Multi-source data fusion-based power spot transaction intelligent quotation auxiliary decision-making method

Abstract

The application discloses a kind of power spot transaction intelligent quotation auxiliary decision-making methods of multi-source data fusion, comprising: physical-data fusion perception, through the system situation information of differentiable power flow calculation and fusion encoder output;Endogenous security and compliance optimization, build the Stackelberg game framework of quotation agent and virtual mechanism designer to generate incentive compatible strategy;Dynamic opponent identification, based on market clearing residual error extraction strategy features and online identification game state;Risk quantification, using survival analysis to evaluate rule mutation risk and adjust value function;Man-machine intention reconciliation, compile expert instruction as constraint and generate feasible strategy through conflict resolution;Closed loop evolution, update model according to market feedback.The application realizes the deep penetration of perception, decision-making, risk management and man-machine cooperation, and improves the physical feasibility, compliance, adaptability and robustness of quotation decision-making.

Description

Technical Field

[0001] This invention relates to the field of power, and specifically to a method for intelligent pricing and decision support in power spot trading that integrates multi-source data. Background Technology

[0002] Currently, most electricity spot market pricing auxiliary decision-making technologies are built around a single aspect or adopt a shallow coupling model, which mainly suffers from the following technical limitations:

[0003] At the information perception level, there are risks of "pseudo-fusion" and physical infeasibility: Traditional methods typically run data-driven prediction models (such as load and renewable energy output prediction based on neural networks) and physical simulation models (such as power flow calculations) independently or sequentially. After the data model outputs predicted values, these are then input into the physical model to verify constraints; if these constraints are not met, external adjustments are required. This process severs the intrinsic connection between data and physical laws, potentially leading to situations where the situational information upon which decisions are based is statistically reasonable but physically unrealizable—a "pseudo-fusion" problem that creates security risks for subsequent optimization.

[0004] Strategy optimization and regulatory compliance are in a passive-aggressive relationship: existing intelligent pricing models mostly take profit maximization as the sole objective, roughly considering rule constraints by setting simple boundaries or penalties. This essentially treats regulation as an external, static obstacle to be passively circumvented, failing to internalize regulatory logic. The intelligent agent cannot proactively understand and adapt to the dynamic intent of regulatory rules, easily getting caught in a "cat-and-mouse game," lacking robustness and long-term stability in complex and changing regulatory environments.

[0005] Modeling opponent behavior relies on strong assumptions and is difficult to update dynamically: Traditional game theory methods often assume that opponents are perfectly rational or adopt fixed behavioral patterns, or attempt to estimate each opponent's private cost function, which is difficult to achieve in practice due to high information asymmetry. The models cannot effectively capture the dynamic evolution of collective game styles in the market, resulting in insufficient adaptability of strategies when facing real and changing market behavior.

[0006] Risk quantification often focuses on routine fluctuations, lacking forward-looking assessments of extreme structural risks: Existing risk models primarily measure and hedge against routine uncertainties such as price volatility and load forecasting errors, but lack effective probabilistic assessment tools for extreme "black swan" events such as sudden policy or rule changes. Decision-making systems often react slowly or are unprepared for such high-risk events.

[0007] Human-machine collaboration remains at a superficial level of interaction, making it difficult to accurately inject expert experience: existing systems' human-machine interfaces mostly use parameter adjustments or natural language commands, which are ambiguous and difficult to directly translate into mathematical constraints that machines can strictly execute. When expert experience conflicts with system models or hard constraints, the lack of automated and interpretable conflict resolution mechanisms severely restricts the effective integration of human wisdom and the improvement of decision-making efficiency.

[0008] The system lacks closed-loop evolution capability: most decision models are statically designed, and the parameters are fixed after training. They cannot learn online and iterate on their own according to the continuous changes in the market environment (i.e., "concept drift"), which leads to the degradation of model performance over time.

[0009] Therefore, there is an urgent need for a next-generation intelligent pricing and decision-making support method for electricity spot trading that can deeply integrate multi-source information, internalize compliance logic, dynamically perceive game theory, quantify extreme risks, achieve deep human-machine collaboration, and possess continuous evolution capabilities, in order to cope with the increasingly complex challenges of the electricity market environment. Summary of the Invention

[0010] To address the aforementioned technical problems, this invention provides a method for intelligent pricing and decision support in electricity spot trading based on multi-source data fusion.

[0011] This invention is achieved through the following technical solution:

[0012] The present invention provides a multi-source data fusion method for intelligent pricing and decision support in electricity spot trading, comprising the following steps:

[0013] Step S1, physical-data fusion sensing, constructing a differentiable power flow calculation layer based on the implicit function theorem, and jointly training a fusion encoder; the fusion encoder takes multi-source heterogeneous data as input, and the physical quantity obtained by the output of the fusion encoder after being calculated by the differentiable power flow calculation layer, the difference between the physical quantity and the measured physical quantity constitutes the physical consistency loss, which is used to drive the training of the fusion encoder, thereby outputting physically consistent system situation information;

[0014] Step S2: Intrinsic security and compliance strategy optimization. Construct a parameterized virtual mechanism designer that simulates the regulatory response of market operators, and form a Stackelberg game framework together with the bidding agent. By solving this two-layer optimization problem, train the bidding agent to output a bidding strategy that balances expected benefits and expected regulatory penalties.

[0015] Step S3: Dynamic opponent behavior identification; calculate the residual sequence between the actual market clearing result and the benchmark prediction result based on public cost information; extract features from the residual sequence to obtain a strategy feature vector representing the market game style; based on the strategy feature vector sequence, use a probabilistic graphical model to update the estimate of the current market game state online.

[0016] Step S4: Systemic risk quantification. Based on historical rule mutation events and their associated covariates, a survival analysis model is constructed to assess the risk of rule mutation. According to the future rule stability probability output by the survival analysis model, a risk adjustment factor is calculated, and the discount factor in the multi-period decision value function is adjusted using the risk adjustment factor.

[0017] Step S5, human-machine intent reconciliation, compiles the policy instructions defined by human experts in a formal language into a set of mathematical constraints that can be embedded in the optimization model; verifies the satisfiability of the set of mathematical constraints and the inherent physical and rule constraints, and solves the minimization correction problem when conflicts occur to generate a feasible and semantically close final strategy.

[0018] Step S6, closed-loop evolution: Based on the difference between the execution results of steps S1 to S5 and the subsequent actual market results, a training signal is generated, and at least one of the following is updated: the fusion encoder in step S1, the pricing agent and the virtual mechanism designer in step S2, the probabilistic graphical model in step S3, and the survival analysis model in step S4.

[0019] This method constructs a complete, closed-loop intelligent decision-making system encompassing data perception, strategy optimization, adversary modeling, risk quantification, and human-machine collaboration. Its core innovation lies in overturning the traditional model of treating these processes in isolation or with shallow coupling. Through a series of meticulously designed interdisciplinary methods, it achieves deep integration and collaborative evolution among the various stages.

[0020] Specifically, the physical-data fusion perception layer (S1) lays a solid information foundation for the entire system that conforms to objective physical laws, ensuring that all subsequent decision-making and reasoning are carried out within a physically feasible space.

[0021] Building on this foundation, the endogenous security and compliance strategy optimization layer (S2) draws on the mechanism design concept in economics to internalize external, passive rule compliance into proactive behavioral logic that aligns with the agent's own interests. This is the key to enabling intelligent systems to operate steadily and reliably in complex regulatory environments over the long term.

[0022] The dynamic adversary behavior identification layer (S3) acts like a keen "market radar." It bypasses the traditional problem of directly estimating the elusive cost function of individual adversaries. Instead, it extracts the systematic deviation (i.e., residual) from the macro-output of collective market behavior—the clearing results—from the benchmark cost competition pattern and patterns and characteristics of this deviation. This allows it to dynamically capture the evolution of the overall market game style in a non-intrusive manner.

[0023] The Systemic Risk Quantification Layer (S4) further expands the scope from routine market fluctuations to extreme, structural risks, particularly policy and rule mutation risks. It creatively introduces survival analysis, a tool commonly used in medicine and reliability engineering, to partially transform seemingly unpredictable "black swan" events (rule mutations) into "gray rhino" events that can be probabilistically assessed using historical data and current multi-dimensional indicators, thus enabling the decision-making system to possess forward-looking risk avoidance capabilities.

[0024] The Human-Machine Intent Reconciliation Layer (S5) solves the "last mile" problem between artificial intelligence and human experts. Through formalized specification language and automated conflict resolution mechanism, it transforms unstructured, experience-rich intuitive judgments of humans into machine-understandable and executable instructions in a precise, unambiguous manner, without compromising the system's security bottom line. This achieves true deep human-machine collaboration, rather than simple human-machine interaction.

[0025] Ultimately, the closed-loop evolution layer (S6) endows the entire system with the vitality of self-iteration and self-improvement, enabling it to continuously evolve with the dynamic changes in the market environment and overcome the inherent defects of static models that suffer from performance degradation due to market drift.

[0026] These six steps are progressive and interconnected, together forming a next-generation intelligent decision-making brain capable of adapting to the highly dynamic, highly uncertain, and strictly regulated electricity spot market environment.

[0027] Further, step S1 includes the following steps:

[0028] S1.1 Construct a differentiable power flow calculation layer, encapsulating the traditional power flow numerical solver into a differentiable calculation module. Its differentiability is achieved in the following way: In the forward propagation stage, the traditional numerical solver is called to obtain a high-precision solution of the system state variables; in the backward propagation stage, the iterative process inside the solver is not traced back, but the gradient is calculated by solving the corresponding linear equations based on the linearized model of the system at the solution obtained from the forward propagation, thereby realizing the differentiable embedding of physical laws.

[0029] S1.2, Construct a physical consistency-driven fusion encoder. Design a deep neural network as the fusion encoder. Its architecture is configured according to the input data type and is used to process spatiotemporal sequence and topology data. The output of the fusion encoder is the predicted value of the injected power and topology parameters of the power grid nodes. This predicted value is directly fed to the differentiable power flow calculation layer as its input.

[0030] S1.3, implement end-to-end joint training, using the previously constructed differentiable power flow calculation layer as a physical constraint bridge to calculate the physical consistency loss; the physical consistency loss is the difference norm between the predicted physical quantity output by the differentiable power flow calculation layer and the measured physical quantity from the measurement system; the total loss function of the fusion encoder is composed of the weighted sum of the data reconstruction loss and the physical consistency loss, and through the backpropagation of the total loss function, the parameters of the fusion encoder are optimized so that the fused situation information simultaneously satisfies the dual constraints of data fidelity and physical laws.

[0031] This step details how to embed physical laws into data-driven models using "hard constraints," which is the core technology for solving the "pseudo-fusion" problem. In traditional methods, physical models (such as power flow calculations) and data-driven models (such as neural networks) typically operate independently or are sequentially connected in the process. For example, the data model first predicts the inputs (such as load and power generation), then these predictions are input into the physical model to calculate the state (such as voltage), and finally, the physical constraints are checked. If they are not met, the predictions need to be adjusted manually or by external algorithms, a cumbersome and fragmented process.

[0032] The differentiable power flow calculation layer proposed in this scheme is a revolutionary bridge. Its key lies in achieving differentiable solutions to the complex nonlinear system of power flow equations. Specifically, during forward propagation, it utilizes mature, high-precision numerical methods (such as the Newton-Raphson method) to solve for definite system state variables (such as node voltage magnitudes and phase angles), ensuring the accuracy of physical calculations. During backpropagation to calculate gradients, it does not need to, nor does it trace, the complex iterative process within Newton's method, but cleverly utilizes implicit function theorems. At the solution points obtained from the forward calculation, the power flow equations are satisfied. Taking the total differential of this equation yields a linear equation (i.e., the Jacobian matrix equation) concerning the differential components of the state variables and the differential components of the input variables. Solving this linear equation efficiently yields the gradient of the output state variables with respect to the input parameters (∂state / ∂input). This implementation is equivalent to installing a "gradient calculator" for the physical laws themselves, allowing the entire physical calculation module to be seamlessly embedded into the end-to-end training process, much like a simple activation function in a neural network.

[0033] The physical consistency-driven fusion encoder is a deep network responsible for extracting information from multi-source heterogeneous "big data," but its training objective differs fundamentally from that of traditional autoencoders. Its output is no longer simply reconstructing the input data, but is forced to predict power grid parameters (node ​​injected power, topology connections, etc.). After calculation by the aforementioned differentiable physical layer, the resulting physical state quantities (such as voltage phase angle) must be as close as possible to the actual physical quantity measurements obtained through advanced measurement systems (such as PMUs). This is analogous to training a student not only to memorize textbook text (data reconstruction) but also to understand and apply physical laws to explain and predict real physical phenomena (physical consistency). Through joint training using a weighted combination of "data reconstruction loss" and "physical consistency loss," the encoder is ultimately forced to learn a unique internal representation: this representation not only retains the effective information in the original data to the greatest extent but also naturally and inherently satisfies the fundamental physical laws of power grid operation (such as power conservation and voltage-current relationships). The resulting system situation information provides an accurate and physically feasible "digital twin" environment for all subsequent decision-making processes.

[0034] Further, in step S1.3, the total loss function of the fusion encoder Loss from data reconstruction With the aforementioned physical consistency loss The weighted summation is as follows:

[0035] ,in, and For hyperparameters;

[0036] The data reconstruction loss It is a loss function based on the difference between the input and output of the fusion encoder;

[0037] The physical consistency loss Let be the difference norm.

[0038] The design of the loss function acts as a "command stick" guiding the model's learning direction. The total loss function here embodies the concept of multi-task learning, aiming to enable the fusion encoder to master two skills simultaneously. The data reconstruction loss is responsible for ensuring the encoder possesses strong information compression and feature extraction capabilities. For example, if the input data includes weather cloud maps, load time series curves, and unit status reports, This enables the encoder's intermediate or final output to reconstruct the key features of these data, ensuring that the encoder does not "forget" important patterns in the original data.

[0039] and The physical consistency loss is the core driving signal of this scheme, which measures the encoder's ability to "understand the physical world". Specifically, the predicted node injection power and topology parameters output by the encoder are sent to the differentiable power flow layer (Diff-PF Layer) for calculation to obtain a set of predicted physical state quantities (such as the voltage phase angle of each node).

[0040] At the same time, the system obtains the actual voltage phase angle measurement value at the same moment from real-time measurement systems such as the synchronous phasor measurement unit (PMU). This refers to the difference norm between the two (e.g., mean squared error, MSE). This loss term directly penalizes fusion results that appear reasonable on the data but violate physical laws. Hyperparameters and The adjustment is like adjusting the trade-off between "data fit" and "physical accuracy".

[0041] In practice, initial support may be given. Larger weights force the model to first learn to obey physical laws; as training progresses, these weights can be adjusted to optimize overall decision-making performance. Through this loss function design, the backpropagation gradient signal flows back to the encoder from both the data and physical levels, guiding its parameter updates. Ultimately, this results in an intelligent perception module capable of producing physically reliable situational information, fundamentally eliminating the risk of subsequent decisions being optimized based on physically infeasible scenarios.

[0042] Further, step S2 includes the following steps:

[0043] S2.1, Construct a virtual market mechanism designer and establish a parameterized regulatory response function to simulate the regulatory behavior of market operators on pricing strategies; the regulatory response function calculates market power monitoring indicators based on the input pricing strategy and maps them to the expected intensity of regulatory penalties;

[0044] S2.2, Establish a Stackelberg game optimization framework, with the bidding agent as the leader and the virtual market mechanism designer as the follower, and construct a two-layer optimization model; wherein, the upper-layer optimization objective is the comprehensive return of the bidding agent, which is the difference between the expected market return and the expected regulatory penalty intensity adjusted by weights; the lower-layer optimization is that, given a bidding strategy, the virtual market mechanism designer calculates the expected regulatory penalty intensity based on the regulatory response function;

[0045] S2.3, Solve the incentive-compatible strategy by transforming the optimal response problem of the lower-level virtual market mechanism designer into its first-order optimality condition, and embedding the first-order optimality condition as an equilibrium constraint into the upper-level optimization problem, thereby transforming the two-level optimization problem into a single-level mathematical programming problem for solution, and obtaining the incentive-compatible bidding strategy.

[0046] This step aims to achieve "incentive compatibility" between the intelligent pricing system and the market rules environment, rather than passive avoidance. The construction of a virtual market mechanism designer is crucial. It is not actual market operating agency (ISO) software, but rather an internal learning and simulation environment.

[0047] By using parametric modeling (e.g., setting a market power index threshold based on the deviation between bids and costs and a corresponding penalty function), it can simulate the ISO’s possible responses to current bidding strategies under different levels of regulatory stringency, such as issuing warnings, adjusting clearing prices, or conducting ex-post settlement reviews, and quantify these responses into a “predicted regulatory penalty intensity” value.

[0048] The Stackelberg game framework places the bidding agent (leader) and the virtual supervisor (follower) in a dynamic, sequential decision-making relationship. When formulating a strategy, the bidding agent must pre-consider: If I bid this price, how will the virtual supervisor react according to its rules? What penalty will this reaction incur? Then, after weighing expected gains and anticipated penalties, it makes the optimal decision. The virtual supervisor, given the bidding strategy, faithfully executes its rules to calculate the penalty.

[0049] Solving this bilevel optimization problem is computationally challenging. The approach proposed in this paper essentially represents the lower-level (regulator's) optimization problem (i.e., how to impose penalties) using optimality conditions (such as KKT conditions). These optimality conditions are essentially a set of equality and inequality constraints involving variables from both the upper and lower levels. Then, this set of constraints, along with the objective function of the upper-level (quoting agent), constitutes a new, single-level mathematical programming problem with equilibrium constraints.

[0050] By solving this transformed problem, we can directly obtain the pricing strategy in Stackelberg equilibrium. The brilliance of this strategy lies in the fact that it doesn't simply evade regulatory red lines through trial and error, but rather, based on a deep understanding of the logic of regulatory rules, proactively adjusts its behavior to a state where "even if the regulator knows my entire strategy and takes the optimal response, my behavior is still reasonable and minimizes punishment under their rules." This achieves a paradigm shift from a "cat-and-mouse game" to a "compliance game," intrinsically enhancing the agent's compliance and robustness in the real market.

[0051] Further, step S3 includes the following steps:

[0052] S3.1 Calculate the clearing residual sequence. Based on the publicly available cost information for each trading period, calculate the benchmark predicted clearing result. Subtract the actual market clearing result from the benchmark predicted clearing result to obtain the clearing residual sequence that characterizes the collective strategic behavior of the counterparties.

[0053] S3.2, Extracting the strategy feature vector: Feature extraction is performed on the clearing residual sequence; the feature extraction includes: calculating higher-order statistics of the sequence, the higher-order statistics including at least skewness and kurtosis; and / or performing time-frequency domain analysis on the sequence to extract its frequency domain pattern features; combining the extracted features into a low-dimensional strategy feature vector;

[0054] S3.3, Online identification of game states, construction of hidden Markov models with different market game stages as hidden states; using the continuously acquired policy feature vectors as observation sequences, online updating of the probability distribution estimate of the current hidden state using Bayesian filtering algorithm, to complete dynamic opponent behavior identification.

[0055] This step provides a novel and efficient method for dynamically perceiving the behavior of adversaries. Its core idea is "backtracking from results to patterns," avoiding the nearly impossible task of directly modeling the complex and ever-changing internal decision-making mechanisms of each competitor. Calculating the clearing residuals is the first step. The benchmark clearing result is the idealized market clearing outcome (including electricity prices and unit output distribution) assuming all market participants make honest bids based entirely on publicly available, short-term marginal costs (such as fuel costs).

[0056] The deviation from the actual clearing result is the combined effect of the strategic behaviors of all participants (such as holding capacity, raising prices, and following the leader). This residual sequence is a direct external manifestation of the market game state. Feature extraction is the process of transforming continuous residual signals into analyzable discrete features.

[0057] Higher-order statistics such as skewness and kurtosis can capture the asymmetry and sharpness / flatness of the residual distribution. For example, persistent negative residuals (actual electricity prices are lower than benchmark cost prices) may correspond to fierce price wars, and their distribution may be left-skewed; while occasional large positive residuals (soaring electricity prices) will make the distribution exhibit a peak and heavy tails, suggesting that there is exploratory or coordinated price-raising behavior in the market.

[0058] Time-frequency domain analysis (such as Fourier transform or wavelet transform) can reveal periodic or specific frequency patterns in the residual sequence, for example, whether there are strategy cycles on a weekly or monthly basis. Combining these multidimensional features into a low-dimensional "strategy fingerprint" vector enables dimensionality reduction and encoding of complex market states.

[0059] Online game state identification employs the powerful probabilistic graphical model Hidden Markov Model (HMM). It defines unobservable market game "stages" (such as the "free competition period," "probing collusion period," and "convergence period after regulatory intervention") as hidden states. The observed "policy fingerprint" vector is the external manifestation of these hidden states, but it is noisy.

[0060] The function of Bayesian filtering algorithms (such as the forward pass algorithm) is to update the confidence distribution of which hidden state the market is most likely to be in in real time and recursively, based on the latest observed "policy fingerprint," combined with probability estimates of historical states and the inherent transition patterns between states (described by the transition matrix). In this way, the system can dynamically "sense" subtle changes in the market's game atmosphere in a probabilistic manner. For example, it can determine that the market is transitioning from a stable competitive state to a "turbulent period" with increased strategy uncertainty, thus providing key leading signals for its own strategy adjustments.

[0061] Further, step S4 includes the following steps:

[0062] S4.1, Construct a rule-change risk model. Based on the occurrence time of historical market rule-change events and multi-dimensional covariates prior to the events, a Cox proportional hazards model is used for modeling. The Cox proportional hazards model is in the form of... ,in Indicates that given covariates Next moment Instantaneous risk rate As the benchmark risk function, This is a vector of covariate coefficients;

[0063] S4.2, Calculate the survival probability and risk adjustment factor based on the Cox proportional hazards model and the current covariates. Calculate future moments The rule-stable probability, i.e., the survival function. The survival function is applied in a discrete decision cycle. The value at the location It is directly defined as the risk adjustment factor for this cycle;

[0064] S4.3, Execute risk-adjusted decisions: In multi-period pricing strategy optimization, construct the risk-adjusted decision value function. ,in The time discount factor, For the first Expected returns for each cycle; based on the value function Optimize the strategy.

[0065] This step treats the electricity market rules environment as a dynamic "living organism," where "survival" means that the rules remain stable, while "failure" means that the rules undergo abrupt changes (such as introducing a new price ceiling, modifying the market power monitoring algorithm, suspending the market, etc.).

[0066] The Cox proportional hazards model is a standard tool in survival analysis, well-suited for analyzing the relationship between event timing and multiple covariates. The model takes the form of... ,in It is the baseline risk function, which represents the trend of the inherent risk of rule mutation over time without considering the influence of any covariates; This reflects the covariates The multiplier effect on risk.

[0067] For example, if It includes a "market power index", and its corresponding coefficient A positive and significant value means that when the market power index rises, the risk of rule abrupt changes increases exponentially, which aligns with the intuition that regulators are more likely to intervene when there are signs of market manipulation.

[0068] The model is trained based on historical data, specifically the time points of a series of regular mutation events in the past, and the values ​​of various covariates observed just before each event. The coefficient vector can be estimated by maximizing the partial likelihood function. This allows us to obtain a model that can quantify the current level of risk.

[0069] Survival function The calculation is performed given the current covariates. Under these conditions, the rule will survive for at least time in the future. The probability is calculated as follows. This is a curve that gradually decays from 1 (completely stable) to 0 (inevitable mutation). Its application in decision-making is ingenious: in multi-period optimization, future returns are typically represented by a discount factor. To convert it to its current value, This means that the further away the investment is, the lower the return value. This plan will adjust the risk factor. Multiplying this by the time discount factor creates a "risk-adjusted discount factor." This means that if the model predicts a high probability of a regular abrupt change on a future day (i.e.,...), then... (Very small), so even if the expected return for that day is... The risk of high returns is extremely high, and its weight in the current decision objective function will be significantly reduced because the uncertainty (risk) of realizing this high return is extremely high. Conversely, for recent days with stable rules, its return weight is close to that of traditional discounting. In this way, when the system plans long-term strategies, it will automatically and inherently tend to avoid high-risk periods and allocate more resources to recent transactions with higher certainty, thereby achieving a forward-looking and robust decision-making based on risk probability.

[0070] Further, step S5 includes the following steps:

[0071] S5.1 defines and parses strategy instructions, providing a predefined, syntactically rigorous formal specification language for receiving strategy logic defined by traders; the structured statements of the formal specification language can clearly express market conditions, target assets, and specific adjustment actions;

[0072] S5.2, Compilation and Constraint Transformation: The compiler parses the structured statements and transforms them into one or more deterministic mathematical constraints that can be embedded in the mathematical optimization model, thus forming a new constraint set.

[0073] S5.3, Verification and Intent Reconciliation: The newly added constraint set is merged with the inherent physical and rule constraint set, and satisfiability is determined using the constraint solver.

[0074] If it is determined that the constraint can be satisfied, then the newly added constraint set is injected into the optimization model;

[0075] If it is determined to be unsatisfactory, the conflict resolution optimization process is initiated: an optimization problem is constructed with the goal of minimizing the semantic deviation of the structured statement, and the problem is solved under the condition of satisfying all inherent constraints. A corrected feasible strategy is output, and explanatory information describing the reasons for the deviation is generated.

[0076] This step establishes a precise and reliable human-machine communication and collaboration protocol, aiming to seamlessly and unambiguously integrate the experience and wisdom of human experts into automated decision-making processes. Its core is the definition of a formal specification language (DSL). This language differs from natural language; it has strict grammatical and semantic definitions, avoiding ambiguity caused by vague expressions such as "significantly increase" or "closely monitor." It allows traders to express strategy logic in a structured manner, for example: "If (the real-time wind power output of region A is lower than 70% of its day-ahead forecast), and (the time is within one hour before the market closes on the trading day), then (increase the price of gas turbine unit G1 in region A by 5%)."

[0077] These statements contain clear conditions, timelines, and actions. The compiler's role is to automatically translate these human-friendly, logical instructions into mathematical language that the optimization model "understands"—a set of equality or inequality constraints. For example, the above instructions might be translated into conditional constraints in a mixed-integer linear programming (MILP) model. The next crucial step is satisfiability verification.

[0078] The system merges this newly generated set of constraints with the set of inherent, inviolable "hard constraints" in the model (including all physical safety constraints, such as line power flow limits and generator ramp rates; and all core market rule constraints, such as price limits).

[0079] Then, a constraint solver (such as a mathematical programming solver or an SMT solver) is invoked to check whether the merged constraint set contains at least one feasible solution. If it does, it means the trader's intent does not conflict with the system's fundamental rules, and the instruction is adopted and injected into the model. If no feasible solution exists, it indicates an "intent conflict" has occurred.

[0080] At this point, instead of simply displaying an error message like a traditional system saying "the instruction is not feasible," the system initiates an intelligent conflict resolution process. This process is structured as an optimization problem: finding a strategy that requires the "minimum modification" to the trader's original instruction, while ensuring all inherent hard constraints are met.

[0081] The "modification magnitude" here needs to be precisely defined. For example, it can be measured as the difference (norm) between the price adjustment value in the new strategy and the adjustment value required by the original instruction. By solving this optimization problem, the system can automatically generate an "amendment": it is the one among all feasible solutions that best reflects the trader's original intention. Simultaneously, by analyzing which hard constraints are activated (i.e., becoming limiting factors) during the optimization process, the system can generate explanatory reports, such as: "Your instruction caused the L123 flow to exceed the limit by 2% in scenario S, therefore it is recommended to adjust the upward adjustment from 5% to 3%." This greatly improves the efficiency of human-machine collaboration and the utilization rate of expert experience.

[0082] Furthermore, in step S5.3, the conflict resolution optimization process is specifically implemented in the following manner:

[0083] For each mathematical constraint in the newly added constraint set, a corresponding slack variable is introduced, and all slack variables constitute the slack variable vector δ.

[0084] Each mathematical constraint in the newly added constraint set is rewritten as a relaxed constraint containing its corresponding relaxed variable, and all relaxed constraints constitute a relaxed constraint set.

[0085] Construct and solve the following mathematical optimization problem:

[0086] Minimize: the norm |δ| of the slack variable vector δ.

[0087] Constraints: The inherent physical and rule constraint set must be satisfied, and the relaxed constraint set must also be satisfied.

[0088] The values ​​of the decision variables obtained from the solution are used as the modified feasible strategies.

[0089] This is a specific mathematical implementation of the conflict resolution optimization process in step S5.3, providing a systematic and computable "minimum modification" principle. Its core idea is to "soften" the new constraints corresponding to human expert instructions, rather than rigidly treating them as inviolable "edicts."

[0090] In practice, a corresponding slack variable is introduced for each mathematical constraint generated by the instruction compilation. For example, if a constraint after instruction conversion is "gas unit price ≥ 105% of the original plan", then a non-negative slack variable δ can be introduced to rewrite this constraint as "gas unit price + δ ≥ 105% of the original plan".

[0091] If δ=0, the original instruction is strictly executed; if δ>0, it means the actual price can be lower than 105%, but the larger δ is, the greater the deviation from the original instruction. All slack variables constitute a vector δ. The conflict resolution problem is then transformed into: under the premise of satisfying all inherent physical and market rule hard constraints, finding a decision scheme (such as the price of each unit) and simultaneously determining the values ​​of a set of slack variables δ such that a certain norm (such as L1 norm or L2 norm) of the slack variable vector is minimized.

[0092] minimize This objective function directly reflects the intention of "minimizing the semantic deviation from the original instruction". The L2 norm (sum of squares) tends to relax all constraints slightly, while the L1 norm (sum of absolute values) may tend to relax only a few "bottleneck" key constraints.

[0093] By solving this optimization problem, the optimal decision variable values ​​(such as the final price curve) are the "most feasible and closest to the original intention" reconciled strategy. The optimal value of the slack variable δ quantifies the degree to which each instruction constraint needs to be relaxed, which can be directly used to generate explanatory feedback. This method transforms conflict resolution from a process requiring complex heuristic rules or extensive manual intervention into a clear, automatically solvable mathematical optimization problem, ensuring the optimality and consistency of the solution.

[0094] Furthermore, in step S4.1, the multi-dimensional covariates upon which the Cox proportional hazards model is constructed are comprehensive indicators extracted by integrating real-time market operation data, macroeconomic environment information, and public opinion data; specifically, they include at least two of the following categories of indicators:

[0095] Market structure and behavior indicators, including real-time market power indices that reflect market concentration and manipulation potential;

[0096] Market volatility indicators, including market price volatility, which reflects short-term price uncertainty;

[0097] Policy and public opinion indicators, including policy intervention sentiment trend indicators obtained through sentiment analysis of relevant policy announcements and news texts;

[0098] Fundamental supply indicators include fuel supply tightness indicators that reflect the availability and cost pressures of fuels for power generation.

[0099] The effectiveness of the Cox proportional hazards model is highly dependent on whether the selected covariate X can comprehensively and timely capture the risk factors that lead to rule mutations. This approach emphasizes extracting comprehensive cross-domain indicators from multi-source data, rather than single-dimensional data. Market structure and behavioral indicators, such as the real-time market power index, directly reflect market supply and demand forces and potential manipulation risks, and are among the most important focuses of regulatory agencies.

[0100] When a few generating units or power groups can significantly influence marginal prices, the urgency for regulatory intervention increases. Market volatility indicators, such as price volatility, measure the stability of market prices. Abnormal, non-fundamentally driven high volatility is often a signal of market failure or aberrant participant behavior, and is also likely to attract regulatory attention.

[0101] Policy and public opinion indicators are derived through sentiment analysis and theme extraction from official announcements, industry news, and social media discussions using natural language processing technology. For example, if the sentiment towards keywords such as "market power," "price manipulation," and "reform" in recent policy texts and media reports turns negative and their frequency increases, it may indicate that regulatory agencies are considering or about to introduce new regulatory measures. This is a forward-looking risk perception based on unstructured information.

[0102] Fundamental supply indicators, such as fuel supply tightness, reflect the cost pressures on power generation and the overall adequacy of the system's supply. During periods of fuel shortages and high prices, power generators face significant cost pressures and may be more inclined to bid higher prices to cover costs or reduce losses. This can drive up overall electricity prices, increase the social cost of electricity and political pressure, thereby increasing the probability of government intervention in temporary pricing or rule changes.

[0103] By integrating at least two different types of indicators, the Cox model can construct a comprehensive, multi-faceted risk profile. For example, even if the current market power index is low, but policy sentiment is extremely negative and fuel supply is tight, the model may still identify a high risk of rule-breaking. This multi-dimensional, multi-modal covariate design allows the risk quantification model to not only rely on internal market data but also sensitively perceive changes in the external policy environment and macroeconomic fundamentals, greatly enhancing the comprehensiveness and accuracy of risk warnings.

[0104] Furthermore, the method is implemented according to the following workflow within a complete transaction decision cycle:

[0105] On trading day D, step S1 is executed first, based on the latest multi-source heterogeneous data up to that day, physical-data fusion sensing is performed, and physical consistency system status information for the next trading day D+1 is output.

[0106] Subsequently, based on the physical consistency system situation information and recent historical market data, step S3 is executed to dynamically identify the current market game state, and at the same time, based on the current macroeconomic environment data, step S4 is executed to quantify the risk of future rule mutations.

[0107] Next, based on the physical consistency system situation information, and incorporating the market game state identified in step S3 and the risk information quantified in step S4, step S2 is executed to generate a preliminary pricing strategy for trading day D+1.

[0108] After receiving the formal strategy instructions input by the trader based on the review of the preliminary pricing strategy, step S5 is executed to compile and verify the formal strategy instructions, reconcile them with the preliminary pricing strategy, generate the final pricing strategy, and submit it.

[0109] After the market clearing is completed on trading day D+1, the actual clearing result is obtained, and step S6 is executed. Based on the difference between the actual clearing result and the output or decision result of the corresponding step in steps S1 to S5, a training signal is generated to update the parameters of one or more models involved in the method.

[0110] The above describes how to connect the various independent technical modules to form a complete, logically structured industrial workflow. It clearly demonstrates how, in a real trading day, this intelligent quote-assisted decision-making system works collaboratively step by step, starting with data preparation, ultimately generating and submitting quotes, and subsequently learning from them.

[0111] On trading day D, the process begins with step S1, physical-data fusion sensing. It functions like an "intelligence analysis center" that starts work promptly every morning, gathering multi-source data such as the latest weather forecasts, equipment status, network topology, and load predictions. Through a differentiable physical fusion model, it processes and produces a set of physically self-consistent forecast scenarios for the next trading day (D+1). These scenarios serve as the "sandbox" for all subsequent analysis and decision-making.

[0112] Subsequently, the system initiates market "pulse" diagnosis and long-term "climate" assessment in parallel or sequentially. The dynamic counterparty behavior identification module in step S3 analyzes the market clearing residual sequence in the recent period (e.g., the past 5-10 trading days), and like a traditional Chinese medicine practitioner taking a pulse, determines the current "constitution" state of the market game (e.g., a "fiery" price-raising probing period or a "peaceful" cost competition period).

[0113] Meanwhile, the systemic risk quantification module in step S4, much like checking a weather forecast or earthquake warning, comprehensively assesses the probability of a "storm" (sudden change) in the market rules environment over the next few days (e.g., from D+1 to D+7) based on current macroeconomic data. This information is provided to the strategy optimization engine in real time.

[0114] Next, the strategy optimization layer in step S2 begins to work as the "decision center". It takes the physically consistent scenario produced by S1 as its base, incorporates the current opponent behavior characteristics diagnosed by S3 (for example, if the market is judged to be in a probing period, the optimization model will assume that the opponent's bid may be more aggressive), and adopts the risk-adjusted discount factor provided by S4 (for example, if the risk on day D+3 is high, the weight of that day's return in the overall objective is reduced).

[0115] In this integrated environment, a preliminary pricing strategy draft for trading day D+1 is generated by solving the Stackelberg game optimization problem. Once this draft is presented to a human trader, the human-machine intent reconciliation layer in step S5 begins to function as an "intelligent assistant." The trader provides revisions to the draft based on their professional experience (using formal language), and the system immediately compiles, verifies, and resolves any potential conflicts, ultimately seamlessly integrating human intelligence with machine computation to produce a final strategy acceptable to both parties and submit it.

[0116] The entire decision-making process demonstrates efficiency, collaboration, and intelligence. Finally, after the actual market clearing on day D+1, the closed-loop evolution mechanism of step S6 is triggered. The system acts like a rigorous "debriefing meeting," comparing the predictions of each module from the previous day (such as scenario prediction, game state judgment, risk probability, and pricing strategy returns) with the actual results. The resulting error signals are used as training data to fine-tune or periodically retrain relevant models (such as fusion encoders, behavior recognition HMM parameters, risk model coefficients, etc.).

[0117] This process enables the system to learn from each market practice, continuously evolve its forecasting and decision-making capabilities, adapt to long-term changes in the market environment, and truly realize an intelligent agent with a complete cycle of "perception-decision-action-learning".

[0118] The intelligent pricing and decision-making support method for electricity spot trading based on multi-source data fusion provided by this invention has the following significant advantages compared to existing technologies:

[0119] It provides a physically consistent and reliable foundation for system situational awareness: through joint training of a differentiable physical layer and a fusion encoder, it ensures that the power grid state information on which all subsequent decisions are based not only fits the data patterns, but also strictly satisfies the physical laws of the power system. This fundamentally eliminates the risk of making optimization decisions based on physically infeasible scenarios and improves the physical security and feasibility of the decisions.

[0120] It achieves strategy generation with intrinsic security and proactive compliance: By constructing the Stackelberg game framework, regulatory responses are internalized as part of the agent's decision-making model, prompting the bidding strategy to actively form incentive compatibility with regulatory rules while pursuing returns. This enhances the agent's long-term robustness and compliance in a strict regulatory environment and reduces the risk of being punished.

[0121] It achieves dynamic and non-intrusive identification of the collective game style in the market: by analyzing the market clearing residual sequence and extracting strategy features, it can keenly capture subtle changes in the overall game state of the market in a macro and indirect way, without having to estimate the private information of individual opponents, providing a leading signal for timely adjustment of its own strategy and improving the adaptability and pertinence of the strategy.

[0122] It provides a forward-looking probabilistic assessment and avoidance capability for extreme risks such as rule mutations: It innovatively applies a survival analysis model to quantify the probability of future rule stability and integrates this risk information into the value function of multi-period decision-making, enabling the system to automatically and inherently tend to avoid high-risk periods, thereby enhancing the forward-looking nature and robustness of long-term decision-making.

[0123] It achieves precise and unambiguous integration and collaboration between human expert experience and machine intelligence: through formal specification language and automated conflict resolution mechanism, it can accurately transform the strategic intent of human experts into machine-executable constraints, and when it conflicts with the hard constraints of the system, it can intelligently generate feasible modification solutions that are closest to the original intent, which greatly improves the efficiency and decision quality of human-machine collaboration and gives full play to the advantages of human-machine hybrid intelligence.

[0124] It endows the decision-making system with continuous self-optimization and evolution capabilities: through a closed-loop evolution mechanism, the system can automatically update multiple key model parameters based on actual market feedback, enabling it to adapt to dynamic changes in the market environment, overcome the problem of static model performance degradation, and ensure the long-term effectiveness and advancement of the decision-making system.

[0125] A complete intelligent decision-making system with deep integration and synergistic evolution across multiple stages has been constructed: This invention organically integrates six stages—perception, optimization, identification, quantification, reconciliation, and evolution—to form a closed-loop intelligent decision-making system. The stages are deeply coupled and information flows smoothly, jointly improving the overall decision-making efficiency and robustness in the face of a highly dynamic, highly uncertain, and strictly regulated electricity spot market environment. Detailed Implementation

[0126] The present invention will be further described below with reference to specific embodiments:

[0127] Example:

[0128] (I) System Hardware and Software Operating Environment

[0129] The system of this invention is deployed on an integrated high-performance computing platform, and its specific configuration is as follows:

[0130] Hardware platform:

[0131] Data Acquisition and Interface Server: Equipped with an Intel Xeon Silver 4314 processor (16 cores / 32 threads) and 128GB DDR4 memory for high-concurrency access to multi-source data. It features multi-port gigabit / 10-gigabit fiber optic network cards, supporting power system protocols such as IEC 60870-5-104 and IEC 61850, as well as common data interfaces such as RESTful API and Kafka message queues.

[0132] Model training and high-performance computing server: Equipped with four NVIDIA A100 80GB PCIe GPUs for massively parallel training and inference of deep neural networks. It also features dual AMD EPYC 7763 processors (128 cores / 256 threads) and 1TB of DDR4 memory for solving large-scale mathematical programming problems (such as MILP and MPEC).

[0133] Application and Interaction Server: Equipped with an Intel Xeon Gold 6330 processor and 256GB of memory, it deploys business logic, databases (PostgreSQL for storing historical data, Redis for caching real-time data), and human-computer interaction web services (based on Django or Spring Boot framework).

[0134] Networking and Storage: All servers are interconnected via a 100GbE RoCE (RDMA over Converged Ethernet) network, ensuring low-latency data transmission. A distributed Ceph storage cluster provides petabyte-level high-reliability data storage.

[0135] Software environment:

[0136] Operating System: The compute nodes run Ubuntu 20.04 LTS or Rocky Linux 8.5.

[0137] Core development framework:

[0138] Deep learning: PyTorch version 1.12.0 and above, with the functorch library configured to support higher-order differentiation for implementing differentiable physical layers and fusion encoders.

[0139] Automatic Differentiation and Scientific Computing: JAX 0.3.25 and above, used as another efficient solution for implementing a differentiable power flow calculation layer.

[0140] Optimization solutions: Gurobi Optimizer 9.5 or CPLEX 22.1 are used to solve mixed integer linear programming (MILP) and some nonlinear problems; IPOPT 3.14.5 is used to solve large-scale nonlinear programming (NLP) and equilibrium constraint problems (MPEC).

[0141] Constraint Solving and Formal Verification: Z3 Theorem Prover 4.8.15 is used to determine the satisfiability (SAT / SMT) of the compiled constraint set.

[0142] Data Acquisition and Processing: Apache NiFi 2.0 was used for data stream orchestration; Pandas, NumPy, and SciPy were used for numerical computation and feature engineering; NLTK, TextBlob, or dedicated pre-trained models (such as BERT) were used for text sentiment analysis of policy public opinion.

[0143] Human-computer interaction interface: Based on the Vue.js 3.x front-end framework, it provides visual strategy configuration (DSL generator), result display and interactive reporting.

[0144] (II) Detailed Implementation Process of Key Steps

[0145] The steps of the present invention will be described in detail below with reference to a preferred embodiment.

[0146] Taking the operation process of a certain power generation group G on a certain electricity spot market on a certain trading day D as an example

[0147] Step S1: Specific Implementation of the Physical-Data Fusion Sensing Layer

[0148] The goal of this step is to build an end-to-end trainable module called PhysFusionNet.

[0149] S1.1 Implementation of Differentiable Power Flow Computation Layer (Diff-PF Layer)

[0150] For simplicity, this embodiment uses a DC power flow model as an example. The function of this layer is to implement mapping. ,in Inject net active vectors into nodes (except for balanced nodes). This is the imaginary part (electric susceptance matrix) of the nodal admittance matrix. This is the node voltage phase angle vector.

[0151] Traditionally, solving Calculation required (Assuming) (Full rank). But in neural networks, we need to calculate the loss function. For input and gradient and .

[0152] We implement a DifferentiableDCPF class. Its core is to use the implicit function theorem for gradient calculation.

[0153] Forward propagation: receiving and Call an efficient and stable linear equation solver (such as torch.linalg.solve based on LU decomposition) to compute. This step is completely consistent with traditional numerical solutions, ensuring the accuracy of physical calculations.

[0154] Backpropagation: Given the gradient from upstream We need to calculate and .

[0155] According to the DC power flow equation Its total differential: .

[0156] Using the implicit function theorem, we can derive:

[0157]

[0158] in .

[0159] In actual code, we don't need to explicitly write these formulas. Instead, we use the automatic differentiation tool to define custom gradient functions for the forward propagation solution steps. In PyTorch, this can be achieved by extending `torch.autograd.Function`:

[0160] "class DifferentiableDCPF(torch.autograd.Function):

[0161] @staticmethod

[0162] def forward(ctx, P, B):

[0163] # ctx is used to store intermediate variables during forward propagation, which are then used for backward propagation.

[0164] theta = torch.linalg.solve(B, P) # Forward solver

[0165] ctx.save_for_backward(theta, B)

[0166] return theta

[0167] @staticmethod

[0168] def backward(ctx, grad_theta):

[0169] theta, B = ctx.saved_tensors

[0170] # Solve the linear system of equations B^T * grad_P = grad_theta to obtain grad_P

[0171] grad_P = torch.linalg.solve(BT, grad_theta)

[0172] # Calculate grad_B (simplified process to demonstrate the principle)

[0173] # grad_B = -grad_P.unsqueeze(2) @ theta.unsqueeze(1) # Involves three-dimensional tensors; efficiency needs to be considered in the specific implementation.

[0174] # For simplicity, grad_P and the gradient with respect to B can be returned here (possibly None or an approximation).

[0175] grad_B = None # In practice, an efficient Lyapunov equation solver or approximation method can be used.

[0176] return grad_P, grad_B”

[0177] In actual production code, more efficient numerical linear algebra libraries (such as the GPU version of scipy.linalg.solve_lyapunov) will be used to compute the equations. The gradient.

[0178] S1.2 Physically Consistent Driven Fusion Encoder Structure

[0179] PhysFusionEncoder is a multimodal deep neural network.

[0180] enter:

[0181] Time series data: Load, wind power, and photovoltaic output sequences predicted for the past 24 hours and the next 24 hours (shape: [batch, time_steps, num_nodes]), with time series features extracted using a Long Short-Term Memory (LSTM) network or a TransformerEncoder.

[0182] Spatial data: Meteorological forecast grid data (temperature, wind speed, irradiance), spatial features are extracted through convolutional neural networks (CNN), and then mapped to power grid nodes through fully connected layers.

[0183] Topology data: The node-branch connections of the current power grid, represented as a graph. Node features include unit type and capacity, while edge features include branch reactance. Graph neural networks (GNNs), such as graph convolutional networks (GCNs) or graph attention networks (GATs), are used to learn topological embeddings.

[0184] Feature fusion and output: The final hidden state of the LSTM / Transformer, the spatial feature vector output by the CNN, and the node embedding vector output by the GNN are concatenated, and then passed through a multilayer perceptron (MLP). The output layer of this MLP consists of two parts:

[0185] output_P: The net active power injected into each node during a predicted future period (e.g., 96 points on day D+1).

[0186] output_B_params: Predicted grid equivalent admittance matrix parameters (e.g., predicting the on / off state of critical lines, thereby modifying...). ).

[0187] Physical consistency constraint: Reassemble output_P and output_B_params into a sum and send it to the DifferentiableDCPF layer to obtain the predicted voltage phase angle.

[0188] S1.3 End-to-end Joint Training Process

[0189] Data preparation: Collecting historical datasets . It is a moment Multi-source heterogeneous raw data, It is the actual voltage phase angle vector obtained by PMU measurement at the corresponding moment.

[0190] Loss function definition:

[0191] Data reconstruction loss Encourage the encoder to learn meaningful representations. For example, we can add a decoder in the intermediate layer of the encoder to attempt to reconstruct the input load sequence using mean squared error (MSE). .

[0192] Physical consistency loss : Core driver loss. .

[0193] Total loss: In this embodiment, after grid search, the settings are... , Initially, the emphasis is on physical laws, but fine-tuning is possible later.

[0194] Training process: Using the Adam optimizer, initial learning rate The batch size is 32. During training, the gradients of the DifferentiableDCPF layer are backpropagated to output_P and output_B_params, thereby updating the parameters of the entire PhysFusionEncoder. After approximately 200 epochs of training, the encoder learns to output physically highly reliable system situation predictions. ( (Scenario)

[0195] Step S2: Specific Implementation of Intrinsic Security and Compliance Strategy Optimization

[0196] S2.1 Virtual Regulator Modeling

[0197] We design a parameterized regulatory response model. First, we define a real-time market power indicator. ,in It is the quote curve (vector) of the quote agent.

[0198] Indicator Calculation: .in, For the unit The quote, For its publicly disclosed short-term marginal cost, Based on its recent market forecast of the number of contracts won, This represents the total number of generating units. This indicator combines the price-cost deviation with market share.

[0199] Penalty function: .in It is a regulatory threshold (e.g., 0.15). This is the penalty intensity coefficient. This quadratic function simulates the increasing, non-linear penalty intensity imposed by regulators on market behavior.

[0200] S2.2 Construction of Stackelberg Game Optimization Framework

[0201] We formalize the pricing decision-making process as a two-level program.

[0202] Upper layer (leader - quotation agent):

[0203]

[0204] in, Market returns (depending on the scenario) ), It is a compliance trade-off coefficient. This is the optimal response (i.e., penalty value) of the lower-level virtual regulator to the offer.

[0205] Lower layer (Followers - Virtual Overseers):

[0206]

[0207] This lower-level problem is quite simple, and its optimal solution is... However, its significance lies in clarifying regulatory rules as an optimization problem, which facilitates subsequent differentiation.

[0208] S2.3 Solving for incentive-compatible strategies

[0209] We use the recursive differentiation method to solve it.

[0210] Initialization: The policy of the bidding agent is implemented using a policy network. It indicates that its parameters are The input is the situation information generated by S1, and the output is the price quote curve. .

[0211] Lower-level response: For a given ,calculate And thus obtain This step is a forward computation.

[0212] Gradient calculation and upper-layer update: The upper-layer objective function affects the policy network parameters. The gradient is:

[0213]

[0214] in, . It can be obtained through automatic differentiation.

[0215] Iterative optimization: Update using gradient ascent method : Repeat steps 2-4 until the policy network converges. Finally, the policy network... Learn to intrinsically suppress pricing behavior that leads to high market power penalties while pursuing high returns.

[0216] Step S3: Specific Implementation of Dynamic Opponent Behavior Recognition

[0217] S3.1 Calculation of the Clearing Residual Sequence

[0218] For the Historical trading days (or trading sessions):

[0219] Obtain publicly available cost data: Obtain fuel type, heat rate, and daily fuel price delivered to the plant for all participating units from publicly available market information platforms. (Computer Group) Short-run marginal cost (SRMC):

[0220]

[0221] Benchmark clearing calculation: Construct a day-ahead market clearing model (linear programming) with the objective of minimizing total social electricity purchase cost:

[0222]

[0223] Residual calculation: Obtain the actual clearing results for the day: actual output and actual nodal marginal price. Define the residuals in two dimensions:

[0224] Electricity price residual: (Choose between primary nodes or weighted average).

[0225] Output residual: These can be summarized as residuals for key unit groups (e.g., the top three thermal power plants and the total hydropower plant).

[0226] Forming the first The residual vector of the day ,in Indicates possible spatial or category aggregations.

[0227] S3.2 Policy Feature Vector (Policy Fingerprint) Extraction

[0228] For a sliding time window (e.g., the past) Residual sequence (for each trading day) Perform feature engineering:

[0229] Higher-order statistical features:

[0230] Skewness: It measures the asymmetry of the distribution.

[0231] Kuroshi: This measures the sharpness of the distribution.

[0232] Time-frequency domain features: Discrete wavelet transform (DWT) is performed on the residual sequence. Using the db4 wavelet basis function, a three-level decomposition is performed to obtain approximation coefficients (cA3) and detail coefficients (cD1, cD2, cD3). The energy of each level of detail coefficients is calculated as a feature.

[0233]

[0234] This can capture the fluctuation patterns of residual sequences at different time scales (such as intraday and interday).

[0235] Autocorrelation characteristics: Calculate the autocorrelation coefficients of the residual series at lags of 1, 2, and 3 days. .

[0236] Finally, these features are pieced together into one Dimensional policy feature vector (policy fingerprint) .For example: .

[0237] S3.3 Online State Recognition Based on Hidden Markov Model (HMM)

[0238] Model definition: Define a model with A hidden state HMM.

[0239] state This period is dominated by cost competition. Its characteristics include residuals with near-zero mean, low volatility, and no significant autocorrelation.

[0240] state This is the period of exploratory strategies. Characteristics include positive residual skewness (higher electricity prices), peak frequency (occasionally high price spikes), and relatively high energy in mid-to-low-level wavelets (severe short-term fluctuations).

[0241] state Potential period of coordinated equilibrium. Characterized by strong positive autocorrelation in the residual sequence (persistent high price trend) and significant energy in high-level wavelets (medium- to long-term fluctuation pattern).

[0242] Observation: The strategy feature vector observed each trading day Assuming a given state, Follows a multivariate Gaussian distribution .

[0243] Model pre-training: using a sufficiently long historical period (e.g., the past year) The parameters of the Hidden Markov Model (HMM) are estimated using the Baum-Welch algorithm (an expectation-maximization algorithm):

[0244] Initial state distribution .

[0245] State transition probability matrix ,in .

[0246] Observation probability parameters .

[0247] Online Bayesian Update (Filtering): On Trading Day When we obtain the latest strategy fingerprint Then, the forward algorithm is used to update the current state online. Confidence level (filter probability):

[0248] Define forward variables .

[0249] Recurrence formula:

[0250]

[0251]

[0252] in .

[0253] In time Currently in state The probability (filter probability) is:

[0254]

[0255] The system uses the state with the highest probability as the judgment of the current market game stage (e.g.) (In the trial strategy period), the statistical characteristics of historical counterparts' behavior (such as the average price increase) in this state are passed to the strategy optimization layer.

[0256] Step S4: Specific Implementation of Systemic Risk Quantification

[0257] S4.1 Construction and Training of the Cox Proportional Risk Model

[0258] Event definition and data preparation: collecting past data Timing of major market rule changes (such as revising price caps, introducing new market power monitoring rules, or temporary intervention in settlement) (In days). For each event Extract the time point before the event occurred (e.g.) (day) multidimensional covariate vector Meanwhile, for dates on which no event occurred, the observation time is the last day in the sample (right censoring).

[0259] covariates Specific calculations:

[0260] Market power index ( ): Use the definition in S2.1 Calculate the moving average over the past 5 days.

[0261] Price volatility ( : The annualized volatility of the daily returns (logarithmic difference) of the current market electricity price series over the past 20 days.

[0262] Policy intervention sentiment trends ( ): The headlines of news articles on the official websites of national and local energy regulatory agencies over the past week were scored using a sentiment analysis model (such as a fine-tuned model based on BERT) (-1 for extremely negative, 0 for neutral, and 1 for extremely positive), and the average score was taken.

[0263] Fuel supply tightness ( : The inverse of the ratio of the number of days of available inventory of the region's main power generation fuel (such as thermal coal) to the historical average for the same period, and then taking the logarithm.

[0264] Model training: Using the Cox proportional hazards model The coefficients are solved using the partial likelihood estimation method. :

[0265]

[0266] in Indicates that an event has occurred. It is in time The sample set is still at risk. Fitting is performed using the lifelines library or the survival package in R. Assume the fit yields... This indicates that the influence of public opinion on policy has the greatest weight.

[0267] S4.2 Survival Probability Calculation and Risk Adjustment

[0268] Calculate the baseline risk function: Estimate the cumulative baseline risk using the Breslow estimator. .

[0269] Calculate future survival probability: on the trading day Based on currently observed covariates Calculate the future number The probability of survival if the rules remain stable (without mutation) over a given day (relative to today):

[0270]

[0271] It is possible to calculate the next 7 days ( The survival probability curve of ).

[0272] Risk-adjusted decision making: In formulating a plan that covers the future Heaven (as) When formulating a pricing strategy, a risk-adjusted multi-period decision value function is constructed:

[0273]

[0274] in It is the standard time discount factor. It is the first The expected daily return function. The system optimizes pricing. To maximize In this case, lower decision weights will be automatically assigned to future dates with a low probability of survival (high risk).

[0275] Step S5: Specific Implementation of Human-Machine Intent Harmony

[0276] S5.1 Definition and Analysis of Domain-Specific Languages ​​(DSLs)

[0277] This embodiment designs a formal specification language called EML (Electricity Market Language). Its core syntax is defined using Extended Backus Normal Form (EBNF) as follows:

[0278] "Strategy ::= (IfThenStatement | DirectAction) ["," Strategy]

[0279] IfThenStatement ::= "IF" Condition "THEN" Action

[0280] Condition ::= Metric Comparator Value [LogicalOp Condition]

[0281] LogicalOp ::= "AND" | "OR"

[0282] Action ::= Adjustment | Constraint

[0283] Adjustment ::= "ADJUST" Target "BY" Expression ["FOR" Duration]

[0284] DirectAction::= "SET" Target "TO" Expression

[0285] Metric ::= "WIND_POWER" ZoneID | "LOAD" ZoneID | "PRICE" ZoneID | "OUTAGE" GeneratorID | ... / / Predefined metric

[0286] Comparator ::= ">" | "<" | ">=" | "<=" | "==" | "!="

[0287] Value ::= Number | Percentage / / The percentage represents the relative value to the predicted value

[0288] Target ::= "BID" GeneratorID | "BID_ZONE" ZoneID

[0289] Expression ::= Number ["%" | "yuan / MWh"] / / Absolute value or percentage

[0290] Duration ::= "PEAK_HOURS" | "OFF_PEAK_HOURS" | TimeRange

[0291] TimeRange ::= "[" Hour ":" Minute "-" Hour ":" Minute "]" / / Such as "[14:00 - 16:00]"

[0292] ZoneID, GeneratorID ::= String / / "Zone and unit identifiers"

[0293] Semantics: The Condition after IF is converted into a 0 - 1 indicator variable in the policy optimization. The Action after THEN is converted into a constraint of the optimization problem. ADJUST BY X% means increasing or decreasing the percentage based on the benchmark bid. SET TO Y means setting the absolute bid.

[0294] S5.2 Compiler and Constraint Transformation

[0295] The compiler parses the EML statement into an Abstract Syntax Tree (AST), and then traverses the AST to generate mathematical constraints. For example, for the instruction:

[0296] "IF WIND_POWER ZONE_A < 70% AND PRICE ZONE_A > 500 THEN ADJUST BID G1 BY +8% FOR PEAK_HOURS"

[0297] The compiler performs the following transformations:

[0298] Condition variable transformation: Introduce binary variables for each basic condition .

[0299] Let be the time period The actual wind power output ratio of region A is its predicted value.

[0300] Define binary variables , making This is achieved through the Big M method:

[0301]

[0302]

[0303] in It is a very large positive number. It is a very small positive number.

[0304] Similarly, define a binary variable for the condition PRICE ZONE_A > 500. .

[0305] Compound logic: AND corresponds to logical AND. Definition In mixed-integer linear programming (MILP), this is equivalent to the constraint:

[0306]

[0307] Action constraint: Command G1 during the time period The pricing variable is the initial benchmark price given for optimization. The instruction requires "increase by 8%".

[0308] This is transformed into a conditional constraint: if and Belongs to the peak time period set ,So .

[0309] Linearize it using the Big M method:

[0310]

[0311] Output: The compiler ultimately outputs a set of linear equality and inequality constraints. , as a newly added constraint set.

[0312] S5.3 Intent Verification and Conflict Resolution

[0313] Satisfactionability verification: With inherent constraint set merge. Including power flow equations and line capacity Unit climbing slope wait. Including upper and lower price limits etc. Use the computeIIS() function of a mathematical programming solver (such as Gurobi) or an SMT solver (such as Z3) to check. Satisfactionability.

[0314] Conflict resolution optimization: If Infeasible; activate the minimum correction solver. Assume the original instruction corresponds to... Include Each constraint is subject to a slack variable. (Depending on the constraint type, (can be non-negative or free variables), relax them as (or ), forming a relaxed constraint set ,in .

[0315] Optimization problem:

[0316]

[0317] in Promote sparsity (minimize the need to modify constraints). Ensure that the changes are gradual. As weight.

[0318] Solution and Interpretation: Solving this problem yields the optimal decision variables (final strategy) and slack variables. Analysis The component with the largest absolute value is identified as the core constraint causing the conflict. Combining the analysis of the dual variables or infeasible conflict set (IIS) of the optimization problem, the explanation is generated: "The ADJUST BID G1 BY +8% instruction, under peak scenario S7, conflicts with the transmission capacity limit (800 MW) of line L123 (AB), resulting in an expected power flow of 815 MW for this line. It is recommended to modify the upward adjustment to 5%."

[0319] Step S6: Specific Implementation of Closed-Loop Evolution

[0320] On the trading day Once the market clearing process is complete, the system will initiate an offline learning and update process.

[0321] Data Collection and Alignment: Collection Actual data for the day:

[0322] Actual multi-source data (meteorological, load, topology status).

[0323] Actual physical measurement (PMU phasor).

[0324] Actual market clearing results: , .

[0325] Actual regulatory results: Whether there were any warnings, price interventions, or other incidents. .

[0326] Generate training signals and update parameters:

[0327] Update the S1 fusion encoder: Calculate the physical consistency loss Using this as loss, the parameters of PhysFusionEncoder are incrementally learned online with a small number of iterations (e.g., 1 epoch) to adapt them to the latest system operating characteristics.

[0328] Update S3 Behavior Recognition HMM: Use the latest policy fingerprint The observation probability parameters are fine-tuned through Bayesian updates or sliding window re-estimation (rerunning the Baum-Welch algorithm every 30 days). This makes state recognition more adaptable to the current opponent's game characteristics.

[0329] Update S2 policy network and virtual supervisor: The actual daily quote, actual profit, and whether regulatory action was triggered. Constitute an empirical tuple The data is stored in the experience replay buffer. This data is then used periodically for offline reinforcement learning training to update the policy network. parameters And the penalty function parameters of the virtual supervisor To approximate real market feedback.

[0330] Update the S4 Cox risk model: Each day is added as a new observation sample (or right-censored data if there are no regular mutations) to the survival analysis dataset. The Cox model is refitted quarterly or semi-annually using the accumulated new data, and the coefficients are updated. and benchmark risk This allows risk prediction to keep pace with the times.

[0331] Meta-learning coordination: Periodically (e.g., monthly) evaluate the performance of the overall strategy after module updates in recent historical backtesting (e.g., Sharpe ratio, maximum drawdown, compliance rate). Based on this evaluation, a lightweight meta-learning loop automatically adjusts some key hyperparameters, such as the weights of the loss function in S1. Compliance trade-off coefficients in S2 Relaxation weights for conflict resolution in S5 In order to achieve optimal long-term overall performance.

[0332] (III) Work Process

[0333] The following detailed operation of power generation group G on trading day D (Thursday, October 26, 2023) fully demonstrates the implementation process of the method of the present invention.

[0334] Time: Day D, 9:00 AM

[0335] Trigger S1: PhysFusionNet to start automatically.

[0336] Input: Data up to 9:00 AM: Grid data of wind speed / irradiance for the next 48 hours released by the National Meteorological Center (processed by CNN); load forecast curves for each node for the next 24 hours (processed by LSTM); and the day-ahead network topology issued by the dispatch center (processed by GNN), which includes a critical line L123 maintenance plan (out of service from 11:00 to 15:00).

[0337] Processing and Output: The pre-trained PhysFusionEncoder and DifferentiableDCPF layer jointly infer to output a set of 96 physically consistent prediction scenarios (15-minute intervals) for day D+1 (October 27th). Key Output Example: In one high wind power fluctuation scenario, it is predicted that at 14:00 (peak hour), the wind power output in region A will only be 65% of the predicted value. At this time, line L123 has been decommissioned, but the power flow is close to its limit.

[0338] Time: Day D, 10:00 AM

[0339] Trigger S3: The dynamic behavior recognition module starts.

[0340] Input: Read the day-ahead market clearing data for the past 10 trading days (October 16 - October 25).

[0341] Processing: Calculate the residual between the daily baseline clearing (based on the public SRMC) and the actual clearing, and extract the policy fingerprint sequence. Run online HMM filtering.

[0342] Output: The system determines that the current market is in a certain state. (Investigative Strategy Period), Confidence Level 72%. Under this condition, historical data shows that major online retailers tend to quote prices 15%-25% higher than costs during peak hours in similar weather conditions.

[0343] Time: Day D, 10:15 AM

[0344] Trigger S4: Risk Quantification Module Starts.

[0345] Input: Current covariates (Market power index is slightly high, volatility is moderate, policy sentiment is negative, and fuel supply is tight.)

[0346] Processing: Substitute the trained Cox model into the calculation of the survival probability.

[0347] Output: Stable probability curve for the next 7 days. The results show the survival probability on day D+3 (Monday, October 29th). This indicates a high level of risk (possibly related to the policy assessment meeting over the weekend).

[0348] Time: Day D, 10:30 AM to 1:30 PM

[0349] Trigger S2: Strategy optimization engine runs.

[0350] Input: Scenario set from S1, opponent behavior characteristics from S2 (probing period, +20% tendency), and risk adjustment factor from S4 ( ).

[0351] Processing: Policy network within the Stackelberg game framework Conduct large-scale parallel simulation training to maximize risk-adjusted expected returns and avoid regulatory penalties.

[0352] Output: Generate a preliminary pricing strategy draft for day D+1. Example: For Group G's gas turbine unit G1 in Region A, the draft suggests a bid of 520 yuan / MWh between 14:00 and 15:00 (approximately 37% higher than its SRMC of 380 yuan / MWh, but below the threshold for triggering regulatory action).

[0353] Time: Day D, 2:00 PM

[0354] Human-computer interaction: Traders review the draft. Based on experience, he believes the model's prediction of a sharp drop in wind power in the afternoon may still be too optimistic, and hopes to add a defensive order.

[0355] Trigger S5: Traders generate EML instructions through a graphical interface:

[0356] "IF WIND_POWER ZONE_A < 75% AND TIME IN [13:30-15:30] THEN ADJUST BIDG1 BY +8%"

[0357] Compilation and verification: The compiler translates instructions into mathematical constraints. After merging the inherent constraints, the solver found it infeasible. Conflict analysis showed that in the high wind power fluctuation scenario generated by S1 (wind power <65%), if the G1 bid increases by another 8%, the power flow of line L123 will exceed its limit (805 MW > 800 MW).

[0358] Conflict resolution: The system initiates minimum correction optimization, and the solution is obtained. (Corresponding to a 3% adjustment).

[0359] Output and Confirmation: The system pops up a message to the trader: "Instruction conflicts with physical safety. It has been detected that your requested +8% adjustment will overload line L123 in scenarios where wind power output is below 75%. We recommend changing it to ADJUST BID G1 BY +5%. Do you accept?" After reviewing the detailed analysis, the trader clicks "Accept Correction."

[0360] Time: Day D, 2:10 PM

[0361] Final strategy generation: The system injects the revised constraints (ADJUST ... BY +5%) into the optimization model, quickly re-solves, and obtains the final pricing strategy. For unit G1, the final price at 14:00 was determined to be 535 yuan / MWh (that is, based on the draft of 520 yuan / MWh, a human correction of +5% after conflict resolution was added, which is actually about +2.9%).

[0362] Time: Day D, 3:00 PM

[0363] Strategy Submission: The system automatically submits the final price curves for all generating units of Group G on day D+1 through the standard API of the power trading platform.

[0364] Date: D+1 (October 27), after clearing

[0365] Trigger S6:

[0366] Actual data shows that on the afternoon of D+1, the actual output of wind power in area A was 68% of the predicted value, the maximum power flow of line L123 was 798 MW, the market clearing price was 538 yuan / MWh, the G1 unit won the bid, and no regulatory action was triggered.

[0367] The system automatically calculates the errors of each module: S1 scenario prediction error, S3 state judgment accuracy, S2 profit prediction deviation, etc.

[0368] Initiate closed-loop update: fine-tune PhysFusionEncoder using actual physical measurements of the day; add new game data points to the HMM training set; store the decision and result of this game as experience in the buffer for offline learning by the policy network.

[0369] This concludes a complete intelligent pricing decision-making cycle. This embodiment details how the present invention organically integrates multi-source data fusion, embedding of physical laws, dynamic game perception, risk quantification, human-machine collaboration, and closed-loop learning to form a continuously evolving, robust, and reliable intelligent decision-making system. Those skilled in the art can implement the method described in this invention based on the above description and the disclosed software framework.

[0370] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent pricing and decision support in electricity spot trading, characterized in that: Includes the following steps: Step S1, physical-data fusion sensing, constructing a differentiable power flow calculation layer based on the implicit function theorem, and jointly training a fusion encoder; the fusion encoder takes multi-source heterogeneous data as input, and the physical quantity obtained by the output of the fusion encoder after being calculated by the differentiable power flow calculation layer, the difference between the physical quantity and the measured physical quantity constitutes the physical consistency loss, which is used to drive the training of the fusion encoder, thereby outputting physically consistent system situation information; Step S2: Intrinsic security and compliance strategy optimization. Construct a parameterized virtual mechanism designer that simulates the regulatory response of market operators, and form a Stackelberg game framework together with the bidding agent. By solving this two-layer optimization problem, train the bidding agent to output a bidding strategy that balances expected benefits and expected regulatory penalties. Step S3: Dynamic opponent behavior identification, calculate the residual sequence between the actual market clearing result and the benchmark prediction result based on public cost information; extract features from the residual sequence to obtain a strategy feature vector representing the market game style; Based on the strategy feature vector sequence, the estimation of the current market game state is updated online using a probabilistic graphical model; Step S4: Systemic risk quantification. Based on historical rule mutation events and their associated covariates, a survival analysis model is constructed to assess the risk of rule mutation. According to the future rule stability probability output by the survival analysis model, a risk adjustment factor is calculated, and the discount factor in the multi-period decision value function is adjusted using the risk adjustment factor. Step S5, human-machine intent reconciliation, compiles the policy instructions defined by human experts in a formal language into a set of mathematical constraints that can be embedded in the optimization model; verifies the satisfiability of the set of mathematical constraints and the inherent physical and rule constraints, and solves the minimization correction problem when conflicts occur to generate a feasible and semantically close final strategy. Step S6, closed-loop evolution: Based on the difference between the execution results of steps S1 to S5 and the subsequent actual market results, a training signal is generated, and at least one of the following is updated: the fusion encoder in step S1, the pricing agent and the virtual mechanism designer in step S2, the probabilistic graphical model in step S3, and the survival analysis model in step S4.

2. The method according to claim 1, characterized in that, Step S1 includes the following steps: S1.1 Construct a differentiable power flow calculation layer, encapsulating the traditional power flow numerical solver into a differentiable calculation module. Its differentiability is achieved in the following way: In the forward propagation stage, the traditional numerical solver is called to obtain a high-precision solution of the system state variables; in the backward propagation stage, the iterative process inside the solver is not traced back, but the gradient is calculated by solving the corresponding linear equations based on the linearized model of the system at the solution obtained from the forward propagation, thereby realizing the differentiable embedding of physical laws. S1.2, Construct a physical consistency-driven fusion encoder. Design a deep neural network as the fusion encoder. Its architecture is configured according to the input data type and is used to process spatiotemporal sequence and topology data. The output of the fusion encoder is the predicted value of the injected power and topology parameters of the power grid nodes. This predicted value is directly fed to the differentiable power flow calculation layer as its input. S1.3, implement end-to-end joint training, using the previously constructed differentiable power flow calculation layer as a physical constraint bridge to calculate the physical consistency loss; the physical consistency loss is the difference norm between the predicted physical quantity output by the differentiable power flow calculation layer and the measured physical quantity from the measurement system; the total loss function of the fusion encoder is composed of the weighted sum of the data reconstruction loss and the physical consistency loss, and through the backpropagation of the total loss function, the parameters of the fusion encoder are optimized so that the fused situation information simultaneously satisfies the dual constraints of data fidelity and physical laws.

3. The method according to claim 2, characterized in that, In step S1.3, the total loss function of the fusion encoder Loss from data reconstruction With the aforementioned physical consistency loss The weighted summation is as follows: ,in, and For hyperparameters; The data reconstruction loss It is a loss function based on the difference between the input and output of the fusion encoder; The physical consistency loss Let be the difference norm.

4. The method according to claim 1, characterized in that, Step S2 includes the following steps: S2.1, Construct a virtual market mechanism designer and establish a parameterized regulatory response function to simulate the regulatory behavior of market operators on pricing strategies; the regulatory response function calculates market power monitoring indicators based on the input pricing strategy and maps them to the expected intensity of regulatory penalties; S2.2, Establish a Stackelberg game optimization framework, with the bidding agent as the leader and the virtual market mechanism designer as the follower, and construct a two-layer optimization model; wherein, the upper-layer optimization objective is the comprehensive return of the bidding agent, which is the difference between the expected market return and the expected regulatory penalty intensity adjusted by weights; the lower-layer optimization is that, given a bidding strategy, the virtual market mechanism designer calculates the expected regulatory penalty intensity based on the regulatory response function; S2.3, Solve the incentive-compatible strategy by transforming the optimal response problem of the lower-level virtual market mechanism designer into its first-order optimality condition, and embedding the first-order optimality condition as an equilibrium constraint into the upper-level optimization problem, thereby transforming the two-level optimization problem into a single-level mathematical programming problem for solution, and obtaining the incentive-compatible bidding strategy.

5. The method according to claim 1, characterized in that, Step S3 includes the following steps: S3.1 Calculate the clearing residual sequence. Based on the publicly available cost information for each trading period, calculate the benchmark predicted clearing result. Subtract the actual market clearing result from the benchmark predicted clearing result to obtain the clearing residual sequence that characterizes the collective strategic behavior of the counterparties. S3.2, Extracting the strategy feature vector: Feature extraction is performed on the clearing residual sequence; the feature extraction includes: calculating higher-order statistics of the sequence, the higher-order statistics including at least skewness and kurtosis; and / or performing time-frequency domain analysis on the sequence to extract its frequency domain pattern features; combining the extracted features into a low-dimensional strategy feature vector; S3.3, Online identification of game states, construction of hidden Markov models with different market game stages as hidden states; using the continuously acquired policy feature vectors as observation sequences, online updating of the probability distribution estimate of the current hidden state using Bayesian filtering algorithm, to complete dynamic opponent behavior identification.

6. The method according to claim 1, characterized in that, Step S4 includes the following steps: S4.1, Construct a rule-change risk model. Based on the occurrence time of historical market rule-change events and multi-dimensional covariates prior to the events, a Cox proportional hazards model is used for modeling. The Cox proportional hazards model is in the form of... ,in Indicates that given covariates Next moment Instantaneous risk rate As the benchmark risk function, This is a vector of covariate coefficients; S4.2, Calculate the survival probability and risk adjustment factor based on the Cox proportional hazards model and the current covariates. Calculate future moments The rule-stable probability, i.e., the survival function. The survival function is applied in a discrete decision cycle. The value at the location It is directly defined as the risk adjustment factor for this cycle; S4.3, Execute risk-adjusted decisions: In multi-period pricing strategy optimization, construct the risk-adjusted decision value function. ,in The time discount factor, For the first Expected returns for each cycle; based on the value function Optimize the strategy.

7. The method according to claim 1, characterized in that, Step S5 includes the following steps: S5.1 defines and parses strategy instructions, providing a predefined, syntactically rigorous formal specification language for receiving strategy logic defined by traders; the structured statements of the formal specification language can clearly express market conditions, target assets, and specific adjustment actions; S5.2, Compilation and Constraint Transformation: The compiler parses the structured statements and transforms them into one or more deterministic mathematical constraints that can be embedded in the mathematical optimization model, thus forming a new constraint set. S5.3, Verification and Intent Reconciliation: The newly added constraint set is merged with the inherent physical and rule constraint set, and satisfiability is determined using the constraint solver. If it is determined that the constraint can be satisfied, then the newly added constraint set is injected into the optimization model; If it is determined to be unsatisfactory, the conflict resolution optimization process is initiated: an optimization problem is constructed with the goal of minimizing the semantic deviation of the structured statement, and the problem is solved under the condition of satisfying all inherent constraints. A corrected feasible strategy is output, and explanatory information describing the reasons for the deviation is generated.

8. The method according to claim 7, characterized in that, In step S5.3, the conflict resolution optimization process is specifically implemented in the following ways: For each mathematical constraint in the newly added constraint set, a corresponding slack variable is introduced, and all slack variables constitute the slack variable vector δ. Each mathematical constraint in the newly added constraint set is rewritten as a relaxed constraint containing its corresponding relaxed variable, and all relaxed constraints constitute a relaxed constraint set. Construct and solve the following mathematical optimization problem: Minimize: the norm |δ| of the slack variable vector δ. Constraints: The inherent physical and rule constraint set must be satisfied, and the relaxed constraint set must also be satisfied. The values ​​of the decision variables obtained from the solution are used as the modified feasible strategies.

9. The method according to claim 6, characterized in that, In step S4.1, the multi-dimensional covariates upon which the Cox proportional hazards model is constructed are comprehensive indicators derived by integrating real-time market operation data, macroeconomic information, and public opinion data; specifically, they include at least two of the following categories of indicators: Market structure and behavior indicators, including real-time market power indices that reflect market concentration and manipulation potential; Market volatility indicators, including market price volatility, which reflects short-term price uncertainty; Policy and public opinion indicators, including policy intervention sentiment trend indicators obtained through sentiment analysis of relevant policy announcements and news texts; Fundamental supply indicators include fuel supply tightness indicators that reflect the availability and cost pressures of fuels for power generation.

10. The method according to any one of claims 1-9, characterized in that, The method is implemented according to the following workflow over a complete trading decision cycle: On trading day D, step S1 is executed first, based on the latest multi-source heterogeneous data up to that day, physical-data fusion sensing is performed, and physical consistency system status information for the next trading day D+1 is output. Subsequently, based on the physical consistency system situation information and recent historical market data, step S3 is executed to dynamically identify the current market game state, and at the same time, based on the current macroeconomic environment data, step S4 is executed to quantify the risk of future rule mutations. Next, based on the physical consistency system situation information, and incorporating the market game state identified in step S3 and the risk information quantified in step S4, step S2 is executed to generate a preliminary pricing strategy for trading day D+1. After receiving the formal strategy instructions input by the trader based on the review of the preliminary pricing strategy, step S5 is executed to compile and verify the formal strategy instructions, reconcile them with the preliminary pricing strategy, generate the final pricing strategy, and submit it. After the market clearing is completed on trading day D+1, the actual clearing result is obtained, and step S6 is executed. Based on the difference between the actual clearing result and the output or decision result of the corresponding step in steps S1 to S5, a training signal is generated to update the parameters of one or more models involved in the method.

Images