A highway digital asset dynamic valuation method and system based on a causal spatiotemporal model

Abstract

The application relates to the technical field of dynamic asset evaluation, in particular to a highway digital asset dynamic evaluation method and system based on a causal spatiotemporal model.The method comprises the following steps: performing causal spatiotemporal topology reconstruction and feature evolution based on structure discovery according to acquired data; performing cross-domain manifold dimension reduction from modal physical perception to asset value entropy based on the evolution result, including specific business demand-oriented large language model scene semantic embedding, mutual information-based digital asset value entropy quantification network construction, value entropy-guided cross-domain manifold dimension reduction and economic representation extraction; and the application realizes the leap of digital asset management from static storage pricing to dynamic intelligent scheduling and continuous value-added, greatly enhancing the long-term commercial realization efficiency of data assets.

Description

Technical Field

[0001] This invention relates to the field of dynamic asset valuation technology, and in particular to a dynamic valuation method and system for highway digital assets based on a causal spatiotemporal model. Background Technology

[0002] Against the backdrop of the comprehensive deepening of vehicle-road-cloud integration and autonomous driving technologies, highways, as a typical high-frequency data generation scenario, urgently need to transform their accumulated massive radar point clouds and high-definition monitoring records from dormant data into high-utility digital assets. Currently, using large models to process traffic data has become a standard practice in the digital transformation of road networks. However, data asset evaluation for multi-agent collaborative computing is not simply a record of physical states, but a highly nonlinear four-dimensional coupled computational process of "physical perception - causal evolution - utility quantification - agent game scheduling".

[0003] Existing traffic big data models or data scheduling systems face technical bottlenecks when dealing with asset packages that possess high-order spatiotemporal dynamics and multi-agent collaborative attributes. These bottlenecks include causal topological discontinuities and decoupling of physical characteristics from system utility evaluation, resulting in rigid evaluation mechanisms for data resources that are detached from the actual utility of computational tasks.

[0004] On the one hand, traditional spatiotemporal graph convolutional networks or general multimodal large models, when processing road network data, essentially rely on surface feature extraction from two-dimensional topological space or statistically based temporal attention mechanisms. This modeling approach, which emphasizes correlation over causality, ignores the true causal driving mechanism between physical entities on highways and traffic events. For example, when an anomaly occurs on a certain road segment, traditional models will assign equal weight to all surrounding sensor data for extraction and transmission, but cannot accurately locate the true source data of the event through the spatiotemporal causal chain. This lack of micro-causal tracing capability leads to the system's inability to accurately evaluate the network contribution of key data nodes. The evaluation results exhibit the limitation of massive redundant data crowding out computing power, while the evaluation weight of core high-quality data is diluted.

[0005] On the other hand, real high-quality data flow is often accompanied by complex multi-agent task requirements such as federated training of autonomous driving models and collaborative risk avoidance between vehicles, roadside, and the cloud. Most existing system platforms only focus on superficial basic indicators such as data clarity, storage capacity, or API call frequency, severely neglecting the information value entropy generated by specific data nodes in eliminating decision-making uncertainty in downstream agents. The system cannot achieve cross-domain semantic alignment between low-level physical perception features and the task utility attributes of high-level agents, resulting in a severe disconnect between road network physical features and agent collaborative utility in the evaluation. This makes it difficult to dynamically invert the true contribution and high-priority scheduling of core digital assets to the entire multi-agent system under conditions of limited edge computing resources or sudden extreme events. Summary of the Invention

[0006] To address the technical challenges of existing technologies in multi-agent collaborative computing scenarios, such as distorted data causal attribution, disconnect between underlying physical perception and high-level task utility, and rigid static resource evaluation mechanisms, this invention provides a dynamic valuation method and system for highway digital assets based on a causal spatiotemporal model. This method achieves accurate weighted attribution of underlying data by constructing a causal topology discovery network, breaks down the modal barriers between physical characteristics and computational utility using value entropy manifold mapping, and introduces a federated large-scale model double-blind game and stochastic differential equation scheduling mechanism. This enables high-precision, strong causal, and full-cycle adaptive dynamic valuation and collaborative scheduling of highway digital assets.

[0007] Firstly, the present invention provides a dynamic valuation method for highway digital assets based on a causal spatiotemporal model, which adopts the following technical solution: A dynamic valuation method for highway digital assets based on a causal spatiotemporal model includes: Acquire multi-source heterogeneous sensor data; Based on the acquired data, perform causal spatiotemporal topology reconstruction and feature evolution based on structure discovery; Based on evolution results, cross-domain manifold dimensionality reduction from modal physical perception to asset value entropy includes semantic embedding of large language model scenarios for specific business needs, construction of digital asset value entropy quantification network based on mutual information, and cross-domain manifold dimensionality reduction and economic representation extraction guided by value entropy. Based on the representation results, a supply and demand double-blind dynamic pricing game based on a federated large model agent is conducted, including the instantiation of federated agent intelligent agents and the initialization of utility benchmarks, a multi-agent game adversarial network under the condition of supply and demand double-blind, Nash equilibrium convergence and generation of asset dynamic valuation truth value. Digital asset scheduling and value-added evolution under stochastic differential equation constraints based on the true value of dynamic asset valuation; Output the true value of the asset valuation after the value-added process.

[0008] Secondly, a dynamic valuation system for highway digital assets based on a causal spatiotemporal model includes: The data acquisition module is configured to acquire multi-source heterogeneous sensor data; The feature evolution module is configured to perform causal spatiotemporal topology reconstruction and feature evolution based on structure discovery, according to the acquired data. The dimensionality reduction module is configured to perform cross-domain manifold dimensionality reduction from modal physical perception to asset value entropy based on evolution results. This includes semantic embedding of large language model scenarios for specific business needs, construction of digital asset value entropy quantification network based on mutual information, and cross-domain manifold dimensionality reduction and economic representation extraction guided by value entropy. The game module is configured to perform supply and demand double-blind dynamic pricing game based on the representation results, including the instantiation of the federated agent and the initialization of the utility benchmark, the multi-agent game adversarial network under the condition of supply and demand double-blind, the convergence of Nash equilibrium and the generation of the true value of dynamic asset valuation. The value-added module is configured to perform digital asset scheduling and value-added evolution under the constraints of stochastic differential equations based on the true value of asset dynamic valuation. The output module is configured to output the true value of the asset valuation after the increase.

[0009] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned dynamic valuation method for highway digital assets based on a causal spatiotemporal model.

[0010] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide a dynamic valuation method for highway digital assets based on a causal spatiotemporal model.

[0011] In summary, the present invention has the following beneficial technical effects: First, this invention alleviates the problem of rampant spurious correlations caused by the "emphasis on spatial distance and neglect of causal generation" in traditional road network feature extraction, thereby reducing the causal attribution error rate of core data assets. Existing technologies, when processing highway data, often rely on physical distance to construct graph networks, easily leading to forced coupling of opposing lane data with no causal relationship, resulting in distorted attribution. This invention introduces a differentiable causal graph tensor discovery mechanism through a "structure discovery-based causal spatiotemporal topology reconstruction and feature evolution module," forcibly removing spurious correlations from massive heterogeneous data. Experimental data shows that, when facing a complex extreme condition dataset (Complex-HighwayAsset) with numerous sensor outages and bidirectional lane interference, the causal topology discovery error rate of this invention is only 0.11, lower than the 0.31 of the spatially powerful ST-GCN method. This means that the system can extremely accurately remove data noise, restore the true causal propagation chain of traffic events, and provide solid physical evidence for digital asset ownership confirmation.

[0012] Secondly, this invention addresses the modal semantic disconnect between pure physical situational awareness and the commercial digital economy, significantly improving the cross-domain matching accuracy of data assets under specific business needs. Addressing the inherent modal gap between physical behavior and economic value in highway networks, this invention utilizes a "cross-domain manifold dimensionality reduction module for multimodal physical perception to asset value entropy," leveraging scene semantics and mutual information divergence from a large language model to dynamically transform physical data into asset value entropy. Experimental results show that, on the Complex-HighwayAsset dataset, even with samples where the video is blurry but accident liability can be determined—a situation of extreme conflict between appearance and value—the accuracy of this invention's physical-economic cross-domain mapping remains as high as 88.6%, more than half higher than the Late-Fusion method (58.5%), which only performs simple feature concatenation. This design enables the system to keenly capture the ability of cold data to eliminate business uncertainties in extreme scenarios, accurately extracting high-premium economic representations.

[0013] Furthermore, this invention addresses the challenges of rigid subjective pricing and conflicting supply and demand information sources in traditional digital asset transactions, achieving a dynamic valuation consensus that approaches absolute fairness. This invention abandons the simple static allocation of Shapley values ​​or unilateral auction mechanisms in traditional evaluations, introducing a "double-blind dynamic pricing game module based on a federated large model agent," utilizing Nash equilibrium theory to simulate the dynamic negotiation and compromise process between buyers, sellers, and regulators. This mechanism can automatically correct the seller's data premium and the buyer's malicious price suppression, reducing the dynamic valuation equilibrium convergence error to 0.15 on the Complex-HighwayAsset dataset, which is superior to the ST-GCN+MARL model (0.59) lacking a double-blind game mechanism. This demonstrates that this invention possesses superior intelligent arbitration capabilities in handling complex multi-party transaction conflicts and generating objective, fair asset prices that conform to market supply and demand laws.

[0014] Finally, this invention achieves a leap in digital asset management from static storage pricing to dynamic intelligent scheduling and continuous value-added, greatly enhancing the long-term commercial monetization efficiency of data assets. Relying on the "Digital Asset Scheduling and Value-Added Evolution Module under Stochastic Differential Equations Constraints," this invention can not only accurately value current assets but also deduce the optimal data packaging and distribution path to resist market white noise interference. Experimental data verifies that the scheduling strategy generated by this invention based on SDE achieves a long-term return of 87.5% on the Complex-HighwayAsset dataset, nearly doubling the return compared to the conventional multi-agent reinforcement learning ST-GCN model (45.3%). The system can provide the most forward-looking suggestions for computing power allocation and circulation based on the long-term historical growth records and current demand fluctuations of digital assets, realizing a closed-loop commercial model and continuous value-added of data assets throughout the entire lifecycle of the highway network. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of a dynamic valuation method for highway digital assets based on a causal spatiotemporal model according to Embodiment 1 of the present invention; Figure 2 This is a comparison chart of the errors of each model in Embodiment 1 of the present invention under complex and extreme working conditions; Figure 3 This is a graph showing the changes in the cross-domain value mapping accuracy of each model under different scenarios in Embodiment 1 of the present invention; Figure 4 This is a comparison chart of the long-term returns of asset scheduling in various models under complex scenarios in Embodiment 1 of the present invention; Figure 5 This is a heatmap of the comprehensive capabilities of each model in a complex scenario according to Embodiment 1 of the present invention. Detailed Implementation

[0016] The present invention will be further described in detail below with reference to the accompanying drawings.

[0017] Example 1 Reference Figure 1 This embodiment of a dynamic valuation method for highway digital assets based on a causal spatiotemporal model includes: The specific plan is as follows: (1) Causal spatiotemporal topology reconstruction and feature evolution module based on structure discovery This module, serving as the core sensing hub and data cleaning foundation of the entire digital asset valuation system, primarily handles the massive, high-frequency, and noisy multi-source heterogeneous sensor data generated every moment in the highway network. Previous traffic feature extraction techniques typically constructed spatial network topology based on absolute physical distances between devices or simple historical statistical correlations. This easily leads to the forced coupling of data from vehicles on the same physical coordinates but with no actual business causal relationship, such as two-way isolated lanes. This module completely abandons traditional spatial distance maps and designs a bottom-level causal structure discovery mechanism. This mechanism automatically mines and reconstructs the true causal generation chains between heterogeneous data streams using deep learning algorithms, accurately stripping away false correlations and tracing back to the core "causal nodes" driving the road network evolution. Subsequently, the system utilizes this high-purity causal topology, combined with continuous-time neural differential equations, to perform a panoramic extrapolation of the state evolution trajectory of digital assets along the time axis, fundamentally ensuring the high accuracy of data asset ownership confirmation and the rigor of traceability.

[0018] 1) Causal structure graph discovery and adjacency tensor construction of highway network asset nodes In real-world highway network operations, traffic events such as abnormal deceleration leading to congestion and subsequent spread exhibit strong directional and causal chain characteristics. Accurately capturing and quantifying this causal network is a prerequisite for extracting high-value digital assets. The system first connects all sensing device nodes within the road network in parallel as digital asset sources. For any two asset nodes, the system compares their temporal fluctuation characteristics through a pre-trained network. Utilizing a continuously differentiable operator with a temperature annealing mechanism, it transforms the implicit dependencies between high-dimensional features into an explicit directed probability graph. The causal discovery formula is as follows: , in, Representing time From asset nodes Pointing to node The directed causal influence strength probability, and the calculation results of all node pairs together constitute the global causal adjacency tensor required by the system. ; and These represent the underlying physical state feature vectors collected by the two nodes at the current moment; (symbols) This indicates that a feature concatenation operation is performed on these two vectors; The internally learnable parameters are Multilayer perceptual inference network; It is a differentiable discretization operator used to solve the problem of gradient non-differentiability in graph structures during discrete sampling; Annealing temperature hyperparameters used to control the sparsity of the probability distribution; The directed topological physical isolation mask, constructed based on the high-precision map, directly forces invalid paths that are physically blocked to be set to zero.

[0019] The global causal adjacency tensor calculated in this step The underlying features of the nodes will serve as precise topological guidelines, directly passed to the next stage to guide the deep aggregation of features.

[0020] 2) Causal topology-driven spatiotemporal feature aggregation network After obtaining the global causal adjacency tensor output in step 1), Next, the system needs to use this structure to aggregate the originally scattered asset features in a high-order space. To prevent the features of the initiating causal nodes of core events from being overwhelmed by the massive number of normal node features around them, this step uses a causal out-degree matrix to strictly constrain the flow of information transmission, enabling result nodes to adaptively and directionally absorb key early warning information from causal nodes, thereby ensuring the clarity of the asset traceability. The aggregation update formula is designed as follows: , in, The causal high-order spatial feature matrix output in this step is deeply integrated with the causal propagation context of the entire road network. The global causal adjacency tensor generated in 1); It is a global input tensor containing the initial physical characteristics of all nodes at the current moment; It is the inverse of the out-degree diagonal matrix of the causal graph, used to normalize the directed information flow to prevent numerical explosion; This is a learnable linear projection matrix responsible for the transformation of the feature space dimensions; The Mish self-regularized nonlinear activation function was selected to maintain deep negative gradient flow; For the bypass residual term, where It is an adaptive residual coefficient matrix, used to ensure that the original physical properties of the underlying sensing devices are not excessively tampered with.

[0021] The causal high-order spatial feature matrix output in this step This will be seamlessly distributed as the initial physical and causal state benchmark for the next step of time evolution.

[0022] 3) Asset state manifold derivation under continuous-time neural differential equations Because the freshness of real highway digital assets continuously decays over time, and roadside sensors often suffer from inconsistent sampling frequencies or data loss, traditional discrete-time series models cannot handle these irregular and fragmented time segments. Therefore, the system receives causal higher-order features transmitted from 2). As the starting point, the technique of divine constant differential equations is introduced to completely abstract the evolution of asset characteristics over time into a continuous mathematical integral process, thereby smoothly filling data gaps and accurately calculating asset depreciation. Its continuous derivation formula is as follows: , in, This represents the output from the sensing base module to the higher levels of the system, extrapolating to the future target time. The asset state manifold tensor; This is the starting point of the initial high-order feature passed in from the previous step; and These correspond to the current actual observation time and the expected future target time, respectively. For continuous-time variables within the integration domain; Representative at The instantaneous characteristic state at any given moment; It is a nonlinear differential function, given by the control parameters. A deep neural network is used for black-box approximation fitting to compute features in extremely small time slices. The nonlinear decay gradient rate within.

[0023] The final asset state manifold tensor obtained by this module It accurately depicts the spatiotemporal physical trajectory of the continuous evolution of digital assets. This tensor will be used as core physical evidence and directly transmitted to the system module (2) for carrying out the asset value entropy manifold dimensionality reduction and mapping calculation in the digital economy domain.

[0024] (2) Cross-domain manifold dimensionality reduction module for multimodal physical perception to asset value entropy This module serves as a core bridge connecting the underlying physical road network and the high-level digital economy market, aiming to solve the problem of the silo problem of serious modal barriers between "physical status" and "commercial value" in traditional digital asset evaluation. In real traffic data transactions and circulation, pure physical data such as the average speed of a certain road segment or radar cross-section do not have a fixed market price. The real value of the data is highly dependent on the specific business demand scenario. For example, the same traffic flow monitoring video may only have very low basic storage value during the daily off-peak period, but in the event of a major chain-reaction rear-end collision, the video has extremely high "source traceability value" for insurance claims because it can accurately determine liability. If the system only focuses on extracting physical features, it will completely lose the ability to capture the scarce premium of data under extreme conditions. To this end, this module introduces the interdisciplinary concept of "asset value entropy". By quantifying the physical manifold tensor output by the module (1) to eliminate the uncertainty of the system in specific business scenarios, it dynamically transforms the cold spatiotemporal data stream into an economic utility manifold that can directly participate in market pricing. This process bridges the semantic gap between the physical and economic domains, providing a legally and commercially sound economic foundation for subsequent multi-party dynamic valuation games.

[0025] 1) Semantic embedding of large language models for specific business needs Valuation of any digital asset must be grounded in reality and considered within the specific buyer's demand scenario. The core task of this step is to transform unstructured business requirements, such as "autonomous driving companies need radar data for emergency braking of vehicles in rainy or snowy weather" or "traffic management departments need cross-provincial traffic flow prediction results during holidays," into continuous mathematical representations that the system can perform high-dimensional tensor operations on. To accurately capture the hidden commercial intent and asset selection biases within the demand text, the system uses a pre-trained large-scale language model as a semantic encoder to perform deep spatial embedding of the demand input. The scenario semantic embedding formula is as follows: , in, This represents the high-dimensional semantic representation matrix of the business scenario calculated in this step, which highly condenses the business intent and application context of this valuation task; For receiving raw natural language request text sequences from business systems or data buyers; It is a specially designed fine-tuning vector of prompt words in the field of digital economy of transportation. Its function is to force the large model to focus on the core valuation elements such as the timeliness, completeness and scarcity of data when encoding, rather than ordinary text summary. The autoregressive encoding operator, representing the underlying large language model, is responsible for mapping discrete text into high-dimensional latent vectors. As a dimension reduction projection matrix, it is specifically responsible for compressing the ultra-high-dimensional semantic vectors output by large models into a dimensional space that matches the underlying physical features; The bias term vector is used to prevent feature space shift. is the Gaussian error linear unit activation function, used to introduce nonlinear expressions and maintain a smooth transition in the semantic space.

[0026] The scene semantic representation matrix output in this step This will serve as a core benchmark representing the demand side, directly inputting it into the next step to engage in value collision with physical characteristics.

[0027] 2) Digital Asset Value Entropy Quantification Network Based on Mutual Information After acquiring the physical evolution characteristics representing the supply side and the semantic representation of the demand side, the most challenging technical aspect of the system is how to accurately quantify the degree of fit between these two using mathematical methods. This step introduces the mutual information and entropy reduction theories from information theory to construct a value entropy quantification network. Its business logic is that the value of a digital asset in a given business scenario depends on the extent to which the asset can eliminate the uncertainty of that business scenario (i.e., reduce information entropy). The system derives a dynamic value assessment coefficient by calculating the cross-mutual information between the physical and semantic tensors in the latent space and applying an information divergence penalty. The value entropy quantification formula is as follows: , in, This represents the asset value entropy matrix calculated in this step. It is a dynamic and continuous scalar set that accurately reflects the economic utility intensity of the asset at the target time. That is, the asset state manifold tensor (representing physical reality) continuously derived from module (1); This is the semantic representation matrix of the business scenario extracted in module 1) (representing economic demand); and These are cross-domain alignment mapping matrices for physical and semantic features, respectively, used to eliminate underlying distributional differences in heterogeneous data; superscript This represents the transpose of a matrix to satisfy the rules of dot product operation; This is a scaling factor used to prevent gradient saturation caused by excessively large high-dimensional dot product results; Mutual information gain amplification factor; KL divergence is used for extremely rigorous calculations of the probability distribution of physical features. Semantic demand probability distribution Information entropy loss between them; As a divergence penalty hyperparameter, when the physical data and business requirements are mismatched, a large KL divergence will deduct a lot of points from this item; To smoothly approximate the nonlinear activation function of ReLU, we ensure that the output value entropy is always a non-negative value with real economic significance.

[0028] The asset value entropy matrix output in this step This will serve as the core dynamic gating switch, guiding the next step of dimensional reduction operations.

[0029] 3) Value entropy-guided cross-domain manifold dimensionality reduction and economic representation extraction In real valuation games, buyers, sellers, and regulatory platforms cannot directly process the massive raw tensors containing a huge amount of redundant physical information. These tensors must be refined into highly condensed feature vectors with clear economic orientations to significantly reduce the computational overhead of subsequent game theory algorithms. This step uses the asset value entropy generated in the previous step as a dynamic gating mechanism to perform weighted filtering and nonlinear dimensionality reduction on the raw physical tensors. Its essence is to remove redundant information from the physical tensors that has no economic value and is ineffective in the current business scenario, such as the impact of sunny weather on rain and snow prediction, retaining only the core causal anchors with high value entropy, and ultimately merging them into a unified economic representation. The manifold dimensionality reduction formula is as follows: , in, The digital asset economic representation vector representing the output of this stage is a high-order economic manifold that is completely free from the constraints of pure physical attributes and can be directly used for game pricing in the digital asset trading market. It is still the asset state manifold tensor passed from module (1); 2) The asset value entropy matrix calculated; symbol This represents the Hadamard product (element-by-element multiplication). It is a dimension reduction and compression matrix for economic manifolds, responsible for compressing the weighted, massive feature map into a compact vector space; This is a layer normalization operation designed to eliminate fluctuations in absolute economic values ​​across different batches of data, ensuring a stable valuation baseline; the operator... This represents the fusion and addition of vector dimensions, combining the compressed high-value physical features with the original business scenario semantics in 1). Perform deep binding.

[0030] The economic representation vector generated through this module is a refined process. This will serve as a highly valuable valuation tactic, seamlessly transmitted to the system module (3), and officially initiate the dynamic game pricing process among the three parties of "vehicle-road-cloud".

[0031] (3) Supply and demand double-blind dynamic pricing game module based on federated big model agent This module serves as the central hub for commercial value arbitration and transaction decisions in the entire system, aiming to address the core pain points of rigid pricing and supply-demand mismatch in the traditional data trading market. In conventional digital asset transfers, data providers, such as highway operators, tend to offer static high premiums, while data consumers, such as autonomous driving companies and insurance companies, want to reduce procurement costs. Pure machine algorithms lack an understanding of market supply and demand dynamics. To break this valuation deadlock, this module abandons the traditional fixed formula pricing method and designs a multi-agent agent double-blind game mechanism based on a federated learning architecture. This mechanism receives the high-order economic representation output by module (2) as the bargaining chip, and instantiates three large-scale intelligent agents representing the seller, buyer, and regulatory platform. Under the "double-blind" condition of not seeing each other's bargaining chips, the three intelligent agents engage in multiple rounds of tensor-level bidding and compromise at the mathematical level around the scarcity and utility of the asset, ultimately converging within the framework of Nash equilibrium to arrive at a dynamic true value of the asset that takes into account both market supply and demand and regulatory compliance. This mechanism not only gives cold data the warmth of the commercial market, but also achieves a deep assimilation of the subjective and objective aspects of asset pricing.

[0032] 1) Instantiation of the Federal Agent and Initialization of Utility Benchmark Before the dynamic game officially begins, the system must establish the initial bargaining power and utility benchmarks for the three participating intelligent agents (seller's agent, buyer's agent, and platform regulatory agent). The more closely the underlying data they possess aligns with the current business scenario, the higher their bargaining power and confidence in not compromising. The system initializes the three independent federated intelligent agents in parallel and calculates their initial utility functions based on the economic representations passed from the previous module. Their utility initialization formulas are as follows: , in, Representing intelligent agents ( , representing the utility confidence scalar of the seller, buyer and regulator respectively in round 0, which is the initial round. The closer this value is to 1, the stronger their negotiating power and the less willing they are to compromise with other parties in subsequent games. The digital asset economic representation vector obtained from module (2) constitutes the common objective physical basis for the tripartite negotiations; For this specific intelligent agent Built-in business objective function vectors (e.g., buyer's built-in price reduction objective, seller's built-in premium objective), superscript For the transpose operation, by using... The dot product operation is used to measure the alignment between the current data and its own business interests; the denominator part The L2 norm of the solution vector is used to normalize the dot product result using standard cosine similarity. The system pre-sets scaling factors for the aggressiveness of each participant to simulate the risk preferences of entities under different market conditions; As a federal compliance penalty, this item significantly reduces the utility confidence level when a party's business intentions seriously violate the principle of fair market competition. It is a non-linear activation function responsible for smoothly compressing the calculation results into a probability range of 0 to 1.

[0033] The initial utility confidence level generated in this step The agent's initial bid tensor will serve as the core chip for initiating multi-round game competition, and will be directly pushed into the next step of the double-blind game engine.

[0034] 2) Multi-agent game adversarial network under supply and demand double-blind conditions Once all three agents are given initial handicap and utility confidence levels, the game theory system officially enters the substantive adversarial iteration phase. Under the premise of federated double-blind protection of all parties' commercial privacy, agents cannot directly read the other party's target bottom line; they must deduce market consensus by analyzing the bidding differences revealed in each round. The system constructs a global game adversarial loss function, forcing the three parties to continuously adjust their bidding tensors in multiple iterations to minimize this global divergence. The formula for its global game adversarial loss is as follows: , in, For a scalar value calculated in real time, the value at the first moment is precisely quantized. The overall sense of division and disagreement among the three parties at the negotiating table regarding the pricing of digital assets; conditions for seeking peace. This ensures that the system cross-calculates the conflicts between the seller, buyer, and regulator; and Representing intelligent agents and The utility confidence scalar in the current round; fractional term It is an asymmetric weighting mechanism based on utility confidence, with constants. The purpose of this is to absolutely prevent system crashes caused by a denominator of zero; and Representing intelligent agents and The asset pricing high-dimensional tensor advocated in the current round; The square of the Frobenius norm, representing the difference between these two tensors, is the valuation gap as subjectively perceived by each party. These are specially designed regulatory-anchored penalties, among which... Specifically refers to the tensor of the specific pricing proposition given by the seller's agent in the current round. This specifically refers to the reference floor pricing tensor, which possesses absolute fairness and is provided by the regulatory agent in the current round, using the L1 norm. With penalty coefficient This is to prevent potential collusion between buyers and sellers or extreme price gouging.

[0035] The global adversarial loss calculated in this step This clearly indicates the optimal compromise direction in the high-dimensional valuation space, and its partial derivative gradient will be directly used for the Nash equilibrium convergence update in the next step.

[0036] 3) Nash equilibrium convergence and the generation of true value for dynamic asset valuation After calculating the global game disagreement loss, the three agents must modify their offers along the gradient descent direction. This process highly simulates the dynamic game behavior in real business negotiations, where agents gradually probe and make concessions based on the other party's bottom line. Through hundreds of rounds of tensor exchanges, when the rate of change of the global disagreement loss approaches a small system threshold, the engine automatically determines that the negotiation has reached a Nash equilibrium. At this point, the system extracts the final compromise offers from the three parties, uses a soft maximization mechanism for confidence weighting, and generates an indisputable and unique true value for the asset valuation. The offer updates are as follows: , in, and Representing intelligent agents The pricing tensor of the next round versus the current round; The learning rate hyperparameter for system game theory controls the step size of a single concession; partial derivative terms It is precisely the mathematical gradient direction found in the previous step that can maximally resolve the divergence; symbol For the Hadamard product, the part in parentheses For dynamic compromise control gate, The threshold value is a pre-defined rigid matrix of agent personality. When the initial confidence of one party is extremely high, the threshold value approaches zero, instantly blocking back propagation and thus refusing to lower the price.

[0037] Its final valuation formula is as follows: , in, This is the central consensus valuation truth matrix that this module outputs, representing the absolute transaction value of the digital asset in this business scenario. This is the final compromise card for all parties after the Nash equilibrium converges; and The confidence level retained at the convergence time, combined with the natural index. The normalized weighting ensures that bids from parties with strong evidence receive a larger weighting. This is the game theory temperature coefficient, used to differentiate confidence levels.

[0038] The final output of this mechanism is the valuation truth matrix. This will be directly transmitted to the system module (4) as a core economic indicator to guide the commercial scheduling and future value-added evolution of assets.

[0039] (4) Digital asset scheduling and value-added evolution module under stochastic differential equation constraints This module serves as the final business focus and macro-control portal of the entire highway digital asset dynamic valuation and management system. It aims to completely break the dead end of static business in traditional traffic data management, which emphasizes evaluation over application and focuses on the present over the future. After obtaining the extremely accurate and commercially consensus-based dynamic valuation of digital assets output by module (3), the core task of the system shifts from simple pricing to substantive asset operation, distribution, and value-added services.

[0040] In the real-world digital economy ecosystem of highways, the lifecycle value of data assets depends not only on the current one-time ownership transaction but also on complex external environmental factors such as the random fluctuations in future market demand and the scheduling and allocation strategies of edge computing resources. Therefore, this module proactively abandons linear asset appreciation models and discrete Markov chains, introducing stochastic differential equations from advanced mathematics into the asset closed-loop management system. By simulating the continuous Brownian motion trajectory of asset value in a real trading market filled with white noise, the system can deduce the optimal asset packaging and circulation actions. This process fundamentally realizes the stochastic dynamic scheduling and value-added feedback of highway digital assets throughout their entire lifecycle, forming a perfect spatiotemporal valuation and data operation closed loop.

[0041] 1) Market volatility modeling and asset allocation strategy optimization based on stochastic differential equations Faced with rapidly changing business demands and the potential for unforeseen highway anomalies, any static resource allocation and data flow strategy is highly susceptible to leading to the idle waste of high-value assets or the strain on computing power for low-value data. This step models the expected future value increment of the target digital asset as a continuous evolutionary process involving deterministic growth and random fluctuations. The system uses a deep neural network to fit the drift and diffusion terms, searching among a vast array of legitimate scheduling strategies, such as whether to anonymize and upload data to the blockchain, whether to distribute across domains, and whether to reduce frequency for cold backup, to find the optimal scheduling action vector that maximizes the global long-term economic return. Its stochastic differential equation model is as follows: , in, Represents a tiny time period The differential form of the expected potential increase in the commercial value of the digital asset; It is a deterministic drift term function parameterized by a deep neural network, which receives the estimated truth matrix output by the previous module. and the candidate scheduling actions currently being attempted by the system. As a joint input, it is used to assess the expected linear and deterministic value growth that this distribution action can bring; The network is fitted to the diffusion term, and it receives the high-order economic representation vector extracted by module (2). This is used to calculate the market risk and premium volatility faced by the asset in the current specific business scenario; It is a standard Wiener process (i.e., the differential of Brownian motion) used to introduce unpredictable random white noise such as sudden events in the highway network or sudden changes in market demand in the continuous time domain.

[0042] The optimization formula for the strategy is as follows; , in, This is the feature vector of the globally optimal scheduling action derived in this step; Operators are responsible for the system's vast library of legal actions. We then perform rigorous traversal and gradient optimization. The expected value is calculated to comprehensively assess the average anticipated performance of the asset after experiencing various random fluctuation paths; the integral term represents the cumulative increase in the asset's value over an infinite time span, where... It introduces the time discount rate. The continuous discount factor ensures that the valuation system has time-based financial attributes, emphasizing the certainty of current returns rather than the divergent returns of the infinite future. It is a continuous-time variable.

[0043] The optimal scheduling action vector obtained by solving This will serve as the actual data flow control command issued to the highway network edge computing gateway or cloud control platform, and will simultaneously flow into the next step for closed-loop updating of asset files.

[0044] 2) Closed-loop updating of historical archive characteristics of asset appreciation evolution and long-term utility After completing the most forward-looking action strategy optimization and scheduling, the system must record all the high-dimensional tensor results of this valuation game, value discovery, and circulation scheduling as extremely valuable historical digital credit into the underlying asset database. An originally mediocre surveillance video or radar asset, after being successfully traced, valued at a high price, and precisely scheduled to participate in the commercial empowerment of a major traffic event by the system, should have its credit score and expected premium space permanently enhanced.

[0045] This step utilizes the momentum smoothing feature update equation to deeply and physically fuse the valuation results of this federated game with the optimal scheduling strategy, thereby updating the long-term historical record of this asset in the system's underlying data ledger and ultimately completing the spatiotemporal causal closed loop of the entire large model evaluation system. The feature record update formula is as follows: , in, and These represent the high-dimensional feature matrices of the long-term utility history of the digital assets of this specific highway segment in the system's macro distributed ledger before and after the execution of this status value-added update; The system's manually preset translational slip coefficient parameters; symbol The feature channel splicing operator is responsible for converting the final estimated truth matrix generated by the Nash equilibrium of module (3). The optimal scheduling action feature vector obtained in the previous step Forced assembly in both physical and economic dimensions; To maintain the consistency of the latent space dimension of the archive features, the alignment projection update matrix is ​​used to ensure that newly incorporated information can be perfectly compatible with the old ledger. As a hyperbolic tangent nonlinear activation function, it plays a crucial role in preventing numerical overflow and normalizing, and can strictly and smoothly limit the values ​​of the incremental features after fusion and updating within the standardized range.

[0046] This dynamic update mechanism completes the mathematical closed loop of data assets from causal tracing to commercial monetization and then to feedback and accumulation. This enables the system to examine the value fluctuations of the same road segment or similar digital assets when it starts a new round of game valuation in the future, with high-level historical transaction memory. It fundamentally establishes an adaptive and self-growing intelligent valuation system that accompanies the entire life cycle of the highway network.

[0047] Experimental verification To systematically verify the performance advantages of the method of the present invention in performing causal ownership confirmation, dynamic valuation, and value-added scheduling of digital assets under conventional highway operation scenarios and complex extreme traffic conditions, the present invention constructs and uses two highway datasets with different characteristics for comparative experiments: a basic road network perception dataset and a complex extreme condition dataset.

[0048] Dataset description: Base-HighwayAsset (Basic Road Network Sensing Dataset): Collected from standard highway sections under favorable weather conditions, with stable traffic flow and no severe packet loss from sensing devices. This dataset primarily includes tollbooth monitoring data with clear animations, radar point clouds of uniform-speed traffic flow, and corresponding routine traffic flow prediction business requirement texts. It features a clean background and well-defined time series, containing 50,000 basic digital asset samples and corresponding static baseline estimates. It is mainly used to validate the model's basic performance and generalization ability when handling standard traffic data flow tasks.

[0049] Complex-HighwayAsset (Complex Extreme Conditions Dataset): This is a high-difficulty dataset built by this research to address real pain points in the digital economy. It simulates and tests extreme scenarios such as typhoons and heavy rainstorms, multi-vehicle pile-up accidents, and large-scale data loss from some sensors due to power outages at edge gateways. The dataset contains numerous spurious data correlations, such as interference between two-way lanes, and intense supply-demand conflicts, such as high premiums in insurance claims. The dataset contains 80,000 multimodal samples, aiming to verify the core technical advantages of this invention in handling causal structure discovery, physical-economic cross-domain value entropy mapping, and federated large-scale agent game theory.

[0050] Experimental setup: To comprehensively evaluate performance, we selected three representative large-scale models and data pricing methods for comparison with the method of this invention.

[0051] CNN-LSTM + Shapley (Temporal Benchmark Method): This method uses CNN-LSTM to extract temporal physical features and then employs the traditional Shapley value for static data contribution allocation and pricing. It represents a traditional mechanism that ignores the causal topology and dynamic game theory of road network space. Late-Fusion + Auction: This method directly and independently extracts multimodal road network data features, performs simple concatenation at the network end, and uses a traditional one-sided auction mechanism for pricing. It represents a traditional multi-source data trading approach lacking deep cross-domain semantic alignment. ST-GCN + MARL (Spatiotemporal Graph and Multi-Agent Benchmark): This method uses a traditional spatiotemporal graph convolutional network to process road network topology and combines it with standard multi-agent reinforcement learning for resource allocation. It represents a strong competitor in the current transportation field, but it does not remove spurious physical correlations and lacks a value entropy and Nash double-blind game mechanism based on a large model. Finally, there is the Causal-AssetEvoNet method proposed in this invention.

[0052] Evaluation Metrics: The experiment uses five core metrics for evaluation. **Causal Topology Discovery Error Rate (↓):** This measures the error rate of the model in accurately reconstructing the true causal propagation chain of traffic events in complex road networks by removing spurious correlations; a lower value is better. **Physical-Economic Cross-Domain Mapping Accuracy (%) (↑):** This measures the accuracy of the model in matching cold, hard physical perception data with the specific business needs of buyers for value and demand verification; a higher value is better. **Dynamic Valuation Equilibrium Convergence Error (↓):** This measures the deviation between the dynamic asset truth value output by the system after federated game and the actual market supply and demand equilibrium point (absolute truth value); a lower value is better. **Long-Term Asset Scheduling Return Rate (%) (↑):** This measures the commercial return on investment obtained by the system in the long-term real-world transaction market for personalized data scheduling and distribution suggestions generated based on stochastic differential equations; a higher value is better. **Single Game Valuation Time (ms) (↓):** This records the average time from receiving physical data to completing the multi-party game and generating the final scheduling report, reflecting the feasibility of project implementation.

[0053] Table 1. Performance comparison of different methods on the Base-HighwayAsset and Complex-HighwayAsset datasets. The experimental results are shown in Table 1. Figure 2 , Figure 3 , Figure 4 and Figure 5 As shown, through in-depth comparative analysis of the experimental data of each model in Table 1 on the dual dataset, the technical advantages of this invention in handling the complex valuation task of highway digital assets can be clearly revealed. The specific analysis is as follows: First, compared with benchmark models such as CNN-LSTM and ST-GCN, this invention demonstrates advantages in handling road network data cleaning and asset ownership verification. Experimental data shows that even on the ST-GCN model, which has strong spatial processing capabilities, the topology discovery error rate is still as high as 0.31 when faced with a large amount of irrelevant opposing lane data interference in the Complex-HighwayAsset dataset. This is because it relies entirely on spatial distance, leading to a proliferation of false correlations. In contrast, this invention, through a "causal spatiotemporal topology network based on structure discovery," forcibly removes false correlations, compressing the causal topology discovery error rate to an extremely low 0.04 on the Base-HighwayAsset dataset. Even on the extreme Complex-HighwayAsset dataset, with a large number of sensor outages, the error only slightly increases to 0.11, fundamentally ensuring the absolute accuracy of the underlying asset ownership verification evidence.

[0054] Secondly, in the cross-modal value extraction stage, this invention solves the semantic disconnect between physical data and business economic models. While the Late-Fusion model can perform multimodal stitching, its lack of economic guidance causes its physical-economic cross-domain mapping accuracy to plummet to 58.5% when faced with extreme examples where video is blurry but liability can be determined—a situation where appearance and value conflict. This invention, however, introduces an "asset value entropy" mechanism, using the semantics of a large model scenario to perform value-gated dimensionality reduction on the physical tensor. This allows the invention to maintain a cross-domain mapping accuracy as high as 88.6% even in complex, high-noise scenarios, effectively eliminating the modal barriers in the confirmation of digital asset rights.

[0055] Furthermore, this invention solves the challenge of truth value determination under conflicting subjective and objective evaluations, achieving high valuation credibility. This invention abandons traditional static Shapley values ​​or one-sided auction mechanisms, introducing a "two-blind dynamic pricing game based on a federated large-scale model agent." Utilizing Nash equilibrium theory to simulate the dynamic negotiation process between buyers, sellers, and regulators, this mechanism can automatically correct for sellers' data premiums and buyers' malicious price suppression. On the Complex-HighwayAsset dataset, it reduces the dynamic valuation equilibrium convergence error to 0.15, far superior to ST-GCN's 0.59. This demonstrates that this invention possesses superior intelligent arbitration capabilities in handling complex transaction conflicts and generating objective and fair asset prices.

[0056] Finally, this invention achieves a leap from "static pricing" to "dynamic scheduling and value-added operation," enhancing the commercial return on data assets. Based on a strategy optimization mechanism within a stochastic differential equation framework, this invention can not only value current assets but also deduce the optimal data distribution path to resist market white noise interference. Experimental data verifies that the scheduling strategy generated by this invention achieves a long-term return of up to 87.5% on the Complex-HighwayAsset dataset, more than doubling the performance compared to the ST-GCN model using conventional reinforcement learning. The system can provide highly forward-looking asset scheduling suggestions based on the historical growth records of digital assets and current demand fluctuations, realizing a closed-loop commercial value-added model for data assets throughout the entire lifecycle of the highway network.

[0057] Example 2 This embodiment provides a dynamic valuation system for highway digital assets based on a causal spatiotemporal model, including: The data acquisition module is configured to acquire multi-source heterogeneous sensor data; The feature evolution module is configured to perform causal spatiotemporal topology reconstruction and feature evolution based on structure discovery, according to the acquired data. The dimensionality reduction module is configured to perform cross-domain manifold dimensionality reduction from modal physical perception to asset value entropy based on evolution results. This includes semantic embedding of large language model scenarios for specific business needs, construction of digital asset value entropy quantification network based on mutual information, and cross-domain manifold dimensionality reduction and economic representation extraction guided by value entropy. The game module is configured to perform supply and demand double-blind dynamic pricing game based on the representation results, including the instantiation of the federated agent and the initialization of the utility benchmark, the multi-agent game adversarial network under the condition of supply and demand double-blind, the convergence of Nash equilibrium and the generation of the true value of dynamic asset valuation. The value-added module is configured to perform digital asset scheduling and value-added evolution under the constraints of stochastic differential equations based on the true value of asset dynamic valuation. The output module is configured to output the true value of the asset valuation after the increase.

[0058] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device, the aforementioned method for dynamic valuation of highway digital assets based on a causal spatiotemporal model.

[0059] A terminal device includes a processor and a computer-readable storage medium, the processor being configured to implement various instructions; the computer-readable storage medium being configured to store multiple instructions adapted for loading and execution by the processor of the aforementioned dynamic valuation method for highway digital assets based on a causal spatiotemporal model.

[0060] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A dynamic valuation method for highway digital assets based on a causal spatiotemporal model, characterized in that, include: Acquire multi-source heterogeneous sensor data; Based on the acquired data, perform causal spatiotemporal topology reconstruction and feature evolution based on structure discovery; Based on evolution results, cross-domain manifold dimensionality reduction from modal physical perception to asset value entropy includes semantic embedding of large language model scenarios for specific business needs, construction of digital asset value entropy quantification network based on mutual information, and cross-domain manifold dimensionality reduction and economic representation extraction guided by value entropy. Based on the representation results, a supply and demand double-blind dynamic pricing game based on a federated large model agent is conducted, including the instantiation of federated agent intelligent agents and the initialization of utility benchmarks, a multi-agent game adversarial network under the condition of supply and demand double-blind, Nash equilibrium convergence and generation of asset dynamic valuation truth value. Digital asset scheduling and value-added evolution under stochastic differential equation constraints based on the true value of dynamic asset valuation; Output the true value of the asset valuation after the value-added process.

2. The dynamic valuation method for highway digital assets based on a causal spatiotemporal model according to claim 1, characterized in that, The process of performing causal spatiotemporal topology reconstruction and feature evolution based on structure discovery using the acquired data includes: firstly, parallel access to all sensing device nodes within the road network as digital asset sources; for any two asset nodes, comparing temporal fluctuation features through a pre-trained network; and using a continuously differentiable operator with a temperature annealing mechanism to transform the implicit dependencies between high-dimensional features into an explicit directed probability graph. The causal discovery formula is as follows: ,in, Representing time From asset nodes Pointing to node The probability of directed causal influence strength, global causal adjacency tensor ; and These represent the underlying physical state feature vectors collected by the two nodes at the current moment; (symbols) Indicates feature concatenation operation; The internally learnable parameters are Multilayer perceptual inference network; It is a differentiable discretization operator; Annealing temperature hyperparameters used to control the sparsity of the probability distribution; A directed topological physical isolation mask is constructed based on a high-precision map. To prevent the characteristics of the initiating node of a core event from being overwhelmed by the characteristics of a massive number of normal nodes around it, a causal out-degree matrix is ​​used to strictly constrain the flow of information transmission, enabling the result node to adaptively and directionally absorb key early warning information from the cause node, thereby ensuring the clarity of the asset traceability. The aggregation update formula is designed as follows: ,in, This is the causal high-order spatial feature matrix; It is a global causal adjacency tensor; It is a global input tensor; It is the inverse matrix; It is a learnable linear projection matrix; Mish is a self-regularized nonlinear activation function; For the bypass residual term, where It is an adaptive residual coefficient matrix; finally, a divine ordinary differential equation is introduced to completely abstract the evolution of asset characteristics on the time axis into a continuous mathematical integral process, thereby smoothly filling data gaps and accurately calculating asset time-related losses. Its continuous derivation formula is as follows: ,in, This represents the output from the sensing base module to the higher levels of the system, extrapolating to the future target time. The asset state manifold tensor; and These correspond to the current actual observation time and the expected future target time, respectively. For continuous-time variables within the integration domain; Representative at The instantaneous characteristic state at any given moment; It is a nonlinear differential function, given by the control parameters. A deep neural network is used for black-box approximation fitting.

3. The dynamic valuation method for highway digital assets based on a causal spatiotemporal model according to claim 2, characterized in that, The large language model scene semantic embedding for specific business needs includes, in order to accurately capture the hidden business intent and asset selection preferences in the requirement text, calling a pre-trained large language model as a semantic encoder to perform deep spatial embedding on the requirement input. The scene semantic embedding formula is as follows: , in, This represents the high-dimensional semantic representation matrix of the business scenario calculated in this step; For receiving raw natural language request text sequences from business systems or data buyers; It is a specially designed fine-tuning vector for prompts in the field of digital economy in transportation; The autoregressive encoding operator, representing the underlying large language model, is responsible for mapping discrete text into high-dimensional latent vectors. This is a dimension reduction projection matrix used to compress the ultra-high-dimensional semantic vectors output by large models into a dimension space that matches the underlying physical features; The bias term vector is used to prevent feature space shift. The Gaussian error linear unit activation function is used to introduce nonlinear representations and maintain a smooth transition in the semantic space, outputting a scene semantic representation matrix. This will serve as a core benchmark representing the demand side, directly inputting it into the next step to engage in value collision with physical characteristics.

4. The dynamic valuation method for highway digital assets based on a causal spatiotemporal model according to claim 3, characterized in that, The construction of the digital asset value entropy quantification network based on mutual information includes introducing mutual information and entropy reduction theory from information theory to construct a value entropy quantification network. By calculating the cross-mutual information of physical tensors and semantic tensors in the latent space and applying information divergence penalties, a dynamic value assessment coefficient is derived. The value entropy quantification formula is as follows: , in, This represents the calculated asset value entropy matrix; That is, the asset state manifold tensor output by continuous deduction; This is the semantic representation matrix of the extracted business scenario; and These are cross-domain alignment mapping matrices for physical and semantic features, respectively, used to eliminate underlying distributional differences in heterogeneous data; superscript This represents the transpose of a matrix to satisfy the rules of dot product operation; This is a scaling factor used to prevent gradient saturation caused by excessively large high-dimensional dot product results; Mutual information gain amplification factor; KL divergence is used for extremely rigorous calculations of the probability distribution of physical features. Semantic demand probability distribution Information entropy loss between them; This is the divergence penalty hyperparameter; To smoothly approximate the nonlinear activation function of ReLU, the output asset value entropy matrix is... This will serve as the core dynamic gating switch, guiding the next step of dimensional reduction operations.

5. The dynamic valuation method for highway digital assets based on a causal spatiotemporal model according to claim 4, characterized in that, The value entropy-guided cross-domain manifold dimensionality reduction and economic representation extraction includes using the generated asset value entropy as a dynamic gate to perform weighted filtering and nonlinear dimensionality reduction on the original physical tensor. Essentially, it involves removing redundant information in the physical tensor related to the current business scenario, retaining only the core causal anchors with high value entropy, and finally merging them into a unified economic representation. Its manifold dimensionality reduction formula is as follows: , in, The digital asset economic representation vector representing the output of this step; It is the transmitted asset state manifold tensor; The calculated asset value entropy matrix; symbols It represents the Hadamardi (or Hadama) stack; For the dimensionality reduction and compression matrix of the economic manifold; For layer normalization operations to be stable; operator This represents the fusion and addition of vector dimensions, combining the compressed high-value physical features with the original business scenario semantics. Deep binding is performed, resulting in the final economic representation vector. .

6. The dynamic valuation method for highway digital assets based on a causal spatiotemporal model according to claim 5, characterized in that, The instantiation and utility benchmark initialization of the federated agent intelligent agents include establishing the initial bargaining power and utility benchmarks for the three participating intelligent agents before the dynamic game, initializing three independent federated intelligent agents in parallel, and calculating their respective initial utility functions based on the economic representations passed in from the previous module. The utility initialization formulas are as follows: , in, Representing intelligent agents , These represent the seller, the buyer, and the regulator, respectively. It is the obtained digital asset economic representation vector; For this specific intelligent agent Built-in business objective function vector, superscript For the transpose operation, by using... The dot product operation is used to measure the alignment between the current data and its own business interests; the denominator part The L2 norm of the solution vector is used to normalize the dot product result using standard cosine similarity. The system pre-sets scaling factors for the aggressiveness of each participant to simulate the risk preferences of entities under different market conditions; For federal compliance penalties; It is a non-linear activation function responsible for smoothly compressing the calculation results into a probability range of 0 to 1.

7. The dynamic valuation method for highway digital assets based on a causal spatiotemporal model according to claim 6, characterized in that, The multi-agent game adversarial network under the supply and demand double-blind condition includes constructing a global game adversarial loss function, which forces the three parties to continuously adjust their bid tensors in multiple iterations to minimize the global divergence. The formula for the global game adversarial loss is as follows: , in, For real-time calculation of scalar values, used for quantization at the first... The overall sense of division and disagreement among the three parties at the negotiating table regarding the pricing of digital assets; conditions for seeking peace. Ensure the system cross-calculates conflicts between the seller, buyer, and regulator; and Representing intelligent agents and The utility confidence scalar in the current round; fractional term It is an asymmetric weighting mechanism based on utility confidence, with constants. The introduction of this is to absolutely prevent system crashes caused by a denominator of zero; and Representing intelligent agents and The asset pricing high-dimensional tensor advocated in the current round; The square of the Frobenius norm, representing the difference between these two tensors, is the valuation gap as subjectively perceived by each party. These are specially designed regulatory-anchored penalties, among which... Specifically refers to the tensor of the specific pricing proposition given by the seller's agent in the current round. This specifically refers to the reference floor pricing tensor, which possesses absolute fairness and is provided by the regulatory agent in the current round, using the L1 norm. With penalty coefficient This is to prevent potential collusion between buyers and sellers or extreme price gouging.

8. The dynamic valuation method for highway digital assets based on a causal spatiotemporal model according to claim 7, characterized in that, The Nash equilibrium convergence and dynamic asset valuation truth generation process involves, after calculating the global game disagreement loss, the three agents modifying their bids along the gradient descent direction. This simulates the dynamic game behavior in real business negotiations, where agents gradually probe and make concessions based on the other party's bottom line. Through hundreds of rounds of tensor clashes, when the rate of change of the global disagreement loss approaches a small system threshold, the engine automatically determines that the negotiation has reached a Nash equilibrium. The final compromise bids of the three parties are extracted, and a confidence-weighted soft maximization mechanism is used to generate an indisputable and unique asset valuation truth. The bid updates are as follows: ,in, and Representing intelligent agents The pricing tensor of the next round versus the current round; The learning rate hyperparameter for system game theory controls the step size of a single concession; partial derivative terms It is the direction of the mathematical gradient that best resolves disagreements; symbol For the Hadamard product, the part in parentheses For dynamic compromise control gate, Given a pre-defined rigidity matrix for the agent's personality, when one party's initial confidence is extremely high, this gating value approaches zero, instantly blocking backpropagation and thus refusing to lower the price. The final valuation formula is as follows: ,in, This is the central consensus valuation truth matrix that is output to the outside, representing the absolute transaction value of the digital asset in this business scenario; This is the final compromise card for all parties after the Nash equilibrium converges; and The confidence level retained at the convergence time, combined with the natural index. The normalized weighting ensures that the bid from the side with strong evidence receives a larger proportion of the weighting. This is the game theory temperature coefficient, used to differentiate confidence levels.

9. The dynamic valuation method for highway digital assets based on a causal spatiotemporal model according to claim 8, characterized in that, The digital asset scheduling and value-added evolution based on the true value of asset dynamic valuation under stochastic differential equation constraints includes fitting drift and diffusion terms through deep neural networks to search for the optimal scheduling action vector that maximizes global long-term economic returns from a massive number of scheduling strategies. Its stochastic differential equation modeling is as follows: in, Represents a tiny time period The differential form of the expected potential increase in the commercial value of the digital asset; It is a deterministic drift term function parameterized by a deep neural network; Fit the network to the diffusion term; This is a standard Wiener process; its strategy optimization formula is: in, This is the feature vector of the globally optimal scheduling action; Operators are responsible for the system's vast library of legal actions. We then perform rigorous traversal and gradient optimization. The expression represents the expected value of the calculation; the integral term represents the cumulative increase in the asset value over an infinite time span, where... It introduces the time discount rate. Continuous discount factor; As a continuous-time variable; then, using the momentum smoothing feature update equation, the crystallization of this federated game evaluation and the optimal scheduling strategy are deeply physically integrated, completely completing the spatiotemporal causal closed loop of the entire large model evaluation system. Its feature file update formula is: ,in, and These represent the high-dimensional feature matrices of the long-term utility history archive before and after the current state increment update; The system's manually preset translational slip coefficient parameters; symbol For feature channel splicing operations; To maintain the consistency of the latent space dimension of the archive features, the alignment projection update matrix is ​​used. It is a hyperbolic tangent nonlinear activation function.

10. A dynamic valuation system for highway digital assets based on a causal spatiotemporal model, characterized in that, include: The data acquisition module is configured to acquire multi-source heterogeneous sensor data; The feature evolution module is configured to perform causal spatiotemporal topology reconstruction and feature evolution based on structure discovery, according to the acquired data. The dimensionality reduction module is configured to perform cross-domain manifold dimensionality reduction from modal physical perception to asset value entropy based on evolution results. This includes semantic embedding of large language model scenarios for specific business needs, construction of digital asset value entropy quantification network based on mutual information, and cross-domain manifold dimensionality reduction and economic representation extraction guided by value entropy. The game module is configured to perform supply and demand double-blind dynamic pricing game based on the representation results, including the instantiation of the federated agent and the initialization of the utility benchmark, the multi-agent game adversarial network under the condition of supply and demand double-blind, the convergence of Nash equilibrium and the generation of the true value of dynamic asset valuation. The value-added module is configured to perform digital asset scheduling and value-added evolution under the constraints of stochastic differential equations based on the true value of asset dynamic valuation. The output module is configured to output the true value of the asset valuation after the increase.

Images