Power market boundary data adaptive verification and correction method and system

Abstract

The invention provides a power market boundary data adaptive verification and correction method and system, and belongs to the field of power equipment state monitoring. Encoding the multi-modal boundary data into a numerical sequence aggregation matrix, a text representation matrix and a topological graph representation matrix by adopting an MLLM model, and splicing the matrixes into a uniform feature sequence; carrying out physical rule verification, logic consistency verification and time sequence continuity verification by utilizing the matrix and the unified feature sequence, and generating a diagnosis report containing error types, positions and confidence coefficients; positioning historical related data based on the diagnosis report, and combining the error position and the historical related data to construct cue words to generate a plurality of groups of correction schemes; in combination with correction scheme feedback, the MLLM model is subjected to online fine tuning, and the model subjected to fine tuning is adopted to execute boundary data self-adaptive verification and correction. Therefore, the problems of lack of deep fusion of multi-source heterogeneous data, lack of adaptive ability of verification rules and lack of error correction means in current boundary data processing are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power equipment state monitoring, and in particular to a power market boundary data adaptive checking and correction method and system. BACKGROUND

[0002] Power spot market clearing is a core link of power system operation, which needs to determine the power price and generation plan of each period in the future day by solving an optimization problem under the premise of meeting the physical constraints of the power grid (such as line capacity, unit output limit) and market rules, and its accuracy highly depends on the quality of boundary data such as power grid topology, generator bidding curve and load forecast.

[0003] With the development of new power systems, the proportion of new energy is increasing and the operation mode of the power grid is more complex, so the traditional boundary data processing method is difficult to adapt: the boundary data comes from multiple systems such as dispatching and marketing, containing multi-modal data such as numerical, text and atlas, and the traditional method is difficult to find cross-modal logical conflicts, such as inconsistency between text maintenance plan and line state in topology model; the checking relies on fixed rules and thresholds, which cannot adapt to dynamic changes and new energy uncertainty, and is prone to false positives and false negatives; error correction mostly uses original means such as interpolation and filling historical average, ignoring data correlation and physical laws, introducing new bias; and the checking rules and abnormal diagnosis highly depend on expert experience, and the implicit knowledge is difficult to deposit and reuse, so the system lacks the ability to cope with new scenarios. SUMMARY

[0004] To solve the above problems, the present application provides a power market boundary data adaptive checking and correction method and system to solve the problems of lack of deep fusion of multi-source heterogeneous data, lack of adaptive ability of checking rules and lack of error correction means in current boundary data processing.

[0005] To achieve the above purpose, the present application adopts the following technical solutions: In a first aspect, the present application provides a power market boundary data adaptive checking and correction method, comprising: Using the MLLM model, the multi-modal boundary data is respectively encoded into a numerical sequence aggregation matrix, a text representation matrix and a topology graph representation matrix, and then spliced into a unified feature sequence after mapping to a unified vector space; Based on the MLLM model, the matrix and the unified feature sequence are used to detect errors through physical rule checking, logical consistency checking and time series continuity checking, and a diagnostic report containing error type, location and confidence is generated; Based on the diagnostic report, the historical related data is located, the context weight is calculated by the MLLM model, the prompt word is constructed combining the error location and the historical related data, and multiple correction schemes are generated; In combination with the feedback of the personnel on the revision scheme, the MLLM model is fine-tuned online, and the fine-tuned model is used to perform boundary data adaptive checking and correction.

[0006] In a second aspect, the present application provides a power market boundary data adaptive checking and correction system, comprising: The multi-modal data processing module is configured to use the MLLM model to encode the multi-modal boundary data into a numerical sequence aggregation matrix, a text representation matrix and a topological graph representation matrix, respectively, and then splice the matrices into a unified feature sequence after mapping to a unified vector space. The checking module is configured to use the MLLM model to perform error detection through physical rule checking, logical consistency checking and time series continuity checking based on the matrix and the unified feature sequence, and generate a diagnostic report containing error type, location and confidence. The revision scheme generation module is configured to locate historical related data based on the diagnostic report, calculate context weights by the MLLM model, construct prompt words in combination with the error location and the historical related data, and generate multiple revision schemes. The fine-tuning module is configured to fine-tune and optimize the MLLM model online in combination with the feedback of the personnel on the revision scheme, and use the fine-tuned model to perform boundary data adaptive checking and correction.

[0007] In a third aspect, the present application provides a computer program product comprising software code, wherein the programs in the software code perform the steps of the power market boundary data adaptive checking and correction method of the first aspect.

[0008] Compared with the prior art, the present application has the following advantages: 1. The present application uses the MLLM model to realize unified encoding and deep fusion of multi-modal boundary data, can capture deep correlations between different data sources, accurately find cross-modal logical conflicts, and has more comprehensive checking dimensions compared with traditional independent processing methods. Through three types of checking, a diagnostic report containing error details is generated, which can accurately locate the problem; in combination with historical data and context weights, multiple revision schemes are generated, which avoids the introduction of bias by traditional original correction methods and improves data quality. At the same time, the MLLM model is fine-tuned online relying on personnel feedback, continuously optimizing the checking and correction capabilities, solving the problem of difficult deposition and reuse of expert experience, adapting to the dynamic changes of new power systems, and ensuring the accuracy and reliability of power spot market clearing.

[0009] 2. The present application uses the inherent cross-modal attention mechanism of MLLM to realize unified encoding and deep semantic correlation analysis of numerical, text and graph data. It can find deep logical contradictions hidden between different data sources, and has more comprehensive and deeper checking dimensions.

[0010] 3. This invention employs a confidence-based smooth judgment and a self-learning bias tolerance function to replace hard thresholds. This significantly reduces false alarms, flexibly adapts to the volatility of new energy sources and changes in grid operation modes, and demonstrates a high degree of intelligence.

[0011] 4. This invention constructs the correction process as a conditional generation task, generating multiple candidate solutions that conform to physical and market rules based on rich context, and then ranking them. The correction results are not only accurate but also reasonable, leaping from "filling" to "generating," fundamentally improving data quality.

[0012] 5. This invention continuously fine-tunes the model through human-computer interaction feedback (RLHF), enabling the system to learn from expert decisions and adapt to new rules and scenarios. The system possesses self-evolution capabilities, its knowledge base continuously grows, and its ability to cope with unknown problems becomes increasingly stronger.

[0013] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0014] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0015] Figure 1 The main flowchart of an adaptive verification and correction method for electricity market boundary data provided in an embodiment of the present invention is shown. Detailed Implementation

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

[0017] Example 1 like Figure 1 As shown in the figure, this embodiment discloses an adaptive verification and correction method for electricity market boundary data, including the following steps: S1: Using the MLLM model, the multimodal boundary data are encoded into numerical sequence aggregation matrix, text representation matrix and topological graph representation matrix respectively, mapped to a unified vector space and then concatenated into a unified feature sequence; S2: Based on the MLLM model, using the matrix and unified feature sequence, error detection is performed through physical rule verification, logical consistency test and time series continuity test, and a diagnostic report containing error type, location and confidence level is generated; S3: Based on the diagnostic report, locate historical relevant data, calculate the context weight using the MLLM model, combine the error location and historical relevant data to construct prompt words, and generate multiple sets of correction schemes; S4: In combination with the feedback of the modification scheme, the MLLM model is fine-tuned online, and the fine-tuned model is used to perform boundary data adaptive checking and correction.

[0018] Next, in combination with Figure 1 , a power market boundary data adaptive checking and correction method disclosed in the embodiment is described in detail.

[0019] In S1, first, multi-modal data unified representation and alignment are performed.

[0020] The power market and the power system generate thousands of boundary data every day. These boundary data can be divided into three modalities, namely numerical sequences (unit output, market price, etc.), natural text information (such as maintenance plan, device number, etc.), and topology data (power grid topology of devices).

[0021] Boundary checking is a prerequisite for market clearing. In order to intelligently check the boundary data, the data of the three modalities need to be processed. Therefore, the data of the three modalities need to be mapped to a shared high-dimensional semantic space, thereby laying a foundation for subsequent cross-modal reasoning, checking and correction.

[0022] The data structures of the three modalities are different, and their respective features are: Numerical sequence: time dependence, continuity and physical meaning; Natural text: discrete symbol sequence, rich in grammar and semantics; Topology data: non-Euclidean data, emphasizing the connection relationship between nodes and edges.

[0023] Due to the different data structures, the multi-modal boundary data is input into the multi-modal large language model (MLLM) in the embodiment, and a specific encoder (Encoder) is designed for each modality data to convert the original data into a feature vector (Embedding). The output dimensions of all encoders are designed to be the same so as to be fused subsequently.

[0024] 1. Numerical sequence encoding Since the numerical value in the boundary data is a time series, for the i-th numerical sequence, it is represented as: (1) wherein, represents the category of the numerical sequence; represents the length of the numerical sequence; the superscript T represents vector transposition.

[0025] In order to eliminate the dimension influence of different categories of numerical sequences, the following standardization is performed: (2) In the above formula, denotes the mean of the i-th numerical sequence; denotes the standard deviation of the i-th numerical sequence.

[0026] For the normalized numerical sequence, the local patterns and long-term dependencies are captured using the time series Transformer, and the response hidden layer output sequence can be represented as: (3) In the above formula, denotes the time series Transformer as an encoder; denotes the i-th normalized numerical sequence; denotes the real number field; denotes the output length; denotes the dimension of the Transformer feedforward network.

[0027] For the output obtained above, global attention average pooling is performed from the time dimension to obtain the following single sequence representation: (4) In the above formula, denotes the attention pooling process; denotes the single sequence length after global average pooling.

[0028] Then, for all classes of numerical sequences in the boundary data, the above process is performed to finally obtain the following numerical sequence aggregation matrix: (5) where N denotes the total number of classes of numerical sequence class data in the boundary data.

[0029] 2. Natural text encoding For natural text in the boundary data, such as: “Line L12 is scheduled for maintenance outage from 10:00 to 18:00 on 2024-05-27”.

[0030] First, the sentence needs to be converted into a Token sequence using the tokenizer of the pre-trained language model BERT, and the initial embedding of each Token is obtained. Taking the above sentence as an example, its initial embedding can be represented as: (6) where, and denote the start and end markers of the sequence, respectively, which are “:” and “.” in the example sentence; the remaining elements denote the Token processed by the tokenizer; denotes the sequence length; denotes the word embedding dimension.

[0031] The initial embedding is input into the Transformer layer to capture the contextual semantics of the words, resulting in deep features for each word token, and the corresponding text output is: (7) where, represents the core processing procedure in the Transformer, including self-attention layers, feed-forward neural networks, residual connections, and layer normalization.

[0032] For sentence representation, the vector corresponding to the [CLS] token at the beginning of the sequence is taken as the summary representation of the entire sentence, so: (8) Similarly, the above process is performed on all natural texts, and the final representation of all text data can be obtained as: (9) where M represents the total number of natural texts.

[0033] 3. Topological data encoding The power grid topology can be represented as: (10) where, represents the set of nodes, and the total number of nodes is ; represents the set of edges, and the total number of edges is ; represents the features of nodes and edges (such as bus reference voltage, line impedance, etc.).

[0034] For topological data, a message passing neural network (MPNN) is used for aggregation encoding.

[0035] First, for any node , its electrical value is read as the initial feature vector . In each iteration of the MPNN, the node updates its representation by aggregating the information of its neighbors. The message function is used to define the transmission of messages from neighbor nodes to the target node, and the specific function design is: (11) where u represents the neighbor node; v represents the target node; k represents the iteration round; represents the feature of the edge (such as line impedance or capacity, etc.); The function is given by a neural network RNN, which can be represented as: (12) In the above formula, denote the hidden layer weight network coefficients of the RNN; denote the concatenation vector of and denote the drift of the RNN hidden layer.

[0036] For node v, the messages of all neighbor nodes are aggregated by using the attention mechanism, so that (13) where denotes the attention process; denotes the set of all nodes connected to node v.

[0037] Combined with the state of the last round and the aggregated message, the node state is updated, and the feature vector of node v is updated as (14) where, denotes the update function, and here the GRU (Gated Recurrent Unit) is used as the update function.

[0038] Suppose that after K iterations, the algorithm converges or reaches the maximum step, at this time, the final representation of all nodes can be aggregated into a vector representing the entire graph by a readout function, which is the final representation of the topology data, and is specifically (15) where, denotes the readout function; denotes the final representation of the topology data.

[0039] After the above steps, the data of the three modalities are mapped to the same dimensional vector space. In order to realize cross-modal understanding, the model needs to learn to align these representations in the pre-training stage, for example, the text vector describing "line L12 outage" should be as close as possible to the graph vector when line L12 is disconnected in the topology graph.

[0040] Specifically, the alignment of multi-modal data is realized by contrastive learning, the principle of which is to encourage the agent to pull the distance of positive samples (i.e. correct multi-modal representation pairs) in the vector space and push the distance of negative samples (i.e. multi-modal representation pairs that cannot be matched) as far as possible. For this purpose, the following loss function is designed: (16) In the above formula, denotes a similarity function, such as cosine similarity; denotes a positive sample pair; denotes all​​​ Mismatched negative sample pairs Indicates the super-temperature parameter.

[0041] Through the above steps, the agent can obtain the ability to understand the deep semantic association of power data, providing the possibility for subsequent intelligent verification and correction.

[0042] Further, after obtaining the aligned multi-modal data vector, splicing is performed, and position encoding is added, then the initial input feature sequence can be represented as: (17) The initial input feature sequence is input to Iterate to obtain the final feature sequence that integrates all modal information: (18) Wherein, Indicates the iteration step, Indicates the maximum step, then the output of the maximum step Each vector in the

[0043] In S2, adaptive state estimation and consistency verification are performed, which specifically includes the following three types of tasks: 1. Physical rule verification It should be understood that the base power balance refers to the accurate matching between the total generator output and the total load demand and system line loss at any time and any node of the power system, i.e. "total power generation = total power consumption + system line loss", which is the core physical criterion for stable operation of the power system.

[0044] The power clearing of the power spot market needs to be based on this balance relationship to determine the electricity price and generation plan. If there is power imbalance in the boundary data (such as the total generator output and load prediction do not match), it will directly lead to the deviation of the clearing result from the actual carrying capacity of the power grid - light causes abnormal fluctuations in electricity prices, heavy breaks physical constraints such as line capacity and unit output limit, causing frequency deviation, line overload and even power grid collapse.

[0045] Because the base power balance is the bottom line criterion for safe operation of the power system, and small deviations in core data such as generator bids, load forecasts, and new energy output in boundary data may break this balance, it must be the primary task of physical rule verification to ensure that the subsequent clearing calculation is logically correct.

[0046] Specifically, to ensure that the boundary data meets the base power balance and achieve adaptive verification, a smooth probability function can be trained by MLM (multi-modal large language model) to estimate the likelihood that the current data meets the power balance. The smooth probability function is specifically: (19) In the above formula, represents the balance deviation confidence; represents the CDF (Cumulative Distribution Function) of the standard normal distribution; represents the power imbalance deviation, represents the set of all generators, represents the output of the gth unit, represents the set of loads, represents the load power of the lth load, represents the line loss of the entire system, and the generator output, load power and system line loss are all boundary data; represents an adaptive deviation tolerance function, where represents the time label of the sequence, represents global context semantics in is obtained by learning according to the MLLM.

[0047] According to only one proximity threshold value, such as 0.05, can be set, and when , the MLLM gives a power imbalance alarm, reporting “serious power imbalance detected, confidence only ”.

[0048] In this embodiment, the physical rule verification takes the base power balance as the core, and calculates the balance deviation confidence by means of the MLLM learning smooth probability function, replacing the traditional fixed threshold judgment. It can dynamically adapt to the changes of power grid operation mode and the volatility of new energy, accurately identify the imbalance problems of total power generation, total power consumption and line loss, greatly reduce false positives and false negatives, avoid the deviation of clearing results caused by power imbalance, and ensure the stability of power grid frequency and the safety of line.

[0049] 2. Logical consistency verification This verification ensures that the information described in different modalities is logically consistent. This is a key embodiment of system intelligence. The core task is to judge whether the text description and the chart features are consistent.

[0050] To achieve the above task, it is constructed as a matching task based on attention. The model uses the text representation as the query (Query) to query the relevant keys (Key) and values (Value) in the topological graph representation: (20) where, , , respectively, the weight parameter of the query Q, the key K, the value V The corresponding cross-attention mechanism is: (21) In the above formula, is an activation function; T represents vector transposition; denotes the scaling factor between the key and the value.

[0051] Therefore, the consistency score of the text with the graph feature is: (22) wherein, The cosine similarity function is still used. The higher the score , the more consistent the text description is with the topology state. For ease of understanding, an example is given below: Suppose the input text is "line L12 outage", according to the above steps, the model calculates the attention of the text and the topology graph, and the result is: (1) If the attention is concentrated on the node / edge corresponding to the line L12, and the output vector has a high similarity with the text vector (the score is 0.9), it is judged to be consistent; (2) If the graph shows that L12 is in the state of "running", the output vector will have a great difference with the text vector, resulting in = 0.07 (very low), thereby triggering the logic inconsistency alarm.

[0052] In this embodiment, the logic consistency check associates the text and the topology representation through the cross-attention mechanism, calculates the similarity between the two to determine whether they match. Compared with the traditional way of processing multiple modal data independently, it can accurately capture the deep conflict between the text description and the topology state, clearly mark the conflict position and basis, reduce the cost of manual screening, ensure the logical self-consistency of the multi-modal boundary data, and avoid affecting the rationality and accuracy of the clearing decision due to information conflicts.

[0053] 3. Time series continuity check This check is used to detect abnormal jumps or burrs in numerical data (such as power prediction). Traditional methods are generally based on threshold values (such as first-order difference threshold) or traditional time series anomaly detection (such as ARIMA), and this embodiment uses the powerful context learning ability of MLM to learn the normal mode of the time series and predict the next point, and judges the anomaly through the prediction error.

[0054] For the numerical sequence , a part of it can be used as context to let the MLM predict the next value and calculate the probability of prediction error, and the corresponding model is: (23) In the above formula, represents the conditional probability; represents the numerical sequence tth element in the sequence; represents the global context semantics consistent with the meaning in formula (19); is the output vector of the corresponding time t position in represents the weight parameter of the MLLM.

[0055] By the abnormal negative log-likelihood of the above formula, the following index can be constructed: (24) is an index representing potential anomalies, and according to the data validation set, the MLLM can automatically generate a threshold, denoted as , when then is marked as a potential anomaly.

[0056] In this embodiment, the time series continuity test is based on the MLLM learning the normal mode of the numerical sequence, and detects jumps or burrs by predicting the numerical probability and calculating the anomaly index. Compared with traditional threshold or ARIMA method, it can more accurately identify the anomaly of power prediction and other numerical data, combined with the judgment of global context semantics, reduce the dependence on artificial experience, effectively avoid the introduction of outlier bias by abnormal numerical values, and ensure the continuity and reliability of boundary data in time dimension, and improve the accuracy of the result.

[0057] Through the above three types of state estimation and verification, the MLLM outputs a tuple as follows for each detected problem: (25) wherein, represents the error type, including physical conflict, logical inconsistency, time sequence anomaly, and data missing. represents the error position, which is used to accurately locate the error data, such as "Device ID: Line_L12, Time Point: 2024-05-27T14:00"; represents the confidence, i.e. the degree of certainty of the model for this diagnosis; represents the supporting evidence, i.e. the explainability basis for the model to make this judgment. It is usually generated through attention weight, for example, "The judgment that 'Line L12 outage' is inconsistent with the topology state is because in the cross-attention calculation, the model's attention weight for the 'L12' node in the topology graph is as high as 0.95, and the state of the node is 'running'." In this embodiment, physical rule verification refers to the operating rules and constraints of the load power grid. Only when the physical rules are met can the data generated by MLLM be secure and feasible. Logical consistency verification ensures that the data and text generated by MLLM logically conform to causality, especially the electrical laws followed by the power system. Time series continuity verification aims to eliminate possible abnormal and abrupt data in the data sequence, ensuring a smoother and more robust data sequence. Only when all three aspects are met can the multimodal data and text generated by MLLM be ensured to be credible, authentic, and logically consistent. This provides accurate and compliant boundary data support for the clearing of the electricity spot market, preventing the clearing results from deviating from the physical constraints of the power grid or market rules due to data errors, and ensuring the stable operation of the power system and the fairness and efficiency of electricity market transactions.

[0058] In S3, when MLLM diagnoses an anomaly, it performs intelligent data correction based on this and provides corresponding solutions to staff. The relevant problem can be modeled as a conditional probability generation model as follows: (26) In the above formula, This represents the candidate correction data to be determined; This indicates that the corrected data will be used as a statistical variable; Represents the conditional probability of the corrected data statistics; This represents the observed original boundary dataset containing erroneous data; This indicates the context information accepted by MLLM; This represents the constraints in the boundary data, such as power balance conditions and power flow balance conditions. This represents the set of parameters for a given MLLM, namely the neural network weight parameters, learning rate parameters, etc., formed after training with power text.

[0059] The model first uses the diagnostic report tuples The position in the middle and Lock the associated context The specific handling method is as follows: 1. For numerical data errors (such as abnormal bus load). include: (1) Spatiotemporal correlation data: the load values ​​of the bus on / upstream adjacent buses (spatial dimension), and the load data of the bus at other time points (time dimension). (2) Historical data: Load data of the same time and the same week over the past 7 days; (3) Relevant text: Are there any planned activities in the region (extracted from text data); (4) Weather data: (if access) temperature, humidity, etc. information of the area.

[0060] 2. For topology state conflict (such as line state inconsistency), Including: (1) Topology neighbor: the state of the bus connected to the line and other lines.

[0061] (2) Protection logic: related protection device action information (extracted from text alarm).

[0062] (3) Maintenance plan: whether there is a corresponding "maintenance plan" in the text mode.

[0063] Lock the context by the above method After that, a special attention layer can be added in MLLM, or the verification module can be inherited, and the most relevant context information is automatically weighted and selected through the following mechanism: (27) In the formula, represents the most relevant context information data representation given by the attention mechanism; represents the corresponding weight of the i-th context information data representation in the attention mechanism; is the information data representation of the corresponding error position in

[0064] The , error position identifier and problem description obtained by solving formula (27) are collectively constructed as a prompt (Prompt) and input into MLLM. Using strategies such as Top-p Sampling or Beam Search, multiple candidate values are generated from the distribution learned by the model, denoted as .

[0065] For the candidate value set as above, the following process is used to verify and sort the candidate values: Step 1: Substitute the candidate value into the corresponding physical and rule constraints, such as the safe operation constraint conditions in the power system: over-limit, line transmission power cannot exceed capacity; power balance condition, output equals load, etc. Check if it meets the requirements. For example, whether the generated unit output value meets the upper and lower limit constraints, whether the generated topology state will cause an island, etc. Step 2: Calculate the degree of agreement between any candidate value and the context , which is specifically: (28) In the formula, Identifying candidate values corresponding attention representation; Sim is a cosine similarity function.

[0066] Step 3: Calculate the comprehensive score of each candidate value according to the following formula: (29) where is a hyperparameter used to balance the generation probability and the goodness of fit.

[0067] Step 4: Sort all candidate values according to the score size, and attach the generation reason and generation probability of the candidate value, and finally obtain a sorted correction data candidate table.

[0068] Where the generation reason is automatically generated by MLLM according to the attention weight in the generation process (such as: "This value is linearly extrapolated according to the load trend of adjacent buses B1 and B2, and meets the total power balance constraint"), and the candidate value can be represented as a triple, specifically: (30) where, represents the generation probability of MLLM in the generation process; represents the generation reason of .

[0069] In this embodiment, intelligent correction breaks the limitations of traditional single interpolation and mean filling, and generates multiple combination correction schemes with the help of MLLM. Thus, according to the global context semantics, the correction strategy is customized for abnormal data of power and topology, ensuring data continuity while accurately restoring the real operation state of the power system, avoiding the deviation easily introduced by the original means, providing more accurate and practical boundary data for the power spot market clearing, and improving the reliability of the clearing result.

[0070] In S4, after adaptive checking, diagnosis and generation of correction candidate data set, MLLM presents the diagnosis report and sorted correction candidate set to the operator, and the following is an example of the report format: Diagnosis: Load value of

bus X

time T

[0071] Evidence: The value is mutated by +200% compared with the previous period, far exceeding the historical same period fluctuation range (±10%).

[0072] Recommended correction scheme: Scheme 1 [recommended degree: 0.92]: value = 150.0 MW. Reason: based on ARIMA model and spatially correlated load smoothing prediction. ​

[0073] Plan 2 [Recommended: 0.85]: Value = 148.5 MW. Reason: Replaced with historical same-period mean.

[0074] Plan 3 [Recommended: 0.78]: Value = 152.1 MW. Reason: Forecast adjusted for weather factors.

[0075] After the staff obtains the report, they can choose to adopt a certain candidate plan, or mark it as a false positive or reject all plans and manually input the correction.

[0076] The staff's decision will be used as valuable feedback signals for the continuous optimization of the model, and the specific mechanisms include: 1. Reinforcement learning signals: If the staff adopts a plan, the trajectory of that plan will receive a positive reward (Positive Reward); conversely, if all plans are rejected, these trajectories will receive a negative reward (Negative Reward).

[0077] 2. Model fine-tuning: Regularly use these collected (state-action-reward) triple data to fine-tune the model online (Online Fine-Tuning). This is usually achieved through reinforcement learning human feedback (RLHF) or online supervised learning.

[0078] To enable the MLM to implement the above mechanisms, first use the system to continuously collect feedback during human-machine collaboration to build a feedback dataset , which contains the following core elements: 1. State : All input information used by the model when making a diagnosis and generating a correction candidate, including the original data with problems, the context information extracted by the model , the internal representation of the model (which can be the hidden state vector at the corresponding position); 2. Action : A specific candidate correction value generated by the model, such as ; 3. Reward : Set according to the staff's feedback (1) +Positive Reward (Set reward +R1): If the running staff adopts the candidate plan. This is the strongest positive signal; (2) +Small Positive Reward (Set reward +R2): If the running staff does not adopt it, but the manually input value is very close to a certain candidate value (within the Euclidean distance or relative error range). This indicates that the model's direction is correct, but the accuracy is slightly lower.

[0079] (3) -Negative Reward (set reward -R1): If the operator explicitly rejects all the solutions and inputs a completely different value.

[0080] (4) -Strong Negative Reward (set reward -R2): If the operator labels this diagnosis as a "false positive". This means that the model not only acted wrongly, but also the initial "state" judgment was wrong.

[0081] After the construction is completed , in order to realize the smooth updating of the parameters of the MLLM, the Proximal Policy Optimization (PPO) is adopted to train the agent as follows: (31) In the above formula, denotes the parameter set to be optimized of the MLLM; denotes the overall loss function of the Proximal Policy Optimization; and are hyperparameters; denotes the policy loss function, and its specific form is: (32) In the above formula, denotes the ratio of the probability of generating an action by a new policy to the probability of an old policy; denotes the advantage function at the t-th step, which measures how much better it is to take action in state compared to the average case, and is specifically: (33) wherein, is the state value function, which is generated by the state network of reinforcement learning (such as Q-Learning).

[0082] In formula (32), is a hyperparameter, and a typical value is 0.2, which is used to limit the fluctuation range of is a clipping function, which can limit between and to prevent the policy from fluctuating sharply.

[0083] In formula (31), denotes the value loss function, and its form can be designed as: (34) wherein, is the norm of the square difference.

[0084] Finally, the entropy reward in formula (31) is represented as , which is specifically expressed as: (35) , wherein, represents an information entropy function; represents the policy of the agent under the parameter .

[0085] Through the above steps, the online updating of the parameters of the MLLM is realized.

[0086] In this embodiment, on the one hand, the errors of the MLLM are quickly corrected through personnel feedback, and the implicit experience of experts is converted into model knowledge; on the other hand, the model learns the real-time data and the changes in the scene online, and continuously optimizes the checking and correction ability. Thus, the workload of manual checking is reduced, and the system is adapted to the dynamically changing operation scene in the development of the new power system, the ability is iteratively upgraded, and the timeliness and accuracy of the boundary data processing are ensured.

[0087] The specific embodiment is aimed at the problems of difficult fusion of multi-source heterogeneous data, rigid checking rules, primitive correction means, and difficult reuse of expert experience in the existing boundary data processing of the power market. Through the MLLM model, the numerical, textual, and topological multi-modal data are uniformly encoded and semantically aligned, the limitations of traditional independent processing are broken through, and the cross-modal deep logical conflicts can be found. The dynamic confidence judgment is used to replace the fixed threshold, the adaptive ability of checking is improved, and the false positives and false negatives are reduced. Based on the rich context, multiple combination correction schemes are generated, the primitive means such as interpolation and mean filling are abandoned, and the data filling is upgraded to intelligent generation. The MLLM is fine-tuned online combined with personnel feedback, the implicit experience of experts is converted into reusable model knowledge, the system has the ability of continuous evolution, provides high-quality data support for the power spot market clearing, and adapts to the development needs of the new power system.

[0088] Embodiment Two The embodiment provides an adaptive checking and correction system for boundary data of a power market, comprising: a multi-modal data processing module, configured to use an MLLM model to encode multi-modal boundary data into a numerical sequence aggregation matrix, a text representation matrix, and a topological graph representation matrix, respectively, and then splice the matrices into a uniform feature sequence after mapping to a uniform vector space; a checking module, configured to use the matrices and the uniform feature sequence to detect errors through physical rule checking, logical consistency checking, and time series continuity checking based on the MLLM model, and generate a diagnosis report containing error types, positions, and confidence levels; a correction scheme generation module, configured to locate historical related data based on the diagnosis report, calculate context weights by using the MLLM model, construct prompt words combined with the error positions and the historical related data, and generate multiple correction schemes. Fine-tuning module, for fine-tuning the MLLM model online in combination with the feedback of the personnel on the revision scheme, and performing boundary data adaptive checking and correction using the fine-tuned model.

[0089] Embodiment three A computer program product comprising software code characterized in that programs in said software code perform the steps of a method for boundary data adaptive checking and correction in a power market as described in embodiment one above.

[0090] The steps or modules involved in embodiments two to three above correspond to those in embodiment one, and the specific implementation can be seen from the relevant description in embodiment one.

[0091] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. An adaptive verification and correction method for electricity market boundary data, characterized in that, include: The MLLM model is used to encode multimodal boundary data into numerical sequence aggregation matrix, text representation matrix and topological graph representation matrix respectively, and then concatenate them into a unified feature sequence after mapping to a unified vector space. Based on the MLLM model, using the matrix and unified feature sequence, error detection is performed through physical rule verification, logical consistency test and time series continuity test, and a diagnostic report containing error type, location and confidence level is generated. Based on the diagnosis report, historical relevant data is located, and the context weight is calculated by the MLLM model. The prompt words are constructed by combining the error location and historical relevant data, and multiple correction schemes are generated. Based on feedback from personnel regarding the revised scheme, the MLLM model was fine-tuned and optimized online, and the fine-tuned model was used to perform adaptive verification and correction of boundary data.

2. The adaptive verification and correction method for electricity market boundary data as described in claim 1, characterized in that, The process of encoding multimodal boundary data into a numerical sequence aggregation matrix, a text representation matrix, and a topological graph representation matrix specifically includes: Numerical sequence data, natural text, and power grid topology data from the boundary data are respectively input into the encoder; The numerical sequence data, after being normalized, is used to capture temporal features through a Transformer layer and an attention mechanism, and the temporal features of all categories of numerical sequences are aggregated to obtain a numerical sequence aggregation matrix. The natural text is converted into a word sequence using the BERT word segmenter. After capturing the contextual semantics through the Transformer layer, the starting vector of the word sequence is taken as the sentence representation. The sentence representations of all categories of natural text are aggregated to obtain the text representation matrix. A message-passing neural network is used to encode the power grid topology data, and the aggregated encoding is used to obtain the topology graph representation matrix.

3. The adaptive verification and correction method for electricity market boundary data as described in claim 1, characterized in that, The physical rule verification specifically includes: ； in, Indicates the confidence level of the balance deviation; The cumulative distribution function representing the standard normal distribution; Indicates power imbalance deviation. Represents the set of all generators. This represents the output of the g-th unit. Represents a set of loads. This represents the load power of the l-th load. This represents the overall line loss of the system; generator output, load power, and system line loss are all boundary data. Represents the adaptive tolerance function, where Time stamps representing sequences Represents the global context semantics in a unified feature sequence. Global context semantics of MLLM combined with boundary data Dynamic learning generation; when the confidence level of the balance deviation does not exceed a preset threshold, a power imbalance alarm is triggered.

4. The adaptive verification and correction method for electricity market boundary data as described in claim 1, characterized in that, The logical consistency check specifically includes: Using the text representation matrix as the query and the topology graph representation matrix as the matching key and value, the model focuses on the topology nodes or edges associated with the text description through a cross-attention mechanism. Calculate the similarity between the text representation and the attention matching result. If the similarity is lower than the threshold, it is determined that the two are logically conflicting, triggering an inconsistency alarm and marking the conflicting text and topological location.

5. The adaptive verification and correction method for electricity market boundary data as described in claim 1, characterized in that, The time series continuity test specifically includes: Based on the MLLM model, we learn the normal fluctuation pattern of numerical series and predict the value and corresponding probability of the next period using some historical values ​​as context. Calculate the probability of deviation between the actual value and the predicted result, and convert the probability into an anomaly index. If the index exceeds the threshold of the model's self-learning, it is determined that there is an abnormal jump or spike in the value, which is marked as a potential anomaly and the abnormal period is recorded.

6. The adaptive verification and correction method for electricity market boundary data as described in claim 1, characterized in that, The process involves locating historical relevant data based on diagnostic reports, calculating context weights using an MLLM model, constructing prompt words by combining the error location and historical relevant data, and generating multiple sets of correction schemes, specifically including: Based on the error type and location in the diagnostic report, locate the relevant historical data; if the error is a numerical sequence data error, extract the historical values ​​of the same time and week type for the same preset number of days, as well as the concurrent data of adjacent equipment in the same area; if the error is a topology or text data error, extract the historical maintenance records of the equipment and the status data of the associated equipment. The similarity between historical relevant data and error location data is calculated using an MLLM model. Contextual weights are assigned based on the similarity, and the key contextual information with the highest weight is selected. Integrate key contextual information, error location identifiers, and error descriptions to construct structured prompt words, which are then input into the MLLM model; By employing kernel sampling or beam search strategies, the MLLM model samples multiple times from the learned power data distribution to generate multiple sets of initial correction schemes; Each initial modified scheme is substituted into the physical and rule constraints for verification, and schemes that do not meet the constraints are eliminated. The degree of fit between the remaining solutions and the key context is calculated. Combined with the probability when the MLLM generates the solutions, the solutions are comprehensively scored and ranked to obtain the multiple sets of modified solutions.

7. The adaptive verification and correction method for electricity market boundary data as described in claim 1, characterized in that, The process of combining feedback from personnel on the revised scheme and fine-tuning the MLLM model online includes: Collect feedback from personnel on the corrective solutions and construct a feedback dataset; the dataset includes the status of the model during diagnosis and solution generation, the corrective solutions generated by the model, and the rewards set according to the feedback; Based on the feedback dataset, the MLLM model is fine-tuned online using a near-end policy optimization algorithm: An overall loss function is constructed using policy loss, value loss, and entropy reward to limit the probability fluctuation range of the new and old policies, thereby avoiding drastic changes in model parameters.

8. An adaptive verification and correction system for electricity market boundary data, characterized in that, include: The multimodal data processing module is used to encode multimodal boundary data into numerical sequence aggregation matrix, text representation matrix and topological graph representation matrix respectively using MLLM model, and then concatenate them into a unified feature sequence after mapping to a unified vector space. The verification module is used to perform error detection based on the MLLM model, using the matrix and unified feature sequence, through physical rule verification, logical consistency test and time series continuity test, and generate a diagnostic report containing error type, location and confidence level. The correction scheme generation module is used to locate historical relevant data based on the diagnostic report, calculate the context weight by the MLLM model, construct prompt words by combining the error location and historical relevant data, and generate multiple sets of correction schemes. The fine-tuning module is used to fine-tune and optimize the MLLM model online by incorporating feedback from personnel on the correction scheme, and then use the fine-tuned model to perform adaptive verification and correction of boundary data.

9. The adaptive verification and correction system for electricity market boundary data as described in claim 8, characterized in that, The process of encoding multimodal boundary data into a numerical sequence aggregation matrix, a text representation matrix, and a topological graph representation matrix specifically includes: Numerical sequence data, natural text, and power grid topology data from the boundary data are respectively input into the encoder; The numerical sequence data, after being normalized, is used to capture temporal features through a Transformer layer and an attention mechanism, and the temporal features of all categories of numerical sequences are aggregated to obtain a numerical sequence aggregation matrix. The natural text is converted into a word sequence using the BERT word segmenter. After capturing the contextual semantics through the Transformer layer, the starting vector of the word sequence is taken as the sentence representation. The sentence representations of all categories of natural text are aggregated to obtain the text representation matrix. A message-passing neural network is used to encode the power grid topology data, and the aggregated encoding is used to obtain the topology graph representation matrix.

10. A computer program product, comprising software code, characterized in that, The program in the software code executes the steps in the adaptive verification and correction method for electricity market boundary data as described in any one of claims 1-7.

Images