Bus carbon efficiency intervention evaluation method and device driven by large model semantics, medium

Abstract

The application discloses a bus carbon efficiency intervention evaluation method and device driven by a large model, and a medium, comprising: fine-tuning a large language model by using bus multi-source heterogeneous text data, and constructing line-level, station-level and section-level prompt words based on line structural features; inputting the fine-tuned model to generate corresponding semantic embedding vectors, and obtaining semantic features through fusion. Further, the line structural features and the intervention strategy are spliced into joint features, and then fused with the semantic features to form a first embedding vector, which is input into a trained deep causal network to predict the carbon efficiency estimate value after the intervention is implemented. Similarly, the features without implementing the strategy are fused to obtain a second embedding vector, which is input into the same network to obtain the carbon efficiency estimate value after the intervention is not implemented. Finally, by comparing the carbon efficiency improvement rates in the two scenarios, the quantitative evaluation of the effect of the intervention measure is realized.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of urban traffic carbon efficiency, and particularly relates to a bus carbon efficiency intervention evaluation method driven by a large model semantic, a device and a medium. BACKGROUND

[0002] With the acceleration of urbanization and the deepening of the concept of green and low carbon, the public transportation system, as an important part of urban transportation, plays a key role in carbon emission reduction and sustainable development. Unlike the traditional evaluation method focusing on the total amount of carbon emissions, carbon efficiency (i.e. the amount of carbon emissions corresponding to unit passenger mileage) has become an important technical index for measuring the green operation level of the public transportation system, and its improvement is of great significance for promoting the energy efficiency optimization and environmental friendliness of the urban transportation system. However, the influencing factors of public transportation carbon efficiency are extremely complex, including not only structured characteristics such as road facilities, operation strategies, population distribution, but also a large amount of unstructured information (such as historical events, user feedback, etc.).

[0003] The factors affecting the carbon efficiency of public transportation are highly complex:

[0004] 1. Structured characteristic explicit factors: including road infrastructure conditions (such as exclusive lane coverage, intersection signal timing), vehicle technical parameters (energy type, energy consumption level), operation scheduling strategies (departure frequency, line network density), population distribution and commuting characteristics, etc. quantifiable and coded variables.

[0005] 2. Unstructured information implicit factors: covering historical operation event records (such as major events, traffic accidents, temporary road closures, etc.), public information and user feedback (social media reviews, complaint and suggestion texts), descriptive content in management and operation reports, etc. information with rich semantics but difficult to structure directly.

[0006] Currently, the effect evaluation methods of intervention measures for the influencing factors of public transportation carbon efficiency mainly have the following fundamental defects:

[0007] 1. Single data modality, ignoring unstructured semantic information:

[0008] Existing methods (such as structural equation models, multiple regression analysis, models based on traditional traffic simulation or macro statistics) rely too much on structured feature data for modeling. The deep semantic features (e.g. historical events, user feedback, etc.) contained in unstructured information are completely ignored or only simplified by manually extracting a very limited number of labels, resulting in a serious lack of understanding of system complexity.

[0009] 2. Unable to effectively handle unobserved confounding variables, severe estimation bias:

[0010] There are a large number of mixed factors (such as sudden events, user behavior changes, etc.) in actual public transport operation that are difficult to observe or quantify. When only relying on structured features for modeling, these unobserved mixed variables cannot be included in the model for control. Their influence will be incorrectly attributed to the intervention variables or structured features being studied, resulting in systematic bias (confounding bias) in the causal effect estimate of the actual carbon efficiency benefit of the intervention measures (such as the establishment of a dedicated lane, adjustment of ticket prices), making it difficult to fully reflect the true complexity of the system, introducing bias in the causal effect estimate, and thus weakening the accuracy of causal inference, seriously weakening the credibility of the evaluation conclusion.

[0011] In summary, the current evaluation method for public transport carbon efficiency intervention measures is limited to structured features and lacks the use of key non-structured semantic information. Moreover, there is a lack of effective means to control unobserved confounding variables, resulting in poor accuracy and insufficient interpretability of the evaluation of the effectiveness of public transport carbon efficiency intervention measures. SUMMARY

[0012] To address the deficiencies of the prior art, embodiments of the present application provide a large model semantic driven public transport carbon efficiency intervention evaluation method, device and medium.

[0013] In a first aspect, embodiments of the present application provide a large model semantic driven public transport carbon efficiency intervention evaluation method, the method comprising:

[0014] Fine-tuning a large language model through public transport historical unstructured text data;

[0015] Obtaining public transport line structured features and constructing line-level, station-level and section-level prompt words based thereon; inputting the prompt words into the fine-tuned large language model to obtain semantic embedding vectors corresponding to the line, station and section; and fusing the semantic embedding vectors corresponding to the line, station and section to obtain semantic features;

[0016] Concatenating the public transport line structured features and the intervention strategy to obtain a joint feature representation; fusing the joint feature representation and the semantic features to obtain a first embedding vector; and inputting the first embedding vector into a trained deep causal network to obtain a carbon efficiency estimate value when the intervention strategy is implemented;

[0017] Fusing the public transport line structured features and the semantic features to obtain a second embedding vector; and inputting the second embedding vector into the trained deep causal network to obtain a carbon efficiency estimate value when the intervention strategy is not implemented;

[0018] Quantifying the improvement rate of the carbon efficiency estimate values when the intervention strategy is implemented and when the intervention strategy is not implemented to evaluate the intervention measures.

[0019] In a second aspect, embodiments of the present application provide an electronic device comprising:

[0020] at least one processor; and

[0021] a memory communicatively connected with the at least one processor; wherein,

[0022] the memory stores one or more computer programs executable by the at least one processor, and the one or more computer programs are executed by the at least one processor to enable the at least one processor to perform the above-mentioned large model semantic driven bus carbon efficiency intervention evaluation method.

[0023] In a third aspect, an embodiment of the present application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-mentioned large model semantic driven bus carbon efficiency intervention evaluation method.

[0024] In a fourth aspect, an embodiment of the present application provides a computer program product comprising computer programs / instructions, wherein the computer programs / instructions, when executed by a processor, implement the above-mentioned large model semantic driven bus carbon efficiency intervention evaluation method.

[0025] Compared with the prior art, the present application has the following beneficial effects:

[0026] The present application provides a large model semantic driven bus carbon efficiency intervention evaluation method, which fuses structured features of bus lines and semantic features automatically mined by a large model, uses a deep causal network to predict the fusion features to obtain carbon efficiency estimation values when intervention strategies are implemented and not implemented, and quantifies the improvement rate of the carbon efficiency estimation values when the intervention strategies are implemented and not implemented, thereby realizing evaluation of the intervention measures. The present application fully utilizes the synergistic advantages of multi-modal data modeling and causal reasoning, and not only overcomes the limitations of existing research on data types, but also significantly reduces the interference of confounding variables, improves the accuracy and interpretability of the effect evaluation of bus carbon efficiency intervention measures, and provides support for urban green transportation construction. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0028] Figure 1 is a flowchart of the large model semantic driven bus carbon efficiency intervention evaluation method provided by the embodiments of the present application;

[0029] Figure 2This is a technical roadmap of the large-model semantic-driven public transport carbon efficiency intervention assessment method provided in the embodiments of the present invention;

[0030] Figure 3 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0031] 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.

[0032] It should be noted that, unless otherwise specified, the features in the following embodiments and implementation methods can be combined with each other.

[0033] In existing technologies, the estimation of the causal effects of carbon efficiency in urban public transport systems mainly relies on modeling based on structured features (such as operational data and road network indicators), neglecting the potential impact of semantic features extracted from unstructured data (such as user feedback and historical events) on carbon efficiency. This makes it difficult to fully reflect the complexity of actual operation, especially when there are a large number of unobserved confounding variables, which can easily lead to estimation bias and weaken the accuracy of causal inference.

[0034] To address this, this invention proposes a large-model semantic-driven method for assessing public transport carbon efficiency intervention, deeply fusing structured features with semantic features mined from a large language model. Through multimodal modeling and causal inference, it enhances the ability to identify the causal effects of public transport carbon efficiency and reduces the interference of confounding variables. Figure 1 and Figure 2 As shown, the specific steps are as follows:

[0035] Step S1: Fine-tune the large language model using multi-source heterogeneous unstructured text data from public transportation.

[0036] In this example, large language models such as Qwen2-7B, Qwen3-8B, GLM-4-9B-0414, and DeepSeek-R1-0528-Qwen3-8B can be selected.

[0037] Collect historical unstructured text data related to urban public transport, including but not limited to: historical traffic events (such as congestion, accidents, construction, extreme weather); public transport user complaints, comments, suggestions and feedback; social media information such as news reports, microblogs, and forums; and semi-structured text such as traffic announcements.

[0038] The original text is cleaned and denoised using rule-based and weakly supervised algorithms, and a fine-tuning dataset is constructed using an autoregressive language method. The LoRA (Low-Rank Adaptation) method is employed to fine-tune only the weights within the large language model that are strongly related to traffic semantics (such as the attention layer and projection layer in the Transformer structure). Except for the specified fine-tuning layer, all other parameters are kept frozen to reduce computational resource consumption.

[0039] Step S2: Obtain the structured features of bus routes, and use them to construct route-level prompt words, station-level prompt words, and segment-level prompt words; input them into the fine-tuned large language model to obtain the semantic embedding vectors corresponding to routes, stations, and segments; fuse the semantic embedding vectors corresponding to routes, stations, and segments to obtain semantic features.

[0040] Step S201, the process of obtaining the structured features of bus routes and constructing route-level prompts, station-level prompts, and segment-level prompts based on these features, includes:

[0041] For each bus route, relevant structured data is collected, including but not limited to: bus smart card swipe records, mobile phone signaling data, road network topology, and Points of Interest (POI) information. Based on the above data, for each bus route... Extract key structured features and concatenate them to form the structured features of bus routes. .

[0042] The structural features of the bus routes Including but not limited to: indicators such as private car density, energy type, road type, station density, GDP density, land mix, and congestion index, to ensure that the operational characteristics and regional attributes of public transport routes are fully characterized.

[0043] Based on the structural characteristics of bus routes The system automatically concatenates information such as route name, up and down line station sequence, station latitude and longitude, and road segment sequence to generate semantic text, thereby constructing route-level, station-level, and road segment-level prompts. Furthermore, this example can be combined with a Retrieval-Augmented Generation (RAG) system to fully utilize unstructured data related to the real-time public transportation system, achieving efficient integration of the latest developments, emergencies, and user feedback. Specifically, for each bus route, station, or road segment, the system not only injects knowledge based on historical corpora but also dynamically retrieves the latest text information from channels such as traffic announcements, user complaints, social media, and news, constructing a multi-source knowledge base containing both historical and real-time data. Examples of route-level, station-level, and road segment-level prompts in this example are as follows:

[0044] Line-level prompt: "Please generate a comprehensive line carbon efficiency impact analysis report based on the following comprehensive information of the bus line. The content should cover, but not limited to, the following aspects:

[0045] 1. Historical road condition events (such as congestion, accidents, construction, extreme weather) and their impact on bus operation carbon efficiency;

[0046] 2. User feedback such as complaints, comments, suggestions from passengers on this line, focusing on energy saving and emission reduction, comfort, speed, full load rate, etc.

[0047] 3. Information summary of social media such as news reports, microblogs, forums about this line, especially topics related to environmental protection, green travel, public transportation service optimization, etc.

[0048] 4. Relevant traffic announcements and their guidance or constraint measures for improving public transportation carbon efficiency (such as new energy bus promotion, priority passage strategy, etc.);

[0049] 5. The impact of road infrastructure, signal system, dedicated lane settings along the line on public transportation carbon emissions;

[0050] 6. The impact of vehicle types (such as traditional fuel / new energy), scheduling strategies, interval, passenger load rate, etc. on carbon efficiency;

[0051] 7. The main carbon emission challenges faced by the line and potential measures and suggestions to improve carbon efficiency.

[0052]

Line Information

[0053] Bus line name: Route 1

[0054] Upbound station sequence: D Hub Station --> F Station --> … --> G Station --> H Station

[0055] Upbound route sequence: E Expressway --> I Road --> … --> J Road --> K Street

[0056] Downbound station sequence: H Station --> G Station --> … --> L Station --> D Hub Station

[0057] Upbound route sequence: K Street --> J Road --> … --> J Road --> E Expressway

[0058]

Optional Supplementary Information

[0059] Station-level prompt: "Please generate a station and surrounding area carbon efficiency impact analysis report based on the following detailed information of the bus station. The content should cover, but not limited to:

[0060] 1. The impact of historical traffic events (e.g. congestion, accidents, construction, extreme weather) at the site and surrounding area on public transportation carbon efficiency;

[0061] 2. Passenger complaints, comments, suggestions about the site, focusing on issues related to waiting environment, transfer convenience, site congestion, energy consumption, etc.;

[0062] 3. Summary of information about the site and surrounding area from social media, news reports, etc., especially discussions related to green travel, energy saving and emission reduction, etc.;

[0063] 4. Traffic announcements related to the site, especially those related to energy saving, public transportation priority, site renovation, etc.;

[0064] 5. The positive and negative impact of transportation hubs, subways, transfer facilities, and non-motorized lanes near the site on public transportation carbon efficiency;

[0065] 6. Factors affecting the efficiency, energy consumption, and scheduling of public transportation vehicles at the site;

[0066] 7. Suggestions for improving the carbon efficiency of public transportation at the site and surrounding area.

[0067]

Site Information

[0068] Site Name: D Hub Station

[0069] Location: A City, B District, C Street

[0070] Latitude and Longitude: (116.4906, 39.9056)

[0071] Nearby Places: 0.3 km north: Subway D Station; 0.4 km east: D Long-distance Bus Station

[0072]

Optional Supplementary Information

[0073] Road Section Level Prompt: "Please analyze the carbon efficiency impact of the following road sections based on the detailed information provided. The content should include but not limited to:

[0074] 1. The specific impact of historical traffic events (e.g. congestion, accidents, construction, extreme weather) on public transportation vehicle energy consumption and carbon emissions on this road section;

[0075] 2. Passenger or driver feedback and suggestions about this road section, focusing on traffic flow, public transportation priority, road conditions, etc.;

[0076] 3. Discussions on the road section in social media such as news, microblog, forum, etc., especially on green transportation and low-carbon travel;

[0077] 4. Traffic announcements related to the road section, such as the impact of bus lane setting, traffic restriction measures, etc. on carbon efficiency;

[0078] 5. The role of road section infrastructure such as road type, congestion status, signal setting, slope, etc. on bus energy consumption and carbon emissions;

[0079] 6. The impact of bus operation characteristics on carbon efficiency, such as the type of bus on the road section, driving frequency, average speed, etc.;

[0080] 7. Optimization suggestions for improving the carbon efficiency of buses on the road section.

[0081]

Road section information

[0082] Road section name: E Expressway

[0083] Along the area: A city B1 district C1 street, A city B2 district C2 street

[0084] Road type: Urban expressway

[0085] Congestion status: Smooth

[0086]

Optional supplementary information

[0087] Step S202, input the line-level prompt word, station-level prompt word and road section-level prompt word into the fine-tuned large language model respectively, to obtain the hidden state of all tokens in the last layer of the large language model corresponding to the line level , the hidden state of all tokens in the last layer of the large language model corresponding to the station level , the hidden state of all tokens in the last layer of the large language model corresponding to the road section level .

[0088] Step S203, based on the hidden state of all tokens in the last layer of the large language model corresponding to the line level, station level and road section level, and the total number of tokens output by the last layer of the large language model corresponding to the line level, station level and road section level, based on the average pooling strategy, calculate the semantic embedding vector corresponding to the line, station and road section.

[0089]

[0090]

[0091]

[0092] In the formula, , , The total number of tokens in the last hidden layer output of the large language model corresponding to the line level, site level, and section level, , , The semantic embedding vectors corresponding to the bus line, site, and section, respectively.

[0093] It should be noted that the semantic embedding vectors corresponding to the bus line, site, and section comprehensively reflect multi-dimensional information such as geographic semantics, user perception, and historical events. To improve the stability and consistency of semantic embedding, the same input can be inferred multiple times and the output results can be averaged, or the temperature parameter of the model can be set to zero to reduce the randomness and output volatility of model inference, ensuring that the generated embedding vectors have high representativeness and reproducibility.

[0094] In step S204, the semantic embedding vectors corresponding to the sites are averaged, and the semantic embedding vectors corresponding to the sections are averaged. The expressions are as follows:

[0095]

[0096]

[0097] In the formula, and are the number of sites and the number of sections of the line , respectively. The dimensions of are .

[0098] In step S205, the semantic embedding vectors corresponding to the line , the average semantic embedding vectors corresponding to the sites , and the average semantic embedding vectors corresponding to the sections are fused to obtain the semantic features. The expression is as follows:

[0099]

[0100] In the formula, denotes the vector concatenation operation. The multi-level semantic features of the line, site, and section are integrated to provide high-dimensional and expressive input features for subsequent causal reasoning and carbon efficiency evaluation.

[0101] Step S3, splice the bus line structured feature and the intervention strategy to obtain a joint feature representation; fuse the joint feature representation and the semantic feature to obtain a first embedding vector; input the first embedding vector into the trained deep causal network to obtain a carbon efficiency estimation value when the intervention strategy is implemented;

[0102] Fuse the bus line structured feature and the semantic feature to obtain a second embedding vector; input the second embedding vector into the trained deep causal network to obtain a carbon efficiency estimation value when the intervention strategy is not implemented;

[0103] By quantifying the improvement rate of the carbon efficiency estimation value when the intervention strategy is implemented and not implemented, the evaluation of the intervention measure is realized.

[0104] The expression is as follows:

[0105]

[0106] In the formula, denotes the carbon efficiency estimation value of a given line when the intervention strategy is implemented, denotes the carbon efficiency estimation value of a given line when the intervention strategy is not implemented.

[0107] It should be noted that in this example, the actual benefit of different intervention strategies is represented by quantifying the improvement rate of the carbon efficiency estimation value when the intervention strategy is implemented and not implemented, which assists in the carbon efficiency optimization decision of the urban public transport system. In addition, the semantic features mined by the large language model contain knowledge such as traffic priors, historical patterns, user behaviors and social information, which improves the accuracy of inverse propensity score modeling and causal effect estimation, and reduces the bias caused by unobserved confounding variables.

[0108] Further, in this example, the process of training the deep causal network includes:

[0109] Step S100, splice the bus line structured feature and the intervention strategy to obtain a joint feature representation.

[0110] The bus line structured feature is spliced with the intervention strategy to form a joint feature representation , which is input into the effect structure encoder for encoding, and the expression is as follows:

[0111]

[0112] ​​Further, the intervention strategy includes: bus lane setting, priority control of bus signal lights, vehicle scheduling scheme optimization and departure interval optimization.

[0113] At step S200, the encoded joint feature representation and the encoded semantic feature are fused to obtain an effect embedding representation; and the effect embedding representation is input into a deep causal network to obtain a carbon efficiency estimate value.

[0114] Further, encoding the semantic feature includes: using a large language model to mine the semantic feature Inputting the effect semantic encoder Dimensionality reduction processing is performed, and the expression is as follows:

[0115]

[0116] To fully exploit the complementarity of structured features and semantic features of bus routes, a gated multi-modal fusion mechanism is used to fuse the encoded joint feature representation and the encoded semantic feature into a unified effect embedding representation, specifically:

[0117]

[0118]

[0119] wherein, is a gating weight vector, represents a Sigmoid activation function, represents a Hadamard product, represents a weight matrix, represents a bias term, represents an effect embedding representation.

[0120] Further, in this example, a deep causal network (DCN) is constructed based on the idea of multi-task learning. The backbone network of the deep causal network adopts a residual multilayer perceptron (RMLP), and outputs through an individual effect estimate branch, an average processing effect branch and a global trend branch, respectively.

[0121] Specifically, the process of inputting the effect embedding representation into the deep causal network to obtain the carbon efficiency estimate value includes:

[0122] The effect embedding representation is input into the backbone network in the deep causal network to obtain a deep embedding vector; the expression is as follows:

[0123]

[0124] wherein, denotes a deep embedding vector;

[0125] The deep embedding vector is output respectively through an individual effect estimation branch, an average treatment effect branch and a global trend branch, to obtain an individual effect estimation value, an average treatment effect estimation value and a global trend effect estimation value; the expressions are as follows:

[0126]

[0127]

[0128]

[0129] In the formula, denotes a bus line an individual effect estimation value, denotes a bus line an average treatment effect estimation value, denotes a bus line a global trend effect; is an individual effect estimation layer, used to depict the individualized intervention effect of each line; is an average treatment effect layer, reflecting the overall intervention effect of all lines in the same city; is a global trend modeling layer, combining a time variable to capture the dynamic change trend of carbon efficiency.

[0130] The individual effect estimation value, the average treatment effect estimation value and the global trend effect estimation value are weighted and summed to obtain a carbon efficiency estimation value, and the expression is as follows:

[0131]

[0132] In the formula, , , is a weight coefficient.

[0133] Step S300, setting a joint loss function, the joint loss function is a weighted sum of inverse propensity weighted mean square error loss and adversarial regularization loss; wherein the inverse propensity weighted mean square error loss is calculated based on inverse propensity weight, the number of bus lines, carbon efficiency estimation value and carbon efficiency real value; the inverse propensity weight is set through a multi-branch deep propensity network.

[0134] Wherein, the expression of the joint loss function is as follows:

[0135]

[0136] In the formula, denotes a joint loss function; represents the inverse propensity weighted mean squared error loss, aiming to correct the bias of causal effect estimation due to confounding variables or sampling bias; represents the regularization coefficient; represents the adversarial regularization loss, which discriminates the difference between the distribution of carbon efficiency estimation value and the distribution of carbon efficiency true value through KL divergence, and is used to improve the robustness of the deep causal network to unobserved confounding variables.

[0137] wherein the expression of the inverse propensity weighted mean squared error loss is as follows:

[0138]

[0139] In the formula, is the carbon efficiency true value, represents the carbon efficiency estimation value; is the number of bus routes in the city . represents the inverse propensity weight.

[0140] wherein the expression of the adversarial regularization loss is as follows:

[0141]

[0142] In the formula, is the KL divergence, and are the observed distribution and the model distribution respectively.

[0143] Further, in the present example, in order to quantify the improvement of the semantic features of the large model on the accuracy of the DCN model, the following improvement rate index is defined:

[0144]

[0145] In the formula, represents the relative improvement range of the mean squared error of carbon efficiency prediction after introducing the semantic features of the large model, is the carbon efficiency estimation result using only the structured features of the bus route.

[0146] Further, the obtaining process of the inverse propensity weight comprises:

[0147] a probabilistic structure encoder based on a probabilistic multi-layer perceptron (PMLP) to the structured features of the bus route dimensionality augmentation;

[0148] a probabilistic semantic encoder based on a multi-head self-attention mechanism to the semantic features dimension reduction;

[0149] The structured features of the bus line after dimension increase and the semantic features after dimension reduction are fused by adopting a gated multi-modal fusion mechanism to obtain a multi-modal embedding vector ;

[0150] The multi-modal embedding vector is input into a multi-branch deep propensity network (MDPN) to obtain the outputs of a structure branch, a semantic branch and a global fusion branch, the outputs of the structure branch, the semantic branch and the global fusion branch are fused by adopting attention weighting, and after being processed by a Softmax activation function, a probability distribution of implementing the intervention measure is obtained ; the reciprocal of the probability distribution is taken as the inverse propensity weight .

[0151] In step S400, the deep causal network and the multi-branch deep propensity network are collaboratively trained by a joint loss function.

[0152] In summary, the application provides a bus carbon efficiency intervention evaluation method driven by a large model semantic, which fuses structured features of a bus line and semantic features automatically mined by a large model, predicts the fused features by a deep causal network to obtain carbon efficiency estimation values when intervention strategies are implemented and not implemented, and evaluates the intervention measures by quantifying the improvement rate of the carbon efficiency estimation values when the intervention strategies are implemented and not implemented. The application fully utilizes the synergistic advantages of multi-modal data modeling and causal reasoning, overcomes the limitations of existing research on data types, significantly reduces the interference of confounding variables, improves the accuracy and interpretability of the evaluation of the effect of bus carbon efficiency intervention measures, and provides support for urban green transportation construction.

[0153] Correspondingly, the application also provides an electronic device, which includes one or more processors, a memory for storing one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the large model semantic driven bus carbon efficiency intervention evaluation method as described above. Figure 3 As shown in FIG. 1, a hardware structure diagram of the large model semantic driven bus carbon efficiency intervention evaluation method provided by the embodiment of the application is in any device with data processing capability. In addition to the processor, the memory and the network interface shown in FIG. 1, any device with data processing capability in which the device in the embodiment is usually provided according to the actual function of the device with data processing capability, and can also include other hardware, which will not be described here. Figure 3

[0154] ​Correspondingly, the application further provides a computer readable storage medium, which stores computer instructions, and the instructions are executed by a processor to implement the above-mentioned large model semantic driven bus carbon efficiency intervention evaluation method. The computer readable storage medium can be an internal storage unit of any device with data processing capability, such as a hard disk or a memory. The computer readable storage medium can also be an external storage device, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. Further, the computer readable storage medium can include both the internal storage unit of any device with data processing capability and the external storage device. The computer readable storage medium is used to store the computer program and other programs and data required by the device with data processing capability, and can also be used to temporarily store data that has been output or will be output.

[0155] The above embodiments are only used to illustrate the design ideas and characteristics of the present application, and the purpose is to enable those skilled in the art to understand the present application and implement it, and the protection scope of the present application is not limited to the above-mentioned embodiments. Therefore, any equivalent changes or modifications made according to the principles and design ideas disclosed by the present application are within the protection scope of the present application.

Claims

1. A large-scale model semantic-driven method for evaluating public transport carbon efficiency through intervention, characterized in that, The method includes: Fine-tuning of a large language model using multi-source heterogeneous unstructured text data from public transportation; The structured features of bus routes are obtained, and route-level prompt words, station-level prompt words, and segment-level prompt words are constructed based on them. These are then input into a fine-tuned large language model to obtain semantic embedding vectors corresponding to routes, stations, and segments. The semantic embedding vectors corresponding to routes, stations, and segments are fused to obtain semantic features. The structured features of bus routes are concatenated with the intervention strategy to obtain a joint feature representation; the joint feature representation and semantic features are fused to obtain a first embedding vector; the first embedding vector is input into a trained deep causal network to obtain a carbon efficiency estimate when the intervention strategy is implemented. The structural and semantic features of the bus route are fused to obtain a second embedding vector; the second embedding vector is then input into a trained deep causal network to obtain an estimate of carbon efficiency without intervention. The evaluation of intervention measures is achieved by quantifying the rate of improvement in carbon efficiency estimates when intervention strategies are implemented versus when they are not implemented. The process of training a deep causal network includes: The structured features of bus routes are combined with intervention strategies to obtain a joint feature representation; The encoded joint feature representation and the encoded semantic features are fused to obtain the effect embedding representation; the effect embedding representation is then input into a deep causal network to obtain the carbon efficiency estimate. A joint loss function is set, which is a weighted sum of the inverse bias weighted mean square error loss and the adversarial regularization loss; wherein, the inverse bias weighted mean square error loss is calculated based on the inverse bias weight, the number of bus routes, the carbon efficiency estimate, and the actual carbon efficiency value; the inverse bias weight is set through a multi-branch deep bias network; Deep causal networks and multi-branch deep propensity networks are trained collaboratively using a joint loss function. The process of embedding the effect representation into a deep causal network to obtain an estimate of carbon efficiency includes: The effect embedding representation is input into the backbone of the deep causal network to obtain the deep embedding vector; The deep embedding vectors are output through the individual effect estimation branch, the average treatment effect branch, and the global trend branch, respectively, to obtain the individual effect estimate, the average treatment effect estimate, and the global trend effect estimate; The carbon efficiency estimate is obtained by weighted summation of the individual effect estimate, the average treatment effect estimate, and the global trend effect estimate.

2. The method for evaluating public transport carbon efficiency through large-scale model semantic-driven intervention according to claim 1, characterized in that, The process of obtaining semantic features includes: Input the line-level prompts, station-level prompts, and segment-level prompts into the fine-tuned large language model to obtain the hidden states of all tokens in the last hidden layer of the large language model corresponding to the line-level, station-level, and segment-level prompts. Based on the hidden states of all tokens in the last hidden layer of the large language model corresponding to the line level, station level, and road segment level, and the total number of tokens output by the last hidden layer of the large language model corresponding to the line level, station level, and road segment level, calculate the semantic embedding vectors corresponding to the line, station, and road segment. The semantic embedding vectors corresponding to stations are averaged, and the semantic embedding vectors corresponding to road segments are averaged. The semantic features are obtained by fusing the semantic embedding vectors corresponding to the routes, the average semantic embedding vectors corresponding to the stations, and the average semantic embedding vectors corresponding to the road segments.

3. The method for evaluating public transport carbon efficiency through large-scale model semantic-driven intervention according to claim 1, characterized in that, The process of obtaining the inverse tendency weights includes: The dimensionality-enhanced structural features of bus routes and the dimensionality-reduced semantic features are fused to obtain a multimodal embedding vector; The multimodal embedding vector is input into a multi-branch deep bias network to obtain the probability distribution of implementing intervention measures; the reciprocal of the probability distribution is used as the inverse bias weight.

4. The large-model semantic-driven public transport carbon efficiency intervention assessment method according to claim 1, characterized in that, The adversarial regularization loss uses KL divergence to determine the difference between the distribution of carbon efficiency estimates and the distribution of true carbon efficiency.

5. The large-model semantic-driven public transport carbon efficiency intervention assessment method according to claim 1, characterized in that, The intervention strategies include: setting up dedicated bus lanes, prioritizing bus traffic lights, optimizing vehicle dispatching schemes, and optimizing departure intervals.

6. An electronic device, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores one or more computer programs that can be executed by the at least one processor, the one or more of the computer programs being executed by the at least one processor to enable the at least one processor to perform the large model semantically driven public transport carbon efficiency intervention assessment method as described in any one of claims 1-5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the large-model semantic-driven public transport carbon efficiency intervention assessment method as described in any one of claims 1-5.

8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the large-model semantic-driven public transport carbon efficiency intervention assessment method as described in any one of claims 1-5.

Images