A method for quantitatively locating investment risks in power grid engineering based on dual-drive intelligent agents
By using a dual-driven intelligent agent approach, multi-source heterogeneous data and large language models are employed to quantify and locate investment risks in power grid projects. This addresses the shortcomings of existing technologies in risk probability modeling and location, achieving accurate risk analysis and location, and improving data utilization and the scientific nature of decision-making.
Patent Information
- Application Number
- CN202611309010.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-27
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies lack accurate risk probability modeling and risk source location in power grid engineering investment decisions. Expert judgment has blind spots, and general large language models are prone to making wrong suggestions when processing data from vertical power grid sectors. Existing early warning systems cannot provide the probability distribution and accurate location of risks.
A dual-driven intelligent agent approach is adopted. By collecting multi-source heterogeneous data, performing sparse-dense hybrid retrieval and reordering, knowledge-enhanced context feature vectors are generated. A large language model is used to perform causal reasoning to generate multiple paths in parallel. Logical aggregation and physical constraint verification are performed. Combined with an integrated gradient algorithm, the attribution contribution is calculated and mapped to the power grid supply chain map for risk positioning.
It enables accurate modeling of risk probabilities, effective sensitivity analysis, and precise location of risk sources in power grid engineering investment decisions, avoids subjective bias from a single expert, provides probabilistic recommendations based on global data, and improves data utilization.
Smart Images

Figure CN123117791A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of deep integration of power system automation and artificial intelligence, and more specifically to a method for quantitatively locating investment risks in power grid engineering based on dual-drive intelligent agents. Background Technology
[0002] Currently, power grid projects, especially ultra-high voltage (UHV) DC / AC transmission projects, are among the most complex industrial systems in the world. Their investment decisions not only involve the allocation of funds in the tens of billions of dollars, but are also affected by hundreds of dynamic dimensions such as policy guidance, raw material market fluctuations, supply chain robustness, environmental constraints, and construction cycles.
[0003] For a long time, power grid investment has relied mainly on the intuition and experience of senior experts. This experience is often fragmented and difficult to translate into algorithmic models. Moreover, when faced with unprecedented compound risks such as global supply chain disruptions coupled with sudden changes in carbon neutrality policies, the experts' judgments may have blind spots.
[0004] In recent years, with the development of large language models, although generative AI has made breakthroughs, when processing data in the vertical field of power grids, general models are prone to "hallucination" due to the lack of specific physical constraints and power industry logic, outputting seemingly reasonable but actually wrong engineering suggestions.
[0005] Furthermore, existing early warning systems are mostly simple threshold alarms based on rules, which cannot provide a probability distribution of risk occurrence, let alone accurately locate the "first trigger point of risk" in the intricate supply chain network.
[0006] Therefore, how to achieve accurate risk probability modeling, effective sensitivity analysis, and precise risk source location in power grid engineering investment decision-making is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0007] In view of the above problems, this invention proposes a method for quantitatively locating investment risks in power grid engineering based on dual-drive intelligent agents, aiming to overcome or at least partially solve the above problems.
[0008] To achieve the above objectives, the present invention adopts the following technical solution:
[0009] This invention provides a method for quantifying and locating investment risks in power grid engineering based on dual-drive intelligent agents, comprising: Collect multi-source heterogeneous data related to power grid engineering investment; and perform sparse-dense hybrid retrieval and re-sorting on the multi-source heterogeneous data to obtain contextual data containing textual and numerical features; The text features and numerical features are adaptively fused through a dynamic gating network to generate a knowledge-enhanced context feature vector. The knowledge-enhanced context feature vector is input into the large language model, and multiple causal reasoning paths are generated in parallel through a self-consistent thinking chain. Logical aggregation and physical constraint verification are performed on the multiple causal reasoning paths to obtain the final risk conclusion and key sensitive nodes. Using the key sensitive nodes as input features, the attribution contribution of each input feature is calculated using an integrated gradient algorithm; the attribution contribution is then mapped to material nodes or contractors in the power grid supply chain map to generate risk location results.
[0010] Furthermore, the multi-source heterogeneous data includes data from the power grid's internal ERP system, external futures market price data, policy document data, and geographic information data.
[0011] Furthermore, the sparse-dense hybrid retrieval and reordering of the multi-source heterogeneous data specifically includes: Based on the target question of the investment risk quantification and positioning, a query statement is generated. The SPLADE++ algorithm is used to map the text documents in the multi-source heterogeneous data to an ultra-high-dimensional vocabulary space, and the sparse correlation score between the query statement and each text document is calculated. The Contriever encoder is used to map the numerical sequence data in the multi-source heterogeneous data to a low-dimensional continuous vector space, and the dense similarity score between the query statement and each numerical feature vector is calculated. The sparse correlation score and dense similarity score are weighted and fused by a dynamic weighting factor to obtain a preliminary search result set. The preliminary retrieval result set is then subjected to secondary refinement using the DP-BERT cross-encoder to output the context data.
[0012] Furthermore, the knowledge-enhanced context feature vector Represented as:
[0013] in, Indicates a gating network; Indicates text features Perform nonlinear mapping or dimensional transformation; This indicates a nonlinear mapping or dimensional transformation of numerical features; This indicates element-wise multiplication.
[0014] Furthermore, the step of logically aggregating and physically constraining the multiple causal reasoning paths to obtain risk conclusions and key sensitive nodes specifically includes: A majority voting mechanism is used to statistically aggregate the initial risk conclusions of the multiple causal reasoning paths, and the most frequently occurring initial risk conclusions are selected as candidate risk conclusions. Construct a knowledge base of physical constraints that includes physical limits of the infrastructure construction cycle and mandatory standards for power construction; The temporal logic and resource consumption parameters in the candidate risk conclusions are matched and verified with the physical constraint knowledge base. The candidate risk conclusions that pass the verification are taken as the final risk conclusions, and the material nodes on the path corresponding to the final risk conclusions are extracted as key sensitive nodes.
[0015] Furthermore, the attribution contribution is expressed as:
[0016] in, For the first Attribution contribution of key sensitive nodes; This represents the overall actual risk state vector, i.e., all local features. A high-dimensional vector formed by the joint structure; Represents the overall risk-free baseline state vector; For the first The actual risk status input vector of each key sensitive node; For the first Risk-free baseline state vector of each key sensitive node; F For risk prediction function; The number of discretization steps; For the first A discrete point.
[0017] Furthermore, after obtaining the key sensitive nodes, the method further includes: inputting the key sensitive nodes into the trained recognition model to generate a risk probability density function, and calculating the self-consistency score of investment risk based on the risk probability density function; when the self-consistency score is greater than a preset threshold, triggering manual review.
[0018] Furthermore, the variational lower bound objective function of the recognition model is expressed as:
[0019] in, This is the variational lower bound; To identify encoder parameters in the model; These are the decoder parameters in the generated model; This represents the observation data corresponding to key sensitive nodes, which serves as the input to the identification model; As a potential risk factor; This is the expected value; For reconstruction terms, it means the decoder is based on potential risk factors. Reconstruct the observation data Effect value; Let KL divergence be a metric. This is the approximate posterior distribution output by the encoder; This represents the prior distribution.
[0020] Furthermore, the self-consistency score is expressed as:
[0021] in, Indicates the self-consistency score; This represents the total number of potential risk factors; This represents the model prediction result for the t-th potential risk factor; express The average of the prediction results from the sub-models; Indicates the first Confidence level of cognitive uncertainty in the prediction.
[0022] Furthermore, it also includes: The system receives corrected feedback data output by an expert terminal based on the risk localization result or the risk probability density function, and constructs a fine-tuning loss function based on the corrected feedback data; the fine-tuning loss function includes task loss driven by expert feedback and regularization loss; The parameters of the large language model are updated based on the fine-tuning loss function using a low-rank adaptive algorithm.
[0023] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method for quantitatively locating investment risks in power grid engineering based on dual-drive intelligent agents, which has the following beneficial effects: This invention utilizes massive heterogeneous data for perception and leverages the CoT (Co-Learning Logic) of a large model for logical deduction, thereby achieving accurate risk probability modeling, effective sensitivity analysis, and precise risk source location in power grid engineering investment decisions.
[0024] This invention introduces expert feedback as a reward signal for closed-loop optimization, avoiding subjective bias from a single expert and providing probabilistic recommendations based on global data.
[0025] This invention transforms a large amount of dormant historical texts, contracts, and audit reports within the power grid into quantifiable decision weights, thereby improving data utilization. Attached Figure Description
[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0027] Figure 1 This is a schematic diagram of the process for a method to quantify and locate investment risks in power grid engineering based on dual-drive intelligent agents, provided in an embodiment of the present invention. Detailed Implementation
[0028] 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.
[0029] Example 1: This invention discloses a method for quantitatively locating investment risks in power grid engineering based on dual-drive intelligent agents, such as... Figure 1 As shown, the process includes the following steps: collecting multi-source heterogeneous data related to power grid engineering investment; performing sparse-dense hybrid retrieval and reordering on the multi-source heterogeneous data to obtain contextual data containing textual and numerical features; adaptively fusing the textual and numerical features through a dynamic gating network to generate a knowledge-enhanced contextual feature vector; inputting the knowledge-enhanced contextual feature vector into a large language model to generate multiple causal reasoning paths in parallel through a self-consistent thinking chain; performing logical aggregation and physical constraint verification on the multiple causal reasoning paths to obtain the final risk conclusion and key sensitive nodes; using the key sensitive nodes as input features, calculating the attribution contribution of each input feature using an integrated gradient algorithm; mapping the attribution contribution to material nodes or contractors in the power grid supply chain graph to generate risk location results.
[0030] Next, each of the above steps will be explained in detail.
[0031] I. Multi-source data acquisition: First, multi-source heterogeneous data related to power grid engineering investment is collected. In this embodiment, multi-source heterogeneous data includes, but is not limited to: data from the power grid's internal ERP system, external futures market price data, policy document data, and geographic information data. These data come from diverse sources and have different formats, including both unstructured text and structured numerical time-series data, providing a data foundation for subsequent heterogeneous data fusion.
[0032] II. Sparse-Dense Hybrid Retrieval and Reordering: After acquiring heterogeneous data from multiple sources, the system needs to quickly retrieve contextual information highly relevant to the current investment decision problem. This embodiment employs a dynamic sparsity approach. The dense hybrid retrieval mechanism includes the following sub-steps: (1) Generate query statement: Based on the target problem of investment risk quantification and positioning, the system automatically generates structured query statement. ; (2) Sparse Retrieval: For text documents in multi-source heterogeneous data, such as equipment ledgers, technical specifications, and policy documents in the power grid's internal ERP system, which are highly specialized texts, sparse retrieval is performed using the SPLADE++ algorithm. This SPLADE++ algorithm maps the text to an ultra-high-dimensional vocabulary space through term bottleneck constraints, assigning extremely high activation weights to power-specific terms (such as "ultra-high voltage converter transformer"). Query Statement Sparse correlation score between the document and any text document Represented as:
[0033] in, Used to represent text documents; and They represent the query statements respectively. With text documents Activation intensity in the extended vocabulary space.
[0034] (3) Dense Retrieval: For numerical sequence data in multi-source heterogeneous data, such as data with implicit causal relationships, such as descriptions of futures market price fluctuations, macroeconomic analysis reports, and news information, the system first converts them into a form that facilitates semantic matching, and then uses the Contriever encoder to map the converted numerical sequence data to a low-dimensional continuous vector space; dense similarity score The cosine similarity is used for calculation, and is expressed as:
[0035] In this formula, This represents solving for cosine similarity; This represents the semantic representation vector corresponding to the query statement obtained through deep neural network learning. This represents the semantic representation vector corresponding to the numerical feature vector obtained through deep neural network learning. Based on this, the dense similarity score can identify the semantic relevance between "Chilean environmental policy" and "copper supply disruption", even if there are no overlapping words.
[0036] (4) Dynamic weighted fusion: The sparse relevance score and dense similarity score are weighted and fused by a dynamic weighting factor to obtain a preliminary retrieval result set, which is represented as:
[0037] in, This represents the score after weighted merging; This indicates an adaptive parameter adjustment; it increases when the query contains a large number of technical terms. Reduce when the query focuses on trend analysis Based on the combined scores, the system initially screens out the top 100-500 candidate document segments.
[0038] (5) DP-BERT Reordering: To further improve the accuracy of recall, the system uses a DP-BERT cross-encoder to perform secondary reordering on the initial retrieval result set. Specifically, DP-BERT concatenates the query statement Q with each candidate document and inputs them into the network simultaneously, using a cross-attention mechanism to capture the subtle semantic interactions between the two, effectively distinguishing easily confused concepts such as "transformer oil" and "edible oil". After reordering, the system selects the Top-K documents with the highest scores as the final highly relevant contextual data, ensuring that truly critical power policy and pricing cases are placed at the forefront for use by subsequent modules.
[0039] The top-K results with the highest scores after reordering are input into the alignment layer to associate text and numerical values, generating contextual data. This contextual data can be represented in the form of knowledge-enhanced feature tuples, including: {event semantics, numerical volatility, timestamp alignment, risk probability contribution}. These knowledge-enhanced feature tuples can be used as direct input to the subsequent logical deduction agent, ensuring that each logical statement in the large model has precise numerical support when deriving risk.
[0040] In summary, this invention, in the sparse-dense hybrid retrieval and re-ranking steps, constructs a dual-track parallel computing system by extracting differentiated features from multi-source data. It utilizes sparse relevance scores to locate keywords (e.g., "policy issued on a certain day of a certain month") and dense scores to correlate numerical trends (e.g., "prices fluctuated sharply after that day"), effectively solving the problem that "traditional vector retrieval (Dense Retrieval) easily overlooks key keywords, while keyword retrieval (Sparse Retrieval) lacks semantic understanding."
[0041] III. Cross-modal dynamic gated fusion: After encoding the text and numerical data in the retrieved context data, the corresponding text features are obtained. Sum of numerical characteristics To adaptively fuse information from these two modalities, this embodiment designs a dynamically gated fusion network. :
[0042] in, This represents the Sigmoid activation function, whose output value is in Between these values, the weights represent the text features. The gating weight matrix is a learnable parameter matrix used to learn text features. With numerical characteristics The interaction between them, such as determining how much weight the system should allocate to "policy description" or "price fluctuation" in the current investment environment; This represents the gating bias vector, which provides a basic offset for the gating mechanism, ensuring that the gating still has a basic logical bias when the feature input is zero or extremely small. This indicates a vector concatenation operation, which combines encoded text features and numerical features into a longer feature vector so that... It can be processed uniformly.
[0043] A knowledge-enhanced contextual feature vector is generated by adaptively fusing textual and numerical features through a dynamic gating network. Represented as:
[0044] in, Indicates a gating network; This is a text transformation function, representing the transformation of text features. The purpose of performing nonlinear mapping or dimensional transformation is to adjust the semantic information of the text into a latent feature space that is unified with the numerical information, so as to perform weighted summation; For numerical transformation functions, it represents a nonlinear mapping or dimensional transformation of numerical features, with the aim of aligning the dimensions of the original numerical features (such as futures prices) with those of the text features in a contextual manner. This indicates element-wise multiplication; This knowledge-enhanced contextual feature vector It already includes gating weight information, which can dynamically reflect the relative importance of "policy-driven" (such as policy bans) and "market-driven" (such as rising copper prices) in the current investment environment.
[0045] IV. Logical deduction based on self-consistent thought chain (CoT): This invention introduces a self-consistent framework to address the "single-path illusion" problem in general large models when dealing with complex power grid investment logic. By performing multi-dimensional causal simulation based on the high-quality context produced by the "cross-modal knowledge enhancement layer," it achieves a leap from "data association" to "logical verification." Specifically, it includes the following sub-steps: (1) Multi-path parallel sampling: The inference engine employs a non-zero temperature sampling strategy (e.g., Temperature=0.7), based on the same knowledge-enhanced context feature vector. Simultaneous parallel generation Each independent causal reasoning path generates a path set. ,in, Indicates the first There are 3 causal reasoning paths; each causal reasoning path contains entity nodes and logical relationships; furthermore, each causal reasoning path simulates expert decision-making logic, for example: Path A (Policy Bias Path): Policy Fluctuations Construction permit extension Increased investment. This approach focuses on analyzing the administrative compliance risks arising from policy changes and the resulting increase in construction time and costs.
[0046] Path B (Market Bias Path): Raw Material Fluctuations Supplier default The timeframe is extended. This approach focuses on analyzing supplier defaults and supply chain resilience risks caused by fluctuations in raw material (such as copper and steel) prices.
[0047] (2) Majority voting aggregation: A majority voting mechanism is used to statistically aggregate the initial risk conclusions of multiple causal reasoning paths, and the most frequently occurring initial risk conclusions are selected as candidate risk conclusions. For example, if more than 70% of the reasoning paths point to the same risk conclusion (such as "total investment overrun risk"), the system marks that conclusion as a "high-confidence confirmed conclusion". By analyzing the key nodes that appear in common in different paths (such as "supplier A's delivery delay"), the system identifies the essential links in the risk evolution process.
[0048] (3) Physical constraint verification and output: A physical constraint knowledge base is constructed, encompassing the physical limits of the infrastructure construction cycle and mandatory standards for power construction. The temporal logic and resource consumption parameters in candidate risk conclusions are matched and validated against this physical constraint knowledge base. For example, if the predicted construction period is shortened below the physical limit, the system will determine that the path is invalid. Based on this, only causal chains that conform to power construction standards, have sufficient data support, and are logically consistent are retained. The validated candidate risk conclusions are used as the final risk conclusions, and the material nodes on the paths corresponding to the final risk conclusions are extracted as key sensitive nodes.
[0049] The output content and format of this physical constraint verification are shown in Table 1 below. In this way, the entire process from macro-level policy perception to micro-level material node positioning is completed in sub-second time. It not only tells decision-makers "what happened," but also accurately explains "why it happened" and "how it will evolve in the future."
[0050] Table 1: Output Content and Format of Physical Constraint Verification
[0051] V. Precise Location of Risk Sources Based on Integrated Gradients: After obtaining the risk probability, the system needs to "revert" the abstract risk value back to the specific material nodes or construction units in the physical world. This embodiment uses the Integrated Gradients (IG) algorithm, taking the key sensitive nodes obtained above as input feature dimensions, to calculate the attribution contribution of each feature to the final risk conclusion.
[0052] Let the actual risk state input vector be... (Constituted from the current values of key sensitive nodes), the risk-free baseline state vector is... (For example, when copper prices are at their historical average and supplier deliveries are normal), the risk prediction function is: F Then the first The formula for calculating the ensemble gradient of the input features is:
[0053] This calculation process is actually a mathematical verification of path A (policy bias) or path B (market bias) generated by the "logic deduction layer," by calculating the gradient contribution of each causal node. Which link in the quantization logic chain is the real "fatal cause"? The calculated gradient contribution... Map back to specific nodes in the supply chain map.
[0054] In the specific implementation process, the gradient contribution will be... The formula is integrated, discretized, and summed to obtain the attribution contribution, which is expressed as:
[0055] in, For the first Attribution contribution of key sensitive nodes; This represents the overall actual risk state vector, i.e., all local features. A high-dimensional vector formed by the joint structure; Represents the overall risk-free baseline state vector; For the first The actual risk status input vector of each key sensitive node; For the first Risk-free baseline state vector of each key sensitive node; F For risk prediction function (i.e., risk probability density function (PDF)); The number of discretization steps; For the first A discrete point.
[0056] During the visualization process, the calculated attribution contribution can be presented. The mapping back to specific nodes in the supply chain topology determined by the logic layer. In the investment plan or material supply schedule, specific categories or contractors with different risk contribution levels (red being the highest) are marked with red, orange, and yellow. This positioning result ultimately forms a confirmation conclusion, which is then fed back to the "co-evolution layer" as a direct basis for experts to revise the model or update LoRA parameters.
[0057] In practical applications, especially in ultra-high voltage direct current projects, if the converter valve supplier's production capacity is limited, the algorithm will automatically calculate the contribution weight of this risk to the "overall commissioning time" and mark the supplier node in red in the supply chain map according to the weight.
[0058] Example 2: Based on Example 1, this second embodiment, after obtaining the key sensitive nodes, further includes: risk probability modeling and uncertainty quantification; specifically: This invention does not blindly fit probabilistics to the original data, but rather uses the evolutionary causal logic chain output by the "logic deduction agent layer" as a constraint to achieve a precise mapping from qualitative logic to quantitative distribution. Specifically, the key sensitive nodes obtained above are used as input dimensions of the observed data to construct a variational inference model to quantify risk probability and assess cognitive uncertainty. This includes the following sub-steps: (1) Risk feature modeling based on latent space: Since grid risk is not an isolated variable, but is driven by multiple potential implicit factors (such as policy orientation, global liquidity, and engineering resilience), a variational inference mechanism is introduced to capture these factors that cannot be directly observed.
[0059] Let the observed data be X (i.e., the multidimensional feature vector corresponding to key sensitive nodes), and the potential risk factors be z (i.e., implicit variables that cannot be directly observed, such as "global liquidity," "policy stringency," and "engineering resilience"). The goal of this invention is to infer the posterior distribution. Since direct computation is not feasible, the system constructs an encoder to approximate the true posterior and trains it by maximizing the lower bound of evidence (ELBO); the variational lower bound objective function of this model is:
[0060] in, This is the variational lower bound; To identify encoder parameters in the model; These are the decoder parameters in the generated model; This represents the observation data corresponding to key sensitive nodes, which serves as the input to the identification model; As a potential risk factor; This is the expected value; For reconstruction terms, it means the decoder is based on potential risk factors. Reconstruct the observation data Effect value; Let KL divergence be a metric. This is the approximate posterior distribution output by the encoder; This represents the prior distribution. In this formula, For restructuring items, the aim is to ensure potential risk factors This can fully explain the observed engineering anomalies; The regularization term is designed to force the latent space to conform to a prior distribution (such as a Gaussian distribution), ensuring the smoothness of the risk space and preventing model overfitting.
[0061] (2) Generation and prediction of the risk probability density function (PDF): In the prediction phase, the system starts from the posterior distribution. T potential risk factors z(t) are sampled, and each potential risk factor z(t) is processed by a risk prediction head. This is mapped to a risk prediction value (such as the investment overspending rate). The final risk probability density function is approximated by Monte Carlo sampling to calculate the risk probability. :
[0062] in, This indicates the rate of cost overrun or the probability of construction delays. This represents the input to the recognition model.
[0063] (3) Quantification of uncertainty: To determine the model's confidence level in its current predictions, the system employs Monte Carlo Dropout sampling to calculate cognitive uncertainty. During the testing phase, T random inactivation forward propagations are performed on the identification network or risk prediction head, yielding T prediction results; based on these, the following calculations are performed: Average risk level : ; Cognitive uncertainty confidence level : ; The prediction variance obtained here via MC Dropout As a confidence level for cognitive uncertainty, if the variance is too large, it essentially reflects the generation of uncertainty by the "logic deduction agent layer" when processing this task. There are serious logical conflicts among the reasoning paths, resulting in a low self-consistency score. This self-consistency score is expressed as:
[0064] in, Indicates the self-consistency score; This represents the total number of potential risk factors; This represents the model prediction result for the t-th potential risk factor; express The average of the prediction results from the sub-models; Indicates the first Confidence level of cognitive uncertainty in the prediction.
[0065] When self-consistency score When the value exceeds a preset threshold, the system determines that the current logic chain cannot support deterministic decision-making, thereby forcibly triggering the "manual review" logic, thus avoiding blind investment caused by the "illusion" of large models.
[0066] Example 3: This third embodiment, based on the second embodiment described above, further includes: to incorporate prior knowledge and corrective opinions from experts, the system establishes a human-in-the-Loop and continuous evolution mechanism. (1) Closed-loop processing of expert feedback: The system interface provides an "Expert Confirmation / Correction" function. Experts can review the risk probability density function or risk positioning results obtained above, for example, by lowering the weight of "Supplier A's cash flow risk" (reason: this supplier has a state-owned enterprise background, and the actual risk is relatively small). Each action by the expert will be converted into correction feedback data.
[0067] (2) Construction of fine-tuning loss function: Based on the corrected feedback data, the system constructs an expert feedback-driven task loss. (For example, cross-entropy loss or ranking loss). To preserve the generality of the model while updating the parameters of a large language model, this embodiment employs Low-Rank Adaptive (LoRA) fine-tuning. LoRA inserts two low-rank matrices A and B into the hidden layers of the pre-trained model, updating only the parameters of matrices A and B while keeping the main parameters of the original pre-trained model frozen. The total fine-tuning loss function is... Represented as:
[0068] in, Task loss driven by expert feedback; The regularization coefficient is introduced to prevent overfitting to individual extreme cases of experts. A low-rank matrix The Frobenius norm; A low-rank matrix The Frobenius norm is used to penalize the parameter weights to ensure that the fine-tuned model enhances the power grid's professional judgment while retaining the generalization ability of the basic model.
[0069] The following example, "investment risk in the converter transformer supply chain" in a certain ultra-high voltage direct current transmission project, fully demonstrates the method flow of this invention.
[0070] (1) Initial input: The system monitors in real time that a major copper exporting country has issued a new environmental protection law for mining.
[0071] (2) Hybrid retrieval: The system automatically retrieves data on the impact of similar laws from 10 years ago on copper prices, as well as the degree of dependence on copper materials in current ultra-high voltage projects under construction.
[0072] (3) Reasoning logic: Logical chain 1: Environmental protection bill Production capacity decline Copper prices rise The cost of converter transformers increased by 15%. Risk of total investment exceeding budget.
[0073] Logic Chain 2: Copper Prices Rise Supplier A is under great financial pressure Delivery delay risk Risk of project shutdown.
[0074] (4) Quantitative positioning: Probabilistic modeling shows that the "overspending risk" is 85% and the "work stoppage risk" is 30%. IG positioning shows that the source of risk is "supplier A's cash flow".
[0075] (5) Decision recommendations: It is recommended to lock in copper prices in advance for hedging, or increase the share of alternative supplier B.
[0076] (6) Expert correction: The expert believes that supplier A has a state-owned enterprise background and has less financial risk, so the weight is adjusted. The system records this logical deviation through LoRA and automatically optimizes it in the next inference.
[0077] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0078] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for quantitatively locating investment risks in power grid engineering based on dual-drive intelligent agents, characterized in that, include: Collect multi-source heterogeneous data related to power grid engineering investment; The multi-source heterogeneous data is then subjected to sparse-dense hybrid retrieval and reordering to obtain contextual data containing textual and numerical features. The text features and numerical features are adaptively fused through a dynamic gating network to generate a knowledge-enhanced context feature vector. The knowledge-enhanced context feature vector is input into the large language model, and multiple causal reasoning paths are generated in parallel through a self-consistent thinking chain. Logical aggregation and physical constraint verification are performed on the multiple causal reasoning paths to obtain the final risk conclusion and key sensitive nodes. Using the key sensitive nodes as input features, the attribution contribution of each input feature is calculated using an integrated gradient algorithm. The attribution contribution is mapped to material nodes or contractors in the power grid supply chain map to generate risk positioning results.
2. The method for quantifying and locating investment risks in power grid engineering based on dual-drive intelligent agents as described in claim 1, characterized in that, The multi-source heterogeneous data includes data from the power grid's internal ERP system, external futures market price data, policy document data, and geographic information data.
3. The method for quantifying and locating investment risks in power grid engineering based on dual-drive intelligent agents as described in claim 1, characterized in that, The process of performing sparse-dense hybrid retrieval and reordering on the multi-source heterogeneous data specifically includes: Based on the target question of the investment risk quantification and positioning, a query statement is generated. The SPLADE++ algorithm is used to map the text documents in the multi-source heterogeneous data to an ultra-high-dimensional vocabulary space, and the sparse correlation score between the query statement and each text document is calculated. The Contriever encoder is used to map the numerical sequence data in the multi-source heterogeneous data to a low-dimensional continuous vector space, and the dense similarity score between the query statement and each numerical feature vector is calculated. The sparse correlation score and dense similarity score are weighted and fused by a dynamic weighting factor to obtain a preliminary search result set. The preliminary retrieval result set is then subjected to secondary refinement using the DP-BERT cross-encoder to output the context data.
4. The method for quantifying and locating investment risks in power grid engineering based on dual-drive intelligent agents as described in claim 1, characterized in that, The knowledge-enhanced context feature vector Represented as: in, Indicates a gating network; Indicates text features Perform nonlinear mapping or dimensional transformation; This indicates a nonlinear mapping or dimensional transformation of numerical features; This indicates element-wise multiplication.
5. The method for quantifying and locating investment risks in power grid engineering based on dual-drive intelligent agents as described in claim 1, characterized in that, The process of logically aggregating and physically constraining the multiple causal reasoning paths to obtain risk conclusions and key sensitive nodes specifically includes: A majority voting mechanism is used to statistically aggregate the initial risk conclusions of the multiple causal reasoning paths, and the most frequently occurring initial risk conclusions are selected as candidate risk conclusions. Construct a knowledge base of physical constraints that includes physical limits of the infrastructure construction cycle and mandatory standards for power construction; The temporal logic and resource consumption parameters in the candidate risk conclusions are matched and verified with the physical constraint knowledge base. The candidate risk conclusions that pass the verification are taken as the final risk conclusions, and the material nodes on the path corresponding to the final risk conclusions are extracted as key sensitive nodes.
6. The method for quantifying and locating investment risks in power grid engineering based on dual-drive intelligent agents as described in claim 1, characterized in that, The attribution contribution is expressed as follows: in, For the first Attribution contribution of key sensitive nodes; This represents the overall actual risk state vector, i.e., all local features. A high-dimensional vector formed by the joint structure; Represents the overall risk-free baseline state vector; For the first The actual risk status input vector of each key sensitive node; For the first Risk-free baseline state vector of each key sensitive node; F For risk prediction function; The number of discretization steps; For the first A discrete point.
7. The method for quantifying and locating investment risks in power grid engineering based on dual-drive intelligent agents as described in claim 1, characterized in that, After obtaining the key sensitive nodes, the process further includes: inputting the key sensitive nodes into the trained recognition model to generate a risk probability density function, and calculating the self-consistency score of investment risk based on the risk probability density function; when the self-consistency score is greater than a preset threshold, triggering manual review.
8. The method for quantifying and locating investment risks in power grid engineering based on dual-drive intelligent agents as described in claim 7, characterized in that, The variational lower bound objective function of the recognition model is expressed as: in, This is the variational lower bound; To identify encoder parameters in the model; These are the decoder parameters in the generated model; This represents the observation data corresponding to key sensitive nodes, which serves as the input to the identification model; As a potential risk factor; This is the expected value; For reconstruction terms, it means the decoder is based on potential risk factors. Reconstruct the observation data Effect value; Let KL divergence be a metric. This is the approximate posterior distribution output by the encoder; This represents the prior distribution.
9. The method for quantifying and locating investment risks in power grid engineering based on dual-drive intelligent agents as described in claim 7, characterized in that, The self-consistency score is represented as: in, Indicates the self-consistency score; This represents the total number of potential risk factors; This represents the model prediction result for the t-th potential risk factor; express The average of the prediction results from the sub-models; Indicates the first Confidence level of cognitive uncertainty in the prediction.
10. The method for quantifying and locating investment risks in power grid engineering based on dual-drive intelligent agents as described in claim 7, characterized in that, Also includes: The system receives corrected feedback data output by an expert terminal based on the risk localization result or the risk probability density function, and constructs a fine-tuning loss function based on the corrected feedback data; the fine-tuning loss function includes task loss driven by expert feedback and regularization loss; The parameters of the large language model are updated based on the fine-tuning loss function using a low-rank adaptive algorithm.
