example, in an initial population of 100 individuals, 15 individuals belong to the first non-dominant rank, 25 individuals belong to the second non-dominant rank, and the remaining individuals are distributed in the third rank and below.

[0120] When performing genetic operations based on the initial coding population, a tournament selection method is first used to select parent individuals from the initial population. Specifically, five individuals are randomly selected from the population, their non-dominance levels are compared, and the individual with the highest non-dominance level is selected as the first parent; then five more individuals are randomly selected, and the individual with the highest non-dominance level is selected as the second parent. For the selected two parent individuals, the safety weight similarity of adjacent gene positions in their coding sequences is calculated. The similarity is defined as 1 minus the absolute value of the difference between the safety weights of the two gene positions. For example, if the safety weights of the third and fourth gene positions of an individual are 1.13 and 0.79 respectively, then the similarity between these two positions is 1 - |1.13 - 0.79| = 0.66. Positions with similarity below the threshold of 0.7 are selected as potential crossover points. If multiple positions meet the condition, the position with the lowest similarity is selected first. The crossover probability is dynamically adjusted according to the non-dominance level of the parent individuals; the higher the level, the greater the crossover probability. The base crossover probability is set to 0.8, increasing by 0.1 for first-level individuals, increasing by 0.05 for second-level individuals, and remaining unchanged for third and subsequent levels. At the selected intersection, the gene segments of two parent individuals are swapped to generate two new offspring individuals. For each offspring individual, the safety weight deviation and fluctuation value of each gene locus are calculated. The deviation value is defined as the difference between the safety weight of that gene locus and the average safety weight of all gene loci in the individual, and the fluctuation value is defined as the difference between the safety weight of that gene locus and the average safety weight of its adjacent gene loci. A gene locus selection probability distribution is constructed, with the probability proportional to the product of the deviation value and the fluctuation value. The mutation probability decreases with the number of iterations, initially set to 0.1, decreasing by 0.01 every 10 iterations, with a minimum of 0.01. The mutation probability determines whether to mutate the individual; if mutation is performed, the mutation location is selected according to the selection probability distribution. The mutation operation randomly selects a replacement edge from the candidate graph edge set. The candidate set includes other graph edges that connect the same start and end points as the original edge, or combinations of graph edges that can form an effective bypass route. The safety target value and efficiency target value are recalculated for all mutated individuals, and they are merged with the parent population and then subjected to non-dominated sorting. An elite retention strategy is employed, selecting individuals from highest to lowest non-dominant level, maintaining a population size of N. If the number of individuals at a certain non-dominant level exceeds the retention quota, crowding is calculated (based on the distribution distance between the safety and efficiency target values ​​in the target space), and individuals with high crowding are selected to ensure population diversity and obtain the offspring population. The termination condition for the algorithm's iteration process is set as follows: reaching the preset maximum number of iterations (500 generations), or the change in the safety and efficiency target values ​​of the optimal individual over 20 consecutive generations is less than 0.001. For example, in a certain optimization process, if the change in the target value of the optimal individual from generation 380 to 400 does not exceed the threshold of 0.001, the algorithm terminates prematurely at generation 400, and the optimal individual obtained at this time is the final solution. If the convergence condition is not met by generation 500, the iteration is also forcibly terminated.This termination condition setting ensures that the algorithm has sufficient room for evolution while avoiding the waste of computing resources caused by excessive iteration.

[0121] For individuals in the offspring population, a weighted average score is calculated based on the safety target value and the efficiency target value, using preset weights. These preset weights reflect the decision-maker's preference for safety and efficiency. For high-risk operations like liquefied natural gas transportation, the safety weight is typically set higher, such as 0.7 for safety and 0.3 for efficiency. The overall score is calculated by multiplying the safety target value by the safety weight and then adding the efficiency target value multiplied by the efficiency weight. Since the target value is normalized and a smaller value indicates better performance, the actual calculation uses 1 minus the target value to ensure a higher overall score indicates a better solution. For example, if an individual has a normalized safety target value of 0.2 and a normalized efficiency target value of 0.35, its overall score is 0.7 × (1 - 0.2) + 0.3 × (1 - 0.35) = 0.56 + 0.195 = 0.755. The overall score is calculated for all individuals in the offspring population, and the individual with the highest score is selected; its corresponding route is the optimal transportation route. In practical applications, safety and efficiency weights can be dynamically adjusted according to different time periods and weather conditions to adapt to the needs of different scenarios.

[0122] For example, in a liquefied natural gas (LNG) transportation mission, the initial candidate route set contains 20 routes. After multi-objective optimization, the resulting optimal solution set includes 5 routes located on the non-dominated front, each exhibiting different trade-offs between safety and efficiency. For instance, route A has a safety objective of 0.12 and an efficiency objective of 0.38; route B has a safety objective of 0.18 and an efficiency objective of 0.29; route C has a safety objective of 0.25 and an efficiency objective of 0.22; route D has a safety objective of 0.33 and an efficiency objective of 0.18; and route E has a safety objective of 0.42 and an efficiency objective of 0.15. Using preset weights of 0.7 for safety and 0.3 for efficiency, the comprehensive score was calculated as follows: Route A's comprehensive score was 0.7×(1-0.12)+0.3×(1-0.38)=0.616+0.186=0.802, and Route C's comprehensive score was 0.7×(1-0.25)+0.3×(1-0.22)=0.525+0.234=0.759. After comprehensive score calculation, Route A received the highest score and was selected as the optimal transportation route, reflecting the decision-making principle of prioritizing safety.

[0123] This invention presents a method for multi-objective optimization of multiple candidate transportation routes. By combining safety situation scores with route attributes to construct a multi-objective optimization function, and employing an improved genetic algorithm to solve the problem, a balanced optimization of safety and efficiency is achieved. The invention innovatively introduces the concept of gene position safety weights to guide crossover and mutation operations, making the algorithm's evolution direction more aligned with the safety requirements of liquefied natural gas transportation.

[0124] In one alternative implementation,

[0125] The steps include: calculating the safety weight similarity of adjacent gene loci based on the initial coding population; determining the crossover interval based on the safety weight similarity and adjusting the crossover probability based on the non-dominance level; calculating the safety weight deviation and fluctuation value of gene loci to determine the mutation location; and adjusting the mutation probability according to the number of iterations to perform the mutation operation.

[0126] Calculate the safety weight difference between adjacent gene loci in the gene sequence, divide the safety weight difference by the larger safety weight value among the adjacent gene loci to obtain the safety weight similarity, and determine the gene loci with the safety weight similarity greater than the dynamic similarity threshold and the gene locus distance less than the preset interval length constraint as the cross interval;

[0127] The crossover probability is obtained by adding the lower limit of the crossover probability and the crossover probability adjustment amount. The crossover probability adjustment amount is calculated by multiplying the difference between the upper limit of the crossover probability and the lower limit of the crossover probability by the non-dominated level adjustment coefficient. The non-dominated level adjustment coefficient is the ratio of the difference between the maximum non-dominated level and the current individual's non-dominated level to the maximum non-dominated level.

[0128] The ratio of the difference between the population average safety weight and the safety weight at the current position to the population average safety weight is used as the safety weight deviation value. The ratio of the fluctuation range of the safety weight at the current position to the maximum safety weight is used as the safety weight fluctuation value. The weighted sum of the safety weight deviation value and the safety weight fluctuation value is used as the gene locus evaluation score. Based on the gene locus evaluation score, the mutation position is selected, the mutation probability is adjusted according to the current iteration number, and the mutation operation is performed on the selected position.

[0129] For example, the safety weight difference between adjacent gene loci in a gene sequence is calculated, and the safety weight similarity is obtained by dividing the safety weight difference by the larger safety weight among the adjacent gene loci. This calculation method takes into account relative differences and is more in line with practical application needs. For example, if the safety weights of three consecutive gene loci in an individual's coding sequence are 0.82, 0.75, and 0.90, the safety weight difference for the first pair of adjacent gene loci is |0.82-0.75| = 0.07, and the larger safety weight is 0.82, resulting in a safety weight similarity of 0.07 / 0.82 = 0.085; the safety weight difference for the second pair of adjacent gene loci is |0.75-0.90| = 0.15, and the larger safety weight is 0.90, resulting in a safety weight similarity of 0.15 / 0.90 = 0.167. The dynamic similarity threshold is set to 0.2, which decreases as the number of iterations increases. The calculation method is: initial threshold 0.2 minus the current iteration number divided by the maximum iteration number multiplied by 0.15. The preset interval length constraint is set to 30% of the gene sequence length to ensure that the crossover operation does not lead to the exchange of excessively long segments. Gene positions with a safety weight similarity greater than the dynamic similarity threshold and a gene position spacing less than the preset interval length constraint are identified as crossover intervals. In the example above, if the dynamic similarity threshold is 0.18, the similarity of the first pair of adjacent gene positions (0.085) is less than the threshold and does not meet the condition; the similarity of the second pair of adjacent gene positions (0.167) is also less than the threshold and does not meet the condition. If there are adjacent gene positions in the sequence with a similarity greater than the threshold, these positions are selected as the boundaries of the crossover intervals.

[0130] The lower bound for crossover probability is set to 0.7, and the upper bound is set to 0.95, indicating that the crossover probability ranges from [0.7, 0.95]. If the maximum non-dominance level in the current population is 4, and an individual's non-dominance level is 1, then the non-dominance level adjustment coefficient for that individual is (4-1) / 4 = 0.75. The crossover probability adjustment amount is (0.95-0.7)×0.75 = 0.1875, and the crossover probability for that individual is 0.7 + 0.1875 = 0.8875. In contrast, for an individual with a non-dominance level of 3, the non-dominance level adjustment coefficient is (4-3) / 4 = 0.25, the crossover probability adjustment amount is (0.95-0.7)×0.25 = 0.0625, and the crossover probability is 0.7 + 0.0625 = 0.7625. This design allows individuals with higher non-dominance levels to have a greater crossover probability, promoting the transmission and combination of superior genes. Based on the calculated crossover probability, a random number between [0,1] is generated. If the random number is less than the crossover probability, the crossover operation is performed; otherwise, the parent individuals remain unchanged. The crossover operation selects two different crossover points, swaps the gene segments of the two parent individuals between these two points, and generates two new offspring individuals.

[0131] The average safety weight of a population refers to the average safety weight of all gene loci in all individuals within the current population. For example, if a population contains 100 individuals, each with 8 gene loci, the average safety weight of all 800 gene loci is calculated to be 0.80. The fluctuation range of the safety weight at the current position is defined as the absolute value of the difference between the safety weight at that position and the average safety weight of its two adjacent positions. If the safety weights of three consecutive gene loci of an individual are 0.82, 0.63, and 0.86, then the fluctuation range at the middle position is |(0.63-0.82)+(0.63-0.86)| / 2=0.21. The maximum safety weight refers to the maximum safety weight value of all gene loci in the current population, assumed to be 1.2. For the aforementioned middle position, its safety weight deviation is |0.80-0.63| / 0.80=0.2125, and the safety weight fluctuation is 0.21 / 1.2=0.175. With the deviation value weighted at 0.6 and the fluctuation value weighted at 0.4, the evaluation score for this locus is 0.6 × 0.2125 + 0.4 × 0.175 = 0.1975.

[0132] The higher the evaluation score of a gene locus, the greater the probability of it being selected as a mutation site. In the specific implementation, the evaluation scores of all gene loci are normalized into a probability distribution, and a roulette wheel method is used to select the mutation site.

[0133] The mutation probability is adjusted based on the current iteration count, and a mutation operation is performed on the selected position. The adjustment of the mutation probability follows the principle of decreasing with each iteration, calculated as the initial mutation probability minus the current iteration count divided by the maximum iteration count multiplied by the mutation probability decay coefficient. The initial mutation probability is set to 0.15, the mutation probability decay coefficient is set to 0.1, and the minimum mutation probability is no less than 0.05. For example, if the maximum iteration count is 100 and the current iteration is the 30th, then the current mutation probability is 0.15 - 30 / 100 × 0.1 = 0.12. A random number between [0,1] is generated. If the random number is less than the mutation probability, the mutation operation is performed; otherwise, the individual remains unchanged. The mutation operation targets the selected gene position and selects a different gene value from the candidate gene set to replace the original value. The candidate gene set contains other graph edges that have the same start and end points as the graph edge represented by the current gene position, or graph edge combinations that can form an effective bypass route. If a section of a highway in a liquefied natural gas transportation route is replaced by a national highway due to a change, it is necessary to check whether the replacement will cause the route to become disconnected or form a loop. Only legitimate replacements will be accepted.

[0134] For example, in the route optimization of a liquefied natural gas (LNG) transportation mission, the initial population contains 100 individuals, each with a code length of 12, representing a transportation route containing 12 graph edges. After the above calculations, the safety weight similarity of each gene locus is determined. For instance, the safety weight similarity between the 4th and 5th gene loci is 0.22, exceeding the dynamic similarity threshold of 0.18, and the distance between the two locations is 1, less than the preset interval length constraint of 3.6. Therefore, it is selected as the boundary of the crossover interval. For an individual with a non-dominated level of 2, under the condition of a maximum non-dominated level of 5, its crossover probability is 0.7 + (0.95 - 0.7) × (5 - 2) / 5 = 0.85. The gene locus evaluation score calculation shows that the 7th gene locus of this individual has the highest evaluation score of 0.28, indicating that the safety weight at this position deviates significantly from the population average and fluctuates considerably; therefore, it is selected as the mutation location. In the 50th iteration, the mutation probability is 0.15 - 50 / 100 × 0.1 = 0.1. The randomly generated number is 0.08, which is less than the mutation probability. Therefore, the mutation operation is performed, replacing the edge of the original graph with another path segment from the candidate set. The mutated individuals recalculate the safety target value and efficiency target value, merge them with the original population, and perform non-dominated sorting to select high-quality individuals to form a new generation of population.

[0135] Through 30 iterations, the population gradually converged, the solutions on the non-dominated front became more dense, and the distribution of safety and efficiency objective values ​​became more uniform. The final optimization results yielded multiple sets of non-dominated solutions with different trade-offs. For example, one set of solutions had a safety objective value range of [0.12, 0.42] and an efficiency objective value range of [0.15, 0.38], providing decision-makers with diverse choices. A specific optimization solution, with the encoding sequence [105, 221, 456, 387, 512, 698, 723, 845, 910, 1024, 1135, 1210], corresponds to a transportation route from a liquefied natural gas production base to a storage center, with a safety objective value of 0.18 and an efficiency objective value of 0.25, comprehensively considering the needs of both safety and efficiency.

[0136] This invention improves the crossover and mutation operations in genetic algorithms. It determines the crossover interval by calculating the similarity of safety weights of adjacent gene loci in the gene sequence, and dynamically adjusts the crossover probability based on the non-dominated level, allowing individuals with higher non-dominated levels to obtain a greater crossover probability, thus promoting the combination and transmission of high-quality genes. It innovatively introduces the concepts of safety weight deviation and fluctuation value to accurately identify locations with significant safety risk fluctuations in the route, enabling targeted mutation operations and effectively avoiding the problem of high randomness in mutation location selection in traditional genetic algorithms. Furthermore, it designs a mutation probability adjustment mechanism related to the number of iterations, achieving adaptive balance in the optimization process and significantly improving the convergence speed and solution quality of the algorithm.

[0137] In one alternative implementation,

[0138] The steps for calculating a comprehensive score based on a weighted average of the safety target value and the efficiency target value using preset weights include:

[0139] The route risk coefficient is calculated based on the liquefied natural gas loading capacity, container protection level, and distribution characteristics of emergency resources along the route. The product of the route risk coefficient and the safety target value is used as the corrected safety target value.

[0140] The corrected safety and efficiency target values ​​are normalized, the Euclidean distance between the target values ​​of individuals on the Pareto front is calculated, clustering is performed based on the distance threshold, and the cluster center is determined as the representative individual.

[0141] The main score is calculated for each representative individual based on the weights of the preset safety target value and efficiency target value. At the same time, candidate individuals are selected based on the road segment overlap rate and candidate scores are calculated. The weighted sum of the main score and the candidate scores is used as the comprehensive score of the individual.

[0142] The optimal representative individual and its corresponding alternative individuals are selected based on the comprehensive score, serving as the primary recommended route and alternative routes, respectively, and a dynamic switching threshold based on the safety target value is set.

[0143] For example, liquefied natural gas (LNG) loading capacity is divided into three levels: less than 20 tons is low-level, with a risk coefficient increment of 0.1; 20 to 50 tons is medium-level, with a risk coefficient increment of 0.2; and more than 50 tons is high-level, with a risk coefficient increment of 0.3. Container protection levels are divided into three levels: Level A indicates comprehensive protection measures, with a risk coefficient reduction of 0.15; Level B indicates basic protection measures, with a risk coefficient reduction of 0.05; and Level C indicates limited protection measures, with a risk coefficient increment of 0.05. The distribution characteristics of emergency resources along the route are measured by the average number of emergency rescue stations within every 10 kilometers along the route: 0.5 or more stations is high-density, with a risk coefficient reduction of 0.1; 0.2 to 0.5 stations is medium-density, with no change in risk coefficient; and less than 0.2 stations is low-density, with a risk coefficient increment of 0.1. The base value of the route risk coefficient is 1.0, adjusted based on the assessment results of the above three factors. For example, a transportation mission carries 45 tons of liquefied natural gas, which is considered a medium-volume load, with a risk factor increment of 0.2. Using Class B protective containers reduces the risk factor by 0.05. There are an average of 0.3 emergency rescue stations every 10 kilometers along the route, which is considered a medium-density load, and the risk factor remains unchanged. The risk factor for this route is 1.0 + 0.2 - 0.05 + 0 = 1.15. If the safety target value for this route is 0.18, then the revised safety target value is 0.18 × 1.15 = 0.207.

[0144] The corrected safety and efficiency target values ​​are mapped to the interval [0,1] to make the two dimensions comparable. Specifically, for each target value, the minimum value of that dimension is subtracted, and then divided by the range of that dimension (maximum value minus minimum value). For example, in the non-dominated solution set, the corrected safety target value range is [0.15, 0.42], and the efficiency target value range is [0.18, 0.38]. Then, the normalized corrected safety target value of 0.207 for an individual is (0.207 - 0.15) / (0.42 - 0.15) = 0.211, and the normalized efficiency target value of 0.25 is (0.25 - 0.18) / (0.38 - 0.18) = 0.35. The Euclidean distance between individuals on the Pareto front is calculated, which is the straight-line distance between two individuals in the normalized two-dimensional target space. A distance threshold of 0.15 is set. If the Euclidean distance between two individuals is less than this threshold, they are considered to belong to the same cluster. All individuals in the non-dominated solution set are clustered, and the individual closest to the cluster center is selected as the representative individual for each cluster. The cluster center refers to the average position of all individuals in the cluster within the target space. For example, if a cluster contains 5 individuals with normalized safety target values ​​of 0.21, 0.23, 0.19, 0.22, and 0.20, and efficiency target values ​​of 0.35, 0.33, 0.36, 0.34, and 0.37, then the cluster center position is (0.21, 0.35). The individual closest to this center (with the smallest Euclidean distance) will be the representative individual of that cluster.

[0145] A master score is calculated for each representative individual based on the weighted preset safety and efficiency target values. The preset weights reflect the decision-maker's preference for safety and efficiency. For high-risk operations such as liquefied natural gas transportation, the safety weight is usually set higher, such as 0.7 for safety and 0.3 for efficiency. The master score is calculated by multiplying the safety target value by the safety weight and then adding the efficiency target value multiplied by the efficiency weight. Since the target values ​​are normalized and smaller values ​​indicate better performance, 1 is subtracted from the target value in the actual calculation to ensure that a higher master score indicates a better solution. For example, if a representative individual has a normalized safety target value of 0.211 and a normalized efficiency target value of 0.35, its master score is 0.7 × (1 - 0.211) + 0.3 × (1 - 0.35) = 0.553 + 0.195 = 0.748. The road segment overlap rate refers to the proportion of the shared road segment length between two routes to the total road length. For each representative individual, individuals with a road segment overlap rate below the threshold of 0.6 are selected from the non-representative individuals as candidate individuals. If multiple candidate individuals meet the criteria, the one with the highest main score is chosen. The calculation of the candidate score is similar to that of the main score, but different weights are used, such as a safety weight of 0.6 and an efficiency weight of 0.4, reflecting the relatively higher efficiency requirements of the candidate routes. The comprehensive score is a weighted sum of the main score and the candidate scores, with the main score having a weight of 0.8 and the candidate scores having a weight of 0.2. For example, if a representative individual has a main score of 0.748 and its candidate individual has a candidate score of 0.712, then the comprehensive score is 0.8 × 0.748 + 0.2 × 0.712 = 0.5984 + 0.1424 = 0.7408.

[0146] A comprehensive score is calculated for all representative individuals, and the individual with the highest score is selected as the optimal representative individual, with its corresponding alternative individuals serving as alternative routes. A dynamic switching threshold is used during actual transportation to determine whether to switch from the primary recommended route to an alternative route based on real-time safety conditions. The switching threshold is set to 1.2 times the safety target value of the primary recommended route. For example, if the safety target value of the primary recommended route is 0.18, then the switching threshold is 0.18 × 1.2 = 0.216. During transportation, if the safety target value of the primary recommended route exceeds the switching threshold in real-time monitoring, a switch to an alternative route will be automatically suggested. The calculation of the real-time safety target value takes into account temporary factors such as current weather conditions, traffic flow, and road construction. For example, in rainy or snowy weather conditions, the road safety score will decrease by 20% to 30%; in traffic congestion, the safety score will decrease by 10% to 15%; and in road construction areas, the safety score will decrease by 25% to 35%. These temporary factors are multiplied by the basic safety score to obtain the real-time safety target value.

[0147] For example, for a liquefied natural gas (LNG) transportation mission with a loading capacity of 45 tons, using Class B protective containers, and an average of 0.3 emergency rescue stations per 10 kilometers along the route, the route risk coefficient is 1.15. The non-dominated solution set contains 20 solutions. After normalization and clustering, 5 representative individuals are obtained. Representative individual 1 has a normalized safety objective value of 0.211, an efficiency objective value of 0.35, and a principal score of 0.748; representative individual 2 has a normalized safety objective value of 0.306, an efficiency objective value of 0.243, and a principal score of 0.719; representative individual 3 has a normalized safety objective value of 0.108, an efficiency objective value of 0.486, and a principal score of 0.779; representative individual 4 has a normalized safety objective value of 0.423, an efficiency objective value of 0.172, and a principal score of 0.694; and representative individual 5 has a normalized safety objective value of 0.644, an efficiency objective value of 0.075, and a principal score of 0.618. To find alternative routes for representative individual 3, an individual with a road segment overlap rate of 0.42 was found, with a normalized safety target value of 0.195, an efficiency target value of 0.404, and a candidate score of 0.712. Representative individual 3's overall score is 0.8 × 0.779 + 0.2 × 0.712 = 0.7652, higher than the overall scores of other representative individuals. Therefore, the route corresponding to representative individual 3 is selected as the primary recommended route, and the routes corresponding to its alternative routes are selected as alternative routes. The safety target value of the primary recommended route is 0.165, and the switching threshold is 0.165 × 1.2 = 0.198. During transportation execution, if it is detected that the safety score of a segment of the primary recommended route drops by 25% due to rain or snow, the real-time safety target value will exceed the switching threshold, and a switch to an alternative route will be recommended to ensure transportation safety.

[0148] This invention provides a method for comprehensively evaluating multi-objective optimization results. By incorporating liquefied natural gas (LNG) loading capacity, container protection level, and the distribution characteristics of emergency resources along the route, the safety target value is corrected, making it more aligned with actual transportation risk assessment needs. The invention also designs a route recommendation mechanism that combines primary and backup options and a dynamic switching threshold based on the safety target value, providing decision support for flexibly responding to changes in real-time safety conditions and comprehensively improving the safety, reliability, and adaptability of LNG transportation scheduling schemes.

[0149] A second aspect of the present invention provides a liquefied natural gas transportation dispatching system based on maximizing safety, comprising:

[0150] The first unit is used to acquire multi-dimensional information of liquefied natural gas transport vehicles, use deep learning algorithms to extract features from the multi-dimensional information, extract temporal features through a temporal convolutional network, fuse feature weights from different data sources through an attention mechanism, generate a feature vector reflecting the safety status, and calculate a safety status score based on the feature vector.

[0151] The second unit is used to establish a dynamic graph structure of a road transport network, which includes nodes and road segments, each of which has weight information of traffic attributes; based on the dynamic graph structure of the road transport network, a graph neural network is used to calculate multiple candidate transport routes;

[0152] The third unit is used to perform multi-objective optimization on the multiple candidate transportation routes. The safety situation score and route attributes are used as optimization objectives. A genetic algorithm is used to solve the problem and obtain multiple sets of optimized solutions. The optimal transportation route is selected according to a preset weight. Transportation scheduling instructions are generated based on the optimal transportation route.

[0153] A third aspect of the present invention provides an electronic device, comprising:

[0154] processor;

[0155] Memory used to store processor-executable instructions;

[0156] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0157] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0158] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0159] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A liquefied natural gas transportation scheduling method based on maximizing safety, characterized in that, include: To obtain multidimensional information about liquefied natural gas (LNG) transport vehicles, a deep learning algorithm is used to extract features from this information. Temporal features are extracted using a temporal convolutional network, and feature weights from different data sources are fused using an attention mechanism to generate a feature vector reflecting the safety status. A safety status score is calculated based on this feature vector. Specifically, the process includes: segmenting the multidimensional information temporally using a sliding window; obtaining preprocessed data through normalization; extracting features from the preprocessed data using a temporal convolutional network, obtaining temporal features through causal convolution and multi-layer dilated convolution; constructing a modality correlation matrix based on the temporal features, where the modality correlation matrix represents the correlation strength between different modal features; inputting the modality correlation matrix into multiple parallel feature extraction channels with different receptive field sizes to obtain multi-scale feature representations; and constructing a temporal state memory unit to dynamically fuse the multi-scale feature representations with historical features and calculate the modality. Anomaly score; based on the modal anomaly score, the features are adaptively adjusted, and a dynamic weight matrix is ​​obtained by a softmax function with temperature coefficient adjustment; the temporal features are converted into query vector, key vector, and value vector through a self-attention mechanism, the dot product of the query vector and key vector is calculated and normalized, and then weighted and summed with the value vector to obtain the attention weight matrix; the dynamic weight matrix and the attention weight matrix are subjected to Hadamard product operation to obtain adaptive dynamic weight attention value; multi-head attention calculation is performed based on the adaptive dynamic weight attention value, and the calculation result is residually connected and layer normalized with the original features to obtain fused features; the fused features are processed by a multilayer perceptron to obtain feature weights, the multilayer perceptron containing multiple fully connected layers, and the feature weights are subjected to Hadamard product operation with the hidden layer features output by the intermediate layers of the multilayer perceptron, and a security situation score is obtained after nonlinear transformation; A dynamic graph structure for a road transport network is established, comprising nodes and road segments, each road segment having weight information related to traffic attributes; based on the dynamic graph structure of the road transport network, a graph neural network is used to calculate multiple candidate transport routes; Multi-objective optimization is performed on the multiple candidate transportation routes, taking the safety situation score and route attributes as optimization objectives. A genetic algorithm is used to solve the problem, resulting in multiple sets of optimized solutions. The optimal transportation route is selected according to preset weights. Transportation scheduling instructions are generated based on the optimal transportation route.

2. The method according to claim 1, characterized in that, The steps for establishing a dynamic graph structure for a road transport network, comprising nodes and road segments, with each road segment having weight information related to its traffic attributes, include: A dynamic graph structure for the road transport network is established, mapping road nodes to graph nodes and road segments to graph edges. The feature vectors of the graph nodes include the node's geographical coordinates, node type, and node capacity. The initial weights of the graph edges are calculated using road segment length, average vehicle speed, and traffic flow. Real-time traffic flow data and historical statistical data of the road transport network are obtained, and spatiotemporal alignment and standardization are performed to obtain standardized data; The standardized data is combined with the feature vectors of the graph nodes to calculate the dynamic association strength between the graph nodes. The dynamic association strength is calculated using a learnable parameter matrix and the LeakyReLU activation function to obtain a temporally enhanced node representation. The spatial correlation degree between the graph nodes is calculated based on the spatial query matrix and the spatial key-value matrix. Each row of the spatial query matrix represents the query vector of a node, and each row of the spatial key-value matrix represents the key vector of a node. The size of both the spatial query matrix and the spatial key-value matrix is ​​the number of nodes multiplied by the feature dimension. The spatial correlation degree is then fused with the temporally enhanced node representation to obtain the spatiotemporal fusion feature. The spatiotemporal fusion features between connected graph nodes are extracted, and a weight adjustment factor is obtained through fully connected network processing. The initial weights are multiplied by the weight adjustment factor to obtain the updated graph edge weights. The deviation score between the updated graph edge weights and the historical graph edge weights is calculated. When the deviation score exceeds a preset deviation threshold, a weight decay factor is calculated based on the deviation score. The updated graph edge weights are adjusted using the weight decay factor to obtain the adjusted graph edge weights.

3. The method according to claim 2, characterized in that, Based on the dynamic graph structure of the road transport network, the steps of calculating multiple candidate transport routes using a graph neural network include: Based on the feature vectors of the graph nodes and the adjusted graph edge weights, feature aggregation and transmission are performed to obtain global features; The dynamic association weights between nodes are calculated by the similarity between the query vector and the key vector of the node. The global features are then enhanced based on the dynamic association weights to obtain enhanced features. Construct a gated graph convolutional unit including update gates and reset gates, calculate the parameters of update gates and reset gates based on the enhanced features and historical features, and iteratively update the enhanced features using the parameters of update gates and reset gates to obtain a road network representation with temporal memory capability; An initial transportation route is generated based on the road network representation. The road network representation vectors of two nodes connected by graph edges in the initial transportation route are extracted. The cosine similarity between the two road network representation vectors is calculated as a compatibility score. The compatibility score is multiplied by the safety coefficient of the graph edge to obtain the temporal memory feature score. The product of the temporal memory feature score and the position weight of each graph edge in the initial transportation route is calculated to obtain the graph edge weighted score. The weighted scores of all graph edges in the initial transportation route are accumulated to obtain a comprehensive score. The transportation route with the highest comprehensive score is selected as the benchmark transportation route. The baseline transportation route is branched and expanded. Based on a preset expansion depth parameter, the graph edges in the baseline transportation route are replaced and searched to generate multiple candidate transportation routes.

4. The method according to claim 1, characterized in that, Multi-objective optimization is performed on the multiple candidate transportation routes, using the safety situation score and route attributes as optimization objectives. A genetic algorithm is used to solve the problem, obtaining multiple sets of optimized solutions. The step of selecting the optimal transportation route according to preset weights includes: The multiple candidate transportation routes are mapped to gene coding sequences, where the graph edges in each route are mapped to gene positions in the gene sequence, and the security weight of each gene position is calculated based on the security situation score. A multi-objective optimization function is constructed, which combines the safety situation score and the graph edge weights in the dynamic graph structure of the road transport network to obtain the safety objective value, and combines the time consumption and distance consumption of the route to obtain the efficiency objective value. An initial coding population of size N is generated. The safety objective value and efficiency objective value of each individual in the population are calculated as fitness values. The initial coding population is then sorted non-dominated based on the multi-objective optimization function. Based on the initial encoded population, calculate the safety weight similarity of adjacent gene loci, determine the crossover interval according to the safety weight similarity, and adjust the crossover probability based on the non-dominated level; calculate the safety weight deviation value and fluctuation value of gene loci to determine the mutation position, adjust the mutation probability according to the number of iterations to perform mutation operation; recalculate the safety target value and efficiency target value for the mutated individuals, perform non-dominated sorting and elite retention, and obtain the offspring population; For individuals in the offspring population, a comprehensive score is calculated by weighting the safety target value and the efficiency target value based on preset weights, and the transportation route corresponding to the individual with the highest comprehensive score is selected as the optimal transportation route.

5. The method according to claim 4, characterized in that, The safety weight similarity of adjacent gene loci is calculated based on the initial coding population, the crossover interval is determined according to the safety weight similarity, and the crossover probability is adjusted based on the non-dominance level. The steps involved in calculating the safety weight bias and fluctuation value of gene loci to determine the mutation location, and adjusting the mutation probability based on the number of iterations to perform the mutation operation, include: Calculate the safety weight difference between adjacent gene loci in the gene sequence, divide the safety weight difference by the larger safety weight value among the adjacent gene loci to obtain the safety weight similarity, and determine the gene loci with the safety weight similarity greater than the dynamic similarity threshold and the gene locus distance less than the preset interval length constraint as the cross interval; The crossover probability is obtained by adding the lower limit of the crossover probability and the crossover probability adjustment amount. The crossover probability adjustment amount is calculated by multiplying the difference between the upper limit of the crossover probability and the lower limit of the crossover probability by the non-dominated level adjustment coefficient. The non-dominated level adjustment coefficient is the ratio of the difference between the maximum non-dominated level and the current individual's non-dominated level to the maximum non-dominated level. The ratio of the difference between the population average safety weight and the safety weight at the current position to the population average safety weight is used as the safety weight deviation value. The ratio of the fluctuation range of the safety weight at the current position to the maximum safety weight is used as the safety weight fluctuation value. The weighted sum of the safety weight deviation value and the safety weight fluctuation value is used as the gene locus evaluation score. Based on the gene locus evaluation score, the mutation position is selected, the mutation probability is adjusted according to the current iteration number, and the mutation operation is performed on the selected position.

6. The method according to claim 4, characterized in that, The steps for calculating a comprehensive score based on a weighted average of the safety target value and the efficiency target value using preset weights include: The route risk coefficient is calculated based on the liquefied natural gas loading capacity, container protection level, and distribution characteristics of emergency resources along the route. The product of the route risk coefficient and the safety target value is used as the corrected safety target value. The corrected safety and efficiency target values ​​are normalized, the Euclidean distance between the target values ​​of individuals on the Pareto front is calculated, clustering is performed based on the distance threshold, and the cluster center is determined as the representative individual. The main score is calculated for each representative individual based on the weights of the preset safety target value and efficiency target value. At the same time, candidate individuals are selected based on the road segment overlap rate and candidate scores are calculated. The weighted sum of the main score and the candidate scores is used as the comprehensive score of the individual. The optimal representative individual and its corresponding alternative individuals are selected based on the comprehensive score, serving as the primary recommended route and alternative routes, respectively, and a dynamic switching threshold based on the safety target value is set.

7. A liquefied natural gas transportation and dispatching system based on maximizing safety, used to implement the method of any one of claims 1-6, characterized in that, include: The first unit is used to acquire multi-dimensional information of liquefied natural gas transport vehicles, use deep learning algorithms to extract features from the multi-dimensional information, extract temporal features through a temporal convolutional network, fuse feature weights from different data sources through an attention mechanism, generate a feature vector reflecting the safety status, and calculate a safety status score based on the feature vector. The second unit is used to establish a dynamic graph structure of a road transport network, which includes nodes and road segments, each of which has weight information of traffic attributes; based on the dynamic graph structure of the road transport network, a graph neural network is used to calculate multiple candidate transport routes; The third unit is used to perform multi-objective optimization on the multiple candidate transportation routes. The safety situation score and route attributes are used as optimization objectives. A genetic algorithm is used to solve the problem and obtain multiple sets of optimized solutions. The optimal transportation route is selected according to a preset weight. Transportation scheduling instructions are generated based on the optimal transportation route.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images