Load-side day-ahead declaration decision method based on scenario memory enhancement and convex hull extrapolation

Abstract

The application discloses a load-side day-ahead declaration decision-making method based on scenario memory enhancement and convex hull extrapolation and belongs to the technical field of power market transaction decision-making.The application comprises the following steps: obtaining an information set before a declaration deadline and constructing a time sequence characteristic tensor;extracting global time domain structure and local dynamic characteristics by using a double-channel time sequence encoder and fusing the output;constructing a retrieval clue based on the fused output sequence,obtaining an iterative termination trajectory through mode separation,double-key retrieval and iterative completion,obtaining a weighted gate trajectory through consistency gating;decoupling the fused output sequence and the weighted gate trajectory into a shape component and a level component,correcting the level component through a learnable extrapolation item when an estimated level value exceeds a historical memory range,projecting to a declaration price feasible region to obtain a decision-making offer sequence;using the decision-making offer sequence as a declaration price,constructing a coupled loss function by using a differentiable approximation to perform end-to-end training.The application realizes end-to-end maximization of settlement income in load-side day-ahead declaration decision-making.

Description

Technical Field

[0001] This invention belongs to the field of power market transaction decision technology, and specifically relates to a load-side day-ahead reporting decision method based on contextual memory enhancement and convex hull extrapolation. Background Technology

[0002] In the process of electricity market liberalization, the dual settlement system, centered on the day-ahead and real-time markets, has become the fundamental institution for electricity price formation, influencing how the value of electricity is determined. In the electricity spot market, electricity retailers and wholesale users must... Before 15:00 on the same day, bids for the next 24 hours based on publicly available information must be submitted: bids higher than the previous day's clearing price are successful, and settlement is based on the previous day's price; deviations are settled at the real-time price. Since the full entry of renewable energy into the market, the normalization of negative electricity prices and the dramatic fluctuations in price spreads have simultaneously amplified the risks and arbitrage opportunities of this bid decision. This bid decision differs from typical prediction or optimization problems: whether a bid is successful depends on whether the bid price exceeds the unknown clearing price; profit and loss exhibit threshold-type discontinuity in relation to decision variables; and the rare extreme price spread conditions that determine the main part of the revenue often exceed the historical observation range, rendering methods relying on historical distribution interpolation in principle ineffective.

[0003] The existing technologies employ several approaches. The first approach is a two-stage method from prediction to declaration. This method trains the prediction model according to a statistical accuracy target and then predicts the consumption results afterward based on decision rules. However, settlement revenue is a non-linear function of prediction error; optimal average accuracy does not imply optimal revenue, resulting in a mismatch between prediction and decision at the objective level. The second approach is reinforcement learning, which directly models declaration as maximizing sequential revenue. However, this black-box strategy lacks interpretability, has low efficiency in trial-and-error exploration, and weak generalization ability to out-of-distribution extreme conditions, which precisely determine the main part of the revenue. The third approach is a parametric prediction model, which aims to fit the conditional distribution. This model exhibits a tendency towards mean regression at the edges of the training distribution and systematically distorts during peak and negative price periods when settlement consequences are most severe. The output of the retrieval-based (similar day) method is always a convex combination of historical samples. When the penetration rate of new energy increases and forms a new price range, its output is constrained by the historical convex hull and thus fails in principle. The fourth approach is the mainstream sequence modeling architecture, which neither distinguishes between the daily global structure and intraday multi-scale dynamics, nor fully utilizes... The publicly available boundary condition information was used as evidence for the search.

[0004] In summary, there is an urgent need for a load-side day-ahead reporting decision-making method that can overcome the shortcomings of the existing technologies. This method should integrate the daily global structure and intraday local dynamics through dual-path time-series coding, achieve explicit reuse of rare extreme operating conditions through scenario memory retrieval, and break through historical convex hull constraints to cover new price ranges formed by the increase in the penetration rate of new energy sources. This would maximize settlement revenue in the day-ahead reporting decision-making process. Summary of the Invention

[0005] To address the aforementioned technical issues, this invention proposes a load-side pre-reporting decision-making method based on contextual memory enhancement and convex hull extrapolation. This method achieves end-to-end maximization of settlement revenue in load-side pre-reporting decisions and significantly improves decision generalization performance under extreme operating conditions through contextual memory retrieval and convex hull extrapolation mechanisms.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention proposes a load-side pre-reporting decision-making method based on contextual memory enhancement and convex hull extrapolation, comprising the following steps: Obtain the information set before the application deadline, and construct a time series feature tensor based on the information set; the information set includes the time history boundary condition information of the operating day, meteorological environment information related to the load-side output, lagged statistics and lagged price information of historical market operation data, and time location coding information; The temporal feature tensor is encoded using a dual-path temporal encoder, which includes a long-range path for extracting global temporal structure and a local path for extracting local dynamic features. The long-range path and the local path are fused to obtain a fused output sequence. Based on the fusion output sequence, a retrieval clue is constructed. After pattern separation of the retrieval clue, a master retrieval key is constructed. Boundary driving variables are extracted to construct a driving key. The master retrieval key and the driving key are used to perform a double-key retrieval to obtain an initial retrieval trajectory. The initial retrieval trajectory is fed back to update the retrieval clue and iterative retrieval is performed to obtain the iteration termination trajectory. The gating coefficient is calculated based on the consistency between the retrieval clue and the iteration termination trajectory to obtain the weighted gating trajectory. The fused output sequence and the weighted gated trajectory are decoupled into morphological components and horizontal components. The morphological component is formed by superimposing the initial morphology of the fused output sequence and the residual of the weighted gated trajectory. The horizontal component is estimated by learning through the time-averaged features of the fused output sequence and anchored to the horizontal average value of the weighted gated trajectory. When the estimated horizontal value exceeds the historical memory range, it is corrected by learning extrapolation terms. The corrected horizontal component is superimposed with the morphological component and projected onto the feasible region of the declared price to obtain the decision-oriented output quotation sequence.

[0007] Furthermore, the method also includes: The decision-type output price sequence is used as the bid price. The winning bid event is approximated by a differentiable sigmoid function with a temperature parameter. The discrete indication of the winning bid event is replaced by a continuous soft winning bid probability. The expected settlement is calculated based on the soft winning probability, the declared electricity volume, and the difference between the day-ahead price and the real-time price at each time point; a coupled loss function is constructed, consisting of a weighted sum of a revenue loss term and a prediction error term, where the revenue loss term is the negative value of the expected settlement revenue, and the prediction error term is the weighted sum of squares of the errors between the declared price and the real-time price. The parameters of dual-path encoding, retrieval, decoupling, and differentiable decision-making are jointly optimized end-to-end under the coupling loss function.

[0008] Furthermore, the long-range path for extracting the global temporal structure is specifically as follows: anchor points are selected from the temporal feature tensor at fixed intervals, the anchor point sequence is used to establish global dependencies through multi-head self-attention calculation, the anchor point output is backfilled to all time points through linear interpolation, and then the long-range representation is output after residual connection and layer normalization.

[0009] Furthermore, local pathways for extracting local dynamic features are specifically as follows: For each time step, attention scores are calculated only between the query vector at that time step and the key vectors of the previous set number of historical time steps, including the time step itself. Negative infinity masks are applied to positions where attention is not allowed, so that the attention weights corresponding to the masked positions are normalized to zero. The calculated attention weights are summed with the value vectors at the corresponding time points to obtain the local representation at that time point.

[0010] Furthermore, retrieval clues are constructed based on the fused output sequence, and pattern separation is performed on the retrieval clues, specifically as follows: Search clues via linear projection After high-dimensional nonlinear activation, a fixed proportion of activation components are retained through a competition mechanism to obtain a sparse separation characterization: ; in, Characterization of sparse separation; This is a contention mask used to retain a fixed percentage of the active value, while setting the remaining components to zero. The linear projection matrix for pattern separation; To retrieve clues; It is a non-linear activation function; This represents element-wise product.

[0011] Furthermore, a primary retrieval key is constructed, and boundary driving variables are extracted to construct a driving key. The primary retrieval key and the driving key are then used to perform a dual-key retrieval to obtain the initial retrieval trajectory, specifically: Based on the sparse separation characterization Generate primary search key The boundary driving quantity is extracted from the time history boundary condition information as the driving key. The similarity between the primary search key and the primary keys of each entry in the search memory, and the similarity between the driving key and the driving keys of each entry in the search memory are calculated separately. These similarities are then weighted and summed to obtain the search score for each entry. ; in, For the first The search score for each item; The inverse temperature of the main bond; For driving key weights; Use as the primary search key; For the first in the memory bank The primary search key for each entry; The dimension of the key vector; For driving keys; For the first in the memory bank The driving key for each entry; To retrieve the weight vector; For the first in the memory bank Each entry stores the historical real-time price trajectory; This is the initial search trajectory; For the first The retrieval weight of memory entries.

[0012] Furthermore, the initial search trajectory is fed back to update the search clues, and the iterative search is performed to obtain the iteration termination trajectory. Based on the consistency between the search clues and the iteration termination trajectory, a gating coefficient is calculated to obtain the weighted gating trajectory, specifically: To fix The recursive iteration of the steps implements pattern completion, superimposing the current search trajectory onto the search clues after linear transformation to update the search clues: ; in, Use as the primary search key; For the feedback projection matrix; For the first The retrieval trajectory of each step; It is a dimensionless normalization constant; For layer normalization; Indicates based on primary search key and drive key Perform a double-key retrieval and output the weighted trajectory; Output the termination trajectory after one iteration. Use a fixed number of steps instead of a convergence criterion to eliminate non-differentiable branches; The gating coefficient is: ; in, This is the gating coefficient, with a value between 0 and 1; It is a multilayer perceptron; It is the sigmoid function; This indicates vector concatenation; To retrieve the projection of the clue in the comparison space; The projection of the iteration termination trajectory into the comparison space; The comparison space projection matrix for retrieval clues; The comparison space projection matrix of the trajectory; Represents element-wise product; To retrieve clues; The trajectory is the termination trajectory of the iteration; The gating coefficients are applied to the iteration termination trajectory to obtain the weighted gating trajectory and its gating coefficients.

[0013] Furthermore, the morphological component is composed of the superposition of the initial morphology of the fused output sequence and the residual of the weighted gated trajectory, specifically expressed as follows: ; in, For fusion output sequence; The total number of moments in the temporal feature tensor. This is a combined operation of pooling and mapping. Based on shape, The gating coefficient, Inject the matrix into the residual. For the termination trajectory of the iteration, To calculate the average water value, To remember shape residuals, It is a vector whose elements are all 1s; The level component is estimated by using the time-averaged features of the fused output sequence to obtain the level value through learnable level estimation, and is jointly given by the level average of the iteration termination trajectory as the recall anchor term, specifically expressed as follows: ; in, For horizontal components, For the learnable level estimation function, To fuse the time-averaged features of the output sequence, For anchoring items of memory; For anchoring weight coefficients, The water average value of the iterative termination trajectory.

[0014] Furthermore, when the estimated level value exceeds the maximum value of the historical trajectory level value in the historical memory bank, a correction is made by proportionally amplifying the anchoring through learnable extrapolation terms, specifically: ; in, For exceeding the limit, To stop the gradient operation, This represents the maximum value of the historical trajectory level in the historical memory bank. The normalization constant is For learnable nonnegative extrapolation magnitude, It is a learnable scalar; It is a smooth positive function; The corrected horizontal component is superimposed on the morphological component to obtain the synthesized trajectory: ; in, The output price sequence is for decision-making purposes. This is an operation that truncates each element to the feasible domain of the declared price.

[0015] Secondly, the present invention also proposes a load-side pre-reporting decision-making device based on contextual memory enhancement and convex hull extrapolation, comprising at least one processor and a memory, wherein the memory stores a computer program, and the computer program, when executed by the at least one processor, implements the load-side pre-reporting decision-making method based on contextual memory enhancement and convex hull extrapolation.

[0016] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. One of the above technical solutions has the following advantages or beneficial effects: I. End-to-end alignment of decision-making and benefit objectives This invention, through the design of a differentiable decision layer and a coupled loss function, directly uses the decision-oriented bid sequence as the model output. It employs a temperature-parameterized sigmoid function to perform a differentiable approximation of the winning bid event, enabling the gradient information of threshold-type winning bid events to propagate back from the expected settlement profit to the encoder and retrieval module. This achieves end-to-end joint optimization from information input to bid output. Compared to existing two-stage "prediction and submission" methods, this invention eliminates the systematic mismatch between statistical accuracy and settlement profit objectives, ensuring that each update of the model parameters directly points to the direction of maximizing settlement profit, thus avoiding the value loss of "minimizing prediction error but not achieving optimal profit."

[0017] II. Reuse of Contextual Memory in Rare and Extreme Working Conditions This invention constructs a differentiable contextual memory retrieval module based on a hippocampal three-synaptic circuit. It explicitly stores complete real-time price trajectories from training samples in key-value pairs within a contextual memory bank. Through sparse pattern separation, retrieval cues for different operating conditions are mapped to approximately orthogonal high-dimensional representations, suppressing crosstalk between similar interfering days. A dual-channel retrieval key design explicitly includes day-ahead boundary drivers (system load, renewable energy output, exchange power, etc.) in the definition of "similarity," rather than relying solely on representational similarity. Recursive iterative completion enables the model to gradually reconstruct complete daily trajectories from incomplete boundary evidence. This mechanism allows for the on-demand reuse of memories from rare extreme operating conditions (such as negative electricity prices and peak electricity prices), significantly improving model performance in high-value operating condition regions where training samples are sparse, and resolving the mean regression tendency of parametric prediction models at distribution edges.

[0018] III. Breaking Through Historical Convex Hull Constraints and Effective Decision-Making for New Price Ranges This invention employs a horizontal-morphological decoupling mechanism to decompose the predicted output into morphological and horizontal components. The morphological component, under zero-mean constraints, uses memory retrieval as a priori, utilizing only the morphological information of historical trajectories. The horizontal component is determined jointly by a learnable level estimate and recall anchoring. When the estimated level exceeds the historical memory range, it is corrected in the same direction by proportionally amplifying the anchoring through a learnable extrapolation term. This mechanism enables the model to form effective bid price positions even when the penetration rate of new energy sources continues to rise and price ranges constantly break historical extremes. It fundamentally overcomes the inherent flaw of retrieval-based similarity day methods, whose output is always a convex combination of historical samples, and solves the technical problem of systematic failure of existing methods under extreme out-of-distribution conditions.

[0019] IV. Strict Timing Alignment and Systematic Elimination of Look-Forward Bias This invention strictly aligns the timing of market information disclosure throughout the training and inference processes: the training set contains only samples from dates no later than the target date, and all retrievable entries during inference come from historical windows earlier than the target date; the contextual memory is reconstructed after each training convergence, the model is updated daily, and feature scaling is based solely on samples within the historical window. These multiple mechanisms collectively eliminate the risk of information leakage from three levels: data acquisition, model training, and inference prediction. This ensures that the strategy backtesting results accurately reflect the decision-making performance during online deployment, providing reliable performance assurance for actual trading decisions. Attached Figure Description

[0020] Figure 1 This is a flowchart of the load-side day-ahead reporting decision-making method based on contextual memory enhancement and convex hull extrapolation proposed in Embodiment 1 of the present invention; Figure 2 This is a framework diagram of the load-side pre-dated reporting decision-making method based on contextual memory enhancement and convex hull extrapolation proposed in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the hippocampal trisynaptic episodic memory architecture proposed in Embodiment 1 of the present invention. Figure 3 (a) in the diagram represents the main process of contextual memory retrieval; Figure 3 (b) in the diagram represents the writing protocol and leakage prevention mechanism of the context memory. Figure 4 This is a schematic diagram of the dual-path timing encoder proposed in Embodiment 1 of the present invention; Figure 5 This is a schematic diagram illustrating the decoupling of horizontal and morphological aspects as proposed in Embodiment 1 of the present invention; Figure 6 This is a schematic diagram of the differentiable reporting decision-making proposed in Embodiment 1 of the present invention. Figure 6 (a) in the diagram represents the main process for differentiable reporting and decision-making; Figure 6 (b) in the diagram is a schematic diagram of the input-output relationship of the differentiable decision layer. Figure 7 This is a schematic diagram of the load-side pre-declaration decision-making device based on contextual memory enhancement and convex hull extrapolation proposed in Embodiment 2 of the present invention. Detailed Implementation

[0021] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure of the invention, components and arrangements of specific examples are described below. Furthermore, reference numerals and / or letters may be repeated in different examples. This repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. It should be noted that the components illustrated in the drawings are not necessarily drawn to scale. Descriptions of well-known components, processing techniques, and processes are omitted in this invention to avoid unnecessarily limiting the invention.

[0022] Example 1 Embodiment 1 of this invention proposes a load-side pre-reporting decision-making method based on scenario memory enhancement and convex hull extrapolation. This method addresses the technical problems of existing load-side pre-reporting decision-making methods, which suffer from non-optimal settlement benefits and insufficient generalization ability for extreme conditions due to mismatch between prediction and decision objectives, lack of memory of rare extreme operating conditions, historical convex hull constraints, and look-ahead bias.

[0023] Figure 1 This is a flowchart of the load-side day-ahead reporting decision-making method based on contextual memory enhancement and convex hull extrapolation proposed in Embodiment 1 of the present invention; Figure 2 This is a framework diagram of the load-side pre-reporting decision-making method based on contextual memory enhancement and convex hull extrapolation proposed in Embodiment 1 of the present invention; combined with Figure 1 and Figure 2 The implementation process of this invention will be explained together.

[0024] In step S1, an information set prior to the application deadline is obtained, and a time-series feature tensor is constructed based on the information set. The information set includes time-series boundary condition information for the operating day, meteorological environment information related to load-side output, lagged statistics and lagged price information of historical market operation data, and time location encoding information. According to the prevailing rules of the electricity spot market, load-side entities will bid on the auction day ( Submit 24-hour time-sharing volume and price reports before 15:00 on the day of the report. This invention strictly limits the model input to legally obtainable information before the report deadline, constructing a feature tensor. ( (where is the feature dimension).

[0025] The information set includes: time-history boundary condition information for the operating day, i.e., the boundary condition forecast for 96 points the next day (including system load forecast, new energy output forecast and inter-regional exchange power forecast). Meteorological and environmental information related to load-side output, namely regional temperature, irradiance, and wind speed forecasts; The lagged statistics and lagged price information of historical market operation data, namely the lagged multi-day series of day-ahead and real-time prices and their rolling window statistics (7 / 14 / 30-day windows, all up to...). ); And time location encoding information, namely the sine / cosine cycle encoding of intraday time periods.

[0026] Target date clearing results ( The price to be released at approximately 17:00 on the day of the announcement will not be included in the model input; the real-time price on the target day will only be used as a price spread weighted monitoring signal.

[0027] The above information together constitutes the temporal feature tensor, which serves as the input to the dual-path temporal encoder.

[0028] In step S2, a dual-path temporal encoder is used to encode the temporal feature tensor. The dual-path temporal encoder includes a long-range path for extracting the global temporal structure and a local path for extracting local dynamic features. The long-range path and the local path are fused to obtain a fused output sequence.

[0029] Boundary condition forecasts simultaneously encompass daily global structures (such as global equilibrium patterns and peak-valley configurations) and intraday multi-scale local dynamics (such as minute-level fluctuations in renewable energy output). A single homogeneous attention stream cannot simultaneously represent both types of structures. To address this, this invention constructs a dual-path encoder. Figure 4 This is a schematic diagram of the dual-path temporal encoder proposed in Embodiment 1 of the present invention; the long-range path extracts the global temporal structure through sparse sampling and attention mechanisms, specifically as follows: Let the number of time steps in the time series feature tensor be... (In the embodiments of the present invention) (corresponding to 96 moments at a 15-minute resolution), the number of anchor points is , For fixed intervals, the anchor point index set is: Then anchor sequence attention Represented as: ; in, The query matrix representing the anchor point; The key matrix representing the anchor point; The value matrix representing the anchor points; Indicates the attention head dimension; Anchor point outputs are backfilled to all time points using deterministic linear interpolation in the anchor point index space: for time points... Take the adjacent anchor index: Interpolation weights: Backfilled via linear interpolation: ; in, Indicates time The long-range path output vector; Indicates the first Attention output at each anchor point; Indicates the first The output vector after multi-head self-attention computation; After residual connection and layer normalization, a long-range representation is output. Sparse anchors handle global information exchange, and interpolation backfilling maintains temporal resolution. The computational cost of attention in this pathway is reduced by... Down to .

[0030] Local pathways extract local dynamic features through a temporally constrained attention mechanism, specifically: For time Only allow attention to be directed in front of it. At any given moment, and for itself, local attention calculations are performed using a causal mask: ; in, Indicates time The local path output vector; Indicates time The query vector; Indicates the preceding The key matrix from the previous time step to the current time step; Indicates the preceding A matrix of values ​​from each time point to the current time point; Represents a mask matrix; elements in the mask matrix Satisfy: When hour otherwise The local representation is output after attention calculation.

[0031] The global temporal structure and local dynamic features are fused through a gating mechanism at each time step to obtain the fused output sequence, specifically: After aligning the long-range representation and the local representation through a shared linear layer, they are adaptively fused by gating at each time step. The gating is calculated based on the long-range representation and the local representation at the current time step. The fused output is then processed by residual connection and layer normalization to serve as the fused output sequence.

[0032] ; ; in, Indicates time The gate weight vector; Indicates the activation function; Indicates time The fusion output; Representation layer normalization.

[0033] The residual term is used to balance the relative contributions of global structure and local dynamics, ensuring the complete preservation of information from both pathways. The fused output at each time step... The fused output sequence is used for subsequent retrieval and synthesis steps.

[0034] In step S3, a retrieval clue is constructed based on the fused output sequence. After pattern separation of the retrieval clue, a master retrieval key is constructed, and a driving key is constructed by extracting the boundary driving quantity. The master retrieval key and the driving key are used to perform a double-key retrieval to obtain an initial retrieval trajectory. The initial retrieval trajectory is fed back to update the retrieval clue and iterative retrieval is performed to obtain the iteration termination trajectory. The gating coefficient is calculated based on the consistency between the retrieval clue and the iteration termination trajectory to obtain the weighted gating trajectory. Parametric predictions tend to regress to the mean at the edges of the training distribution, and similar-day retrieval is constrained by the convex hull and lacks explicit processing of retrieval evidence. Inspired by the hippocampal trisynaptic circuit, this invention constructs a "separation, completion, and verification" retrieval mechanism. Figure 3 This is a schematic diagram of the hippocampal three-synaptic episodic memory architecture proposed in Embodiment 1 of the present invention; wherein, Figure 3 (a) in the text is the main process of context memory retrieval, which includes three core steps: DG pattern separation, CA3 double-key retrieval and recursive completion, and CA1 consistency gating. Figure 3 (b) in the text describes the writing protocol and leakage prevention mechanism of the context memory, including daily single write, leave-one mask, snapshot recovery and rolling reconstruction and other safeguards.

[0035] like Figure 3 As shown in (a), in the DG mode separation stage upstream of the main process, the daily search clues are first subjected to high-dimensional linear expansion and nonlinear activation, and then a fixed proportion of activation components are retained through the k-WTA sparse competition mechanism. Similar but different interfering days are mapped into approximately orthogonal representations in high-dimensional space, thereby suppressing search crosstalk, outputting sparse separation representations, and generating the main search key based on this.

[0036] In the CA3 dual-key retrieval and recursive completion process, the context memory is implemented as a key-value buffer, storing the complete real-time price trajectory of historical days. The retrieval is divided into two channels: the primary retrieval key and the driving key. The similarity between these channels and the corresponding keys of each entry in the memory is calculated, and after softmax normalization and weighting, an initial recalled trajectory is obtained. Subsequently, a fixed-step recursive iteration is used to complete the trajectory, feeding back the current recalled trajectory as new retrieval clues. From incomplete boundary evidence, the complete daily trajectory is gradually reconstructed, and the recalled trajectory (24-point real-time price pattern) is output.

[0037] In the CA1 consistency gating process, a comparator is set up to calculate the gating coefficient based on the consistency between the retrieval clue and the recall trajectory: when the retrieval result matches the current working condition, it is allowed to proceed, and the recall trajectory is passed as the gating memory shape prior to the downstream decoupled into a module; when the retrieval result does not match the current working condition, the memory channel is closed, and it degenerates into pure parameterized prediction.

[0038] like Figure 3 As shown in (b), the contextual memory database employs a strict writing protocol and a triple guarantee against leakage: daily single-write ensures only one record is written per training day; a leave-one-out mask is applied to each sample during training, preventing any training sample from retrieving itself; key space consistency is guaranteed by snapshot recovery; cross-round drift is mitigated by daily rolling retraining, with memories being eliminated as they are rebuilt from an empty database each round; and read-only access is allowed during the evaluation and prediction periods. These mechanisms collectively ensure that there is no information leakage during the entire training and inference process of contextual memory retrieval.

[0039] Specifically: (1) DG mode separation: separating search clues via linear projection After high-dimensional nonlinear activation, a fixed proportion of activation components are retained through a competition mechanism to obtain a sparse separation characterization: ; in, Characterization of sparse separation; This is a contention mask used to retain a fixed percentage of the active value, while setting the remaining components to zero. The linear projection matrix for pattern separation; To retrieve clues; It is a non-linear activation function; This represents element-wise product.

[0040] Construct the primary retrieval key and extract the boundary driving variables to construct the driving key. Use the primary retrieval key and the driving key to perform a two-key retrieval to obtain the initial retrieval trajectory, specifically as follows: Based on the sparse separation characterization Generate primary search key The boundary driving quantity is extracted from the time history boundary condition information as the driving key. The similarity between the primary search key and the primary keys of each entry in the search memory, and the similarity between the driving key and the driving keys of each entry in the search memory are calculated separately. These similarities are then weighted and summed to obtain the search score for each entry. ; in, For the first The search score for each item; The inverse temperature of the main bond; For driving key weights; Use as the primary search key; For the first in the memory bank The primary search key for each entry; The dimension of the key vector; For driving keys; For the first in the memory bank The driving key for each entry; To retrieve the weight vector; For the first in the memory bank Each entry stores the historical real-time price trajectory; This is the initial search trajectory; For the first The retrieval weight of memory entries.

[0041] The driving key channel makes the definition of "similar" explicitly include price drivers (system load, renewable energy output, etc.), rather than relying solely on representational similarity.

[0042] The initial search trajectory is fed back to update the search clues, and the search is iterated to obtain the iteration termination trajectory. The gating coefficient is calculated based on the consistency between the search clues and the iteration termination trajectory to obtain the weighted gating trajectory, specifically: To fix The recursive iteration of the steps implements pattern completion, superimposing the current search trajectory onto the search clues after linear transformation to update the search clues: ; in, Use as the primary search key; For the feedback projection matrix; For the first The retrieval trajectory of each step; It is a dimensionless normalization constant; For layer normalization; Indicates based on primary search key and drive key Perform a double-key retrieval and output the weighted trajectory; Output the termination trajectory after one iteration. Use a fixed number of steps instead of a convergence criterion to eliminate non-differentiable branches; To measure the degree of match between the current operating condition and the memory, a comparison gating is constructed: ; in, This is the gating coefficient, with a value between 0 and 1; It is a multilayer perceptron; It is the sigmoid function; This indicates vector concatenation; To retrieve the projection of the clue in the comparison space; The projection of the iteration termination trajectory into the comparison space; The comparison space projection matrix for retrieval clues; The comparison space projection matrix of the trajectory; Represents element-wise product; To retrieve clues; The trajectory is the termination trajectory of the iteration; scalar The intensity of memory calls to downstream synthesis is controlled: when the retrieval results are inconsistent with the current operating conditions, the gating is closed, degenerating into pure parametric prediction. The gating coefficients are applied to the iterative termination trajectory to obtain the weighted gating trajectory and its gating coefficients.

[0043] In step S4, the fused output sequence and the weighted gated trajectory are decoupled into a morphological component and a horizontal component. The morphological component is formed by superimposing the initial morphology of the fused output sequence and the residual of the weighted gated trajectory. The horizontal component is estimated by learning through the time-averaged features of the fused output sequence and anchored to the horizontal average value of the weighted gated trajectory. When the estimated horizontal value exceeds the historical memory range, it is corrected by learning extrapolation. The corrected horizontal component is superimposed with the morphological component and projected onto the feasible region of the declared price to obtain the decision-type output quotation sequence.

[0044] This invention decouples prediction into two channels: horizontal and morphological. The morphological component uses memory retrieval as a prior under zero-mean constraints, while the horizontal component is carried by parameterized states and learnable extrapolation terms, thereby breaking through the historical convex hull constraint. Figure 5 This is a schematic diagram of the horizontal and morphological decoupling proposed in Embodiment 1 of the present invention.

[0045] The prediction head first gives a basic shape from the fused sequence, i.e. To achieve a composite mapping between the fusion sequence and the prediction head, using 15-minute to hourly mean pooling, a 24-point hourly trajectory is output, with zero-mean constraints and superimposed gated memory shape residuals applied. ; in, For fusion output sequence; The total number of moments in the temporal feature tensor. This is a combined operation of pooling and mapping. Based on shape, The gating coefficient, Inject the matrix into the residual. For the termination trajectory of the iteration, To calculate the average water value, To remember shape residuals, It is a vector with all elements being 1; that is, the daily level is first subtracted from the retrieval trajectory (retrieval first, then average removal), and only pure morphological information is injected into the prediction; Given by CA1 gating, the residual is automatically shut down when the retrieval is unreliable.

[0046] The level component is estimated by using the time-averaged features of the fused output sequence to obtain the level value through learnable level estimation, and is jointly given by the level average of the iteration termination trajectory as the recall anchor term, specifically expressed as follows: ; in, For horizontal components, For the learnable level estimation function, To fuse the time-averaged features of the output sequence, For anchoring items of memory; For anchoring weight coefficients, The water average value of the iterative termination trajectory.

[0047] When the estimated level value exceeds the maximum level value of historical trajectories in the historical memory bank, a correction is made by proportionally amplifying the anchoring through learnable extrapolation terms. Specifically: ; in, For exceeding the limit, To stop the gradient operation, This represents the maximum value of the historical trajectory level in the historical memory bank. The normalization constant is For learnable nonnegative extrapolation magnitude, It is a learnable scalar; It is a smooth positive function; Extrapolation is non-zero only when the level exceeds the upper bound of memory; the correction direction is determined by the anchor term of recall. The sign is determined, and the gradient feedback of the extrapolation path to the estimated level value is blocked. Due to anchoring The level of the new price range is generated by extrapolation of boundary evidence and recollection magnitude, rather than by historical interpolation.

[0048] The output of the retrieval method is a convex combination of historical samples, with a daily level. It will inevitably fall into The range; when the actual level on the target day exceeds this range (such as a new price range formed by the increase in new energy penetration, negative prices, and peak operating conditions), any convex combination retrieval has a lower bound of horizontal deviation that cannot be eliminated by retrieval key optimization. This invention breaks through this constraint by decoupling "form based on memory retrieval as a priori, and level carried by parameterized state and learnable extrapolation terms," ​​and the extrapolation amplitude With non-negativity constrained by softplus and indirectly supervised by coupling loss, there is no cumulative drift channel.

[0049] The corrected horizontal component is superimposed on the morphological component to obtain the synthesized trajectory: ; in, The output price sequence is for decision-making purposes. To ensure that the bid price is truncated to the feasible region on an element-by-element basis, the bid price is always within the feasible region of the rule.

[0050] At this point, the model output directly constitutes the 24-hour bid price sequence for the next day, completing the entire decision-making process from information input to bid price generation.

[0051] Step S5: Differentiable reporting decision and end-to-end training Figure 6 This is a schematic diagram of the differentiable bidding decision proposed in Embodiment 1 of the present invention; the present invention transforms the threshold-type winning bid event into a differentiable soft decision-making process, enabling all parameters to be jointly optimized end-to-end under the settlement benefit objective. Among them, Figure 6 (a) in the text represents the main process of differentiable declaration decision-making, which includes four stages: decision-oriented quotation generation, bid-winning determination softening, expected return and coupling training, and end-to-end feedback of settlement error. Figure 6 (b) in the diagram is a schematic diagram of the input-output relationship of the differentiable decision layer.

[0052] like Figure 6 As shown in (a) above, the main process for the decision-making of micro-reporting is as follows: Decision-based price generation: The final output of the model synthesis directly constitutes the bid price at each time of the next day, without the need for additional decision rules or post-processing mapping.

[0053] Softening of the winning bid determination: A temperature-parameterized sigmoid function is used to make the winning bid event differentiable, replacing the original non-differentiable step-like winning bid indication with a continuous soft winning bid probability. Temperature parameter Controlling the degree of softening: The probability of winning a bid converges to the actual winning bid indication.

[0054] Expected Returns and Coupling Training: Expected settlement revenue is calculated based on the soft-winning probability, declared electricity volume, and the difference between the day-ahead price and the real-time price at each time point. Simultaneously, the weighted prediction error based on the price difference between the declared price and the real-time price is calculated. A coupling loss function is constructed by weighted summation of the revenue loss term and the prediction error term, serving as the optimization objective for end-to-end training.

[0055] End-to-end backpropagation of settlement errors: The gradient information of the coupled loss function is backpropagated through the differentiable decision layer to all learnable parameters of the dual-path temporal encoder, the context memory retrieval module, and the horizontal-morphological decoupling module, ensuring that each update of the model parameters directly points to the direction of maximizing the settlement benefit. The relevant parameters of the synthesis and extrapolation terms are simultaneously optimized under the indirect supervision of the coupled loss function.

[0056] like Figure 6 As shown in (b), the differentiable decision layer takes the decision-oriented bid sequence as input, outputs the soft winning probability at each time step through a temperature-parameterized sigmoid function, and then calculates the expected settlement profit and coupling loss. This process does not require a clearing model or market simulation; it only uses day-ahead and real-time prices as supervision signals, enabling end-to-end training within a standard deep learning framework. This mechanism eliminates the systematic bias between the statistical accuracy objective and the settlement profit objective in the traditional two-stage framework, ensuring that the model training objective aligns with the real business objective.

[0057] This invention also uses the decision-type output price sequence as the bid price, and employs a sigmoid function with a temperature parameter to perform a differentiable approximation of the winning bid event, replacing the discrete indication of the winning bid event with a continuous soft winning bid probability: ; in, For the decision-oriented output quotation sequence, the first... The price at that moment; This refers to the current market price; For temperature parameters, For the first The probability of winning a bid at any given time; The first decision-type bid sequence output in step S4 The model output directly constitutes the bid price at any given time. No additional decision rules are required.

[0058] Expected settlement proceeds are explicitly expressed: ; in, Indicates the declared electricity volume; To the expected settlement profit; This represents the revenue loss item. This relaxation does not require a clearing model or market simulation, preserving the explicit and interpretable structure of "bid, award, and settlement".

[0059] The prediction accuracy target uses a price difference-weighted mean square error, which imposes a higher penalty on errors during periods of high price differences: ; in, This represents the prediction error term; Indicates the first Price difference weighting at each moment; This is the price spread weighting factor. Weight range. When the price difference is zero, it degenerates into ordinary mean square error.

[0060] The two losses are generated by learnable coupled weights. coupling: ; in, The learnable coupling weights, with values ​​between 0 and 1, are trained end-to-end along with the model parameters, and their values ​​are bounded and maintained within [0,1] through a bounded mapping. The two loss terms have different dimensions. The learning process can automatically absorb its relative scale. All parameters are in Down-to-end joint optimization. Temperature. A trade-off between relaxation accuracy and gradient adequacy: The soft index indicates convergence to the true step function, but the gradient vanishes. Increasing the gradient results in a sufficient gradient but also increases the relaxation bias.

[0061] During the inference phase, the information set for the target date is input into the trained model, the bid price sequence is calculated forward, and the bid documents are generated according to market rules.

[0062] Model training and strategy backtesting are conducted under an agreement strictly aligned with the timing of market information disclosure. (Target date) The training set contains no more than [number] training sets. The training sample is truncated with a 1-day lag; day-ahead price features are rolled and aligned into the model with a 1-day lag; daily rolling retraining; the context memory is recalculated and reconstructed with the full training sample after each round of retraining convergence and parameter freezing, and is read-only during the evaluation and prediction periods. Early stopping is implemented during training (the validation set is taken from the last few days of the rolling window to monitor coupling loss). Feature scaling is based only on samples within the training window, and all hyperparameters are fixed after rolling validation calibration before the start of backtesting.

[0063] The load-side day-ahead reporting decision-making method based on contextual memory enhancement and convex hull extrapolation provided in Embodiment 1 of the present invention realizes an end-to-end decision-making process from input of raw market data to output of day-ahead reporting price through the complete technical path consisting of steps S1 to S5. Step S1 constructs a temporal feature tensor through multi-source information fusion, providing a complete information foundation for subsequent encoding. Step S2 decouples the daily global structure from intraday local dynamics through a dual-path temporal encoder, overcoming the limitation of a single attention stream in simultaneously expressing multi-scale temporal features. Step S3 achieves explicit storage and on-demand reuse of rare extreme conditions through three-synaptic contextual memory retrieval, solving the problem of mean regression at the edge of the training distribution for parameterized models. Step S4 overcomes historical convex hull constraints through horizontal-morphological decoupling and learnable extrapolation, enabling the model to still make effective decisions in new price ranges formed by the increase in new energy penetration. Step S5 achieves end-to-end maximization of settlement revenue through differentiable decision-making and coupled training objectives, eliminating the systematic bias of mismatch between prediction and decision objectives. Furthermore, the inference protocol, which strictly aligns with the timing of market information disclosure, eliminates look-ahead bias from a mechanistic perspective. These five steps work synergistically to achieve end-to-end maximization of settlement revenue in the load-side pre-reporting decision-making process and the robustness of strategy generalization under extreme conditions.

[0064] Example 2 To fully verify the load-side day-ahead reporting decision-making method based on contextual memory enhancement and convex hull extrapolation provided in Embodiment 1 of the present invention, Embodiment 2 of the present invention provides a specific example to further illustrate the application process of the method described in Embodiment 1 in the actual electricity spot market reporting scenario.

[0065] Example 2 uses real operating data from a certain electricity spot market as the implementation object, including: The current and real-time 96-point prices (load side according to the unified settlement point electricity price of the user side); The daily release of the 96-point boundary condition forecast for the following day; Regional weather forecast; Intraday time encoding.

[0066] The out-of-sample backtesting interval is a continuous time period, executed daily on a rolling basis: for each target day Only at no later than The sample training is based on the model decision output (truncated to [-80, 1300] yuan / MWh), and the settlement is calculated based on the actual clearing price.

[0067] Suppose that a certain electricity sales company operates on a certain day within the out-of-sample backtesting interval ( day) the day before ( On the day of the event, the electricity sales company needs to submit its time-of-use (TOU) bid price for the next 24 hours to the power trading center. The electricity sales company has deployed the load-side day-ahead bid decision-making device based on contextual memory enhancement and convex hull extrapolation as described in Embodiment 1 of this invention. The device is integrated into the company's transaction decision support system in the form of an intelligent computing all-in-one machine.

[0068] Set the number of applications (Fixed throughout the day), the feasible price range is: Yuan / MWh, price difference weighting factor sigmoid temperature parameter Number of recursive iterations Episodic memory bank capacity Bar, primary key inverse temperature driving key weight Long-distance anchor point interval (Corresponding to 8 anchor points), local window (Including itself, there are 5 time points in total), the DG expansion factor is 6x, and the k-WTA retention ratio is... The driving bond feature dimension is 8, and the normalization constant is... Coupling weights It is a learnable scalar (initial value 0.5).

[0069] Taking a specific operating day within the out-of-sample backtesting interval as an example, the execution process is as follows: Step S1: At 14:30 on the same day (30 minutes before the application deadline), the device automatically starts the data acquisition program to obtain the information set available before the application deadline. This includes: system load forecast, renewable energy output forecast, and inter-regional power exchange forecast for 96:00 the following day (15-minute interval); regional temperature, irradiance, and wind speed forecasts; and forecasts up to [date missing]. The 7-day, 14-day, and 30-day lagged sequences of day-ahead and real-time prices, along with their rolling window statistics, and the sine / cosine period encodings for 96 time points within the day, are used to construct a feature tensor. ,in For the feature dimension (in this embodiment) ).

[0070] Step S2: The device converts the feature tensor Input a dual-path temporal encoder. Let the number of time points in the temporal feature tensor be... Number of anchor points Fixed interval (i.e., take the first) (Time as anchor point), local window (Including itself, there are 5 moments in total).

[0071] In the long-range pathway, the eight anchor points were calculated using multi-head self-attention and then backfilled to all 96 time points via linear interpolation to obtain the long-range representation. In local pathways, attention is only allowed to the four time steps preceding it and to itself at each time step, resulting in a local representation. The two-path representations are then subjected to time-by-time gated adaptive fusion to obtain the fused output sequence. The computational cost of this step is determined by the full-time attention... Down to The calculation took approximately 0.3 seconds.

[0072] Step S3: The device constructs retrieval clues based on the fused output sequence. Separated by DG sparse mode (DG expansion factor 6x, k-WTA retention ratio top- ) to obtain sparse separation characterization ,based on Generate primary search key Simultaneously, eight-dimensional boundary driving quantities, such as system load and new energy output, are extracted from the time-history boundary condition information as driving keys. The contextual memory bank already stores 2048 historical daily trajectories and their corresponding primary search keys and driving keys. The similarity between the primary search key and the primary key of each entry in the memory bank, and the similarity between the driving key and the driving key of each entry in the memory bank, are calculated separately. These similarities are then weighted and summed to obtain the search score for each entry. ; The initial retrieval trajectory is obtained by weighting and summing the historical trajectories in the memory bank according to the scores. Then with The recursive iteration completes the pattern by superimposing the current search trajectory onto the search clues after a linear transformation and then re-searching, thus obtaining the final iteration trajectory. The gating coefficients are calculated based on the consistency between the search clues and the iteration termination trajectory, resulting in the weighted gating trajectory and its gating coefficients. .

[0073] This step retrieves the complete real-time price trajectory pattern of a target day by searching the contextual memory bank for the most similar historical trading days (such as weather and load characteristics of a certain day in history that are highly similar to the target day), providing a priori memory for subsequent synthesis.

[0074] Step S4: The device decouples the fused output sequence and the weighted gated trajectory. Specifically, the fused output sequence, after pooling and mapping to obtain the basic shape at multiple time steps and centering it, is superimposed with the centered residual of the weighted gated trajectory to form the morphological component. ; The time-averaged features of the fused output sequence are estimated using a learnable level estimate, and the average level of the iteration termination trajectory is used as the recall anchor. ; When the estimated level exceeds the historical memory range (e.g., the target day's new energy output hits a historical high, and the predicted price level is lower than the historical low), a correction is made in the same direction using learnable extrapolation terms: ; The corrected horizontal component and the morphological component are superimposed and projected onto the feasible region of the declared price. Yuan / MWh: ; Obtain the decision-making output quotation sequence for the next 24 hours. This step allows for the reuse of price patterns from historical memory, while also covering price levels not seen in history through extrapolation mechanisms.

[0075] Step S5: The device will output a decision-making quotation sequence. As the bid price, a price with temperature parameters is used. The sigmoid function relaxes the winning bid event into a continuous soft winning bid probability: ; If the quoted price for a certain period is 200 yuan / MWh, and the market price for that period is 150 yuan / MWh, then the probability of winning the soft bid is... This means that there is a 99.3% probability of winning the bid during this period, and the settlement will yield the profit from the difference between the previous day's and the real-time price. The expected settlement profit is: ; The coupling loss function is: ; in, ; This step is performed only during the model training phase and is used to calculate the coupling loss function and update the model parameters. In this embodiment, the coupling weights... As a learnable scalar (initial value 0.5), it is trained end-to-end along with the model parameters. During training... The automatic convergence to approximately 0.38 indicates that the model has automatically found the optimal balance between the profit / loss term and the prediction error term.

[0076] During the inference phase, the device directly executes the forward computation of steps S1 to S4, outputting the decision-oriented price sequence. Generate application documents that conform to the trading center's format requirements, which are then submitted by the trader. Submissions must be completed before the deadline of 15:00 on the same day.

[0077] In the above process, steps S1 to S4 constitute a complete forward inference chain, step S5 is used for offline training, and finally online deployment is performed. The entire training and inference process is strictly aligned with the market information disclosure timeline: the training set contains only data no later than [date missing]. The sample data, during inference, allows for the retrieval of entries entirely from historical windows earlier than the target date. The previous day's clearing results announced at 17:00 on the same day will not be included in the model input. The daily real-time price serves only as a post-event monitoring signal, thus eliminating forward bias from the mechanism.

[0078] Taking a specific operating day within the out-of-sample backtesting interval as an example, the forecast for total renewable energy output on that day reached a record high (wind power forecast reached 1.8 times the historical average, and photovoltaic forecast reached 1.5 times the historical average). The system predicted that the day-ahead market price level might be lower than the historical low. Step S3 retrieved the real-time price trajectory pattern of the historically highest renewable energy output day (but lower than the current day) from the scenario memory database. Step S4 used a learnable extrapolation term to extrapolate the level estimate downwards, while the pattern channel reused the price curve pattern of similar historical days. The final synthesized bid price retained the pattern characteristics of historical extreme days and covered price levels not seen in history. This bid price was successfully won in the day-ahead market, capturing the negative electricity price gains caused by the large renewable energy output in the real-time market the following day (which occurred in the real-time market during the midday period). With a negative electricity price of RMB / MWh, this invention achieved a return improvement of approximately 18.7% compared to the two-stage prediction baseline method and approximately 9.2% compared to the scenario stochastic programming baseline method. In the out-of-sample backtesting period, the invention achieved a relative return capture rate of 92.6% on the Oracle, significantly outperforming the 78.3% of the two-stage prediction method and the 84.1% of the scenario stochastic programming method.

[0079] In summary, Embodiment 2 of the present invention, through the specific execution process of steps S1 to S5 described above, fully demonstrates the actual deployment and decision-making process of the method of the present invention in a real electricity spot market operating data environment. This example verifies the effectiveness of the method of the present invention in the following key aspects: The multi-source information fusion in step S1 can automatically complete data collection and feature construction within 30 minutes before the application deadline, meeting the strict requirements of market rules on the legality and availability of information; the dual-path temporal encoder in step S2 reduces the computational load by using sparse anchor point attention. Down to The method completes global and local multi-scale temporal feature extraction at 96:00 the next day within 0.3 seconds, verifying its computational feasibility within the application deadline. In step S3, contextual memory retrieval recalls complete price patterns for similar operating conditions from 2048 historical trajectories, providing effective prior memory for subsequent synthesis. In step S4, level-morphological decoupling, under extreme operating conditions where new energy output reaches record highs (wind power reaches 1.8 times the historical average, photovoltaic reaches 1.5 times the historical average), uses learnable extrapolation terms to extrapolate the level estimate downwards, successfully covering historically unseen price levels (negative midday electricity price of -50 yuan / MWh), verifying the effectiveness of the convex hull extrapolation mechanism. In step S5, the coupling loss function uses learnable weights... The automatic balancing of the gain loss term and the prediction error term (converging to approximately 0.38) verifies the adaptability of end-to-end training; finally, the inference protocol completes the entire forward computation in 0.8 seconds, verifying the engineering feasibility of decoupling offline training and online inference.

[0080] Embodiment 2 of this invention achieves full automation from raw data input to application document output. A 92.6% Oracle relative return capture rate within the out-of-sample backtesting interval, and an 18.7% return improvement compared to the two-stage prediction method, further quantitatively verify the significant superiority of this invention over existing technologies. This demonstrates that the invention achieves the expected technical results in terms of return capture capability and strategy generalization robustness under extreme operating conditions in scenarios with high penetration rates of new energy.

[0081] Example 3 Embodiment 3 of the present invention also provides a load-side pre-reporting decision-making device based on contextual memory enhancement and convex hull extrapolation. Figure 7 This is a schematic diagram of the load-side pre-reporting decision-making device based on contextual memory enhancement and convex hull extrapolation proposed in Embodiment 2 of the present invention. The device includes: At the hardware level, the electronic device 700 includes a processor 710, and optionally, an internal bus 720, a network interface 730, and memory. The memory may include main memory, such as high-speed random-access memory (RAM), or it may also include non-volatile memory, such as at least one disk drive. Of course, the electronic device may also include other hardware required for other business operations.

[0082] The processor 710, network interface 730, and memory can be interconnected via an internal bus 720. This internal bus 720 can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be categorized as an address bus, data bus, control bus, etc. For ease of illustration, only a single bidirectional arrow is used in this diagram, but this does not imply that there is only one bus or one type of bus. The memory is used to store programs. Specifically, the program can include program code, which includes computer operation instructions. The memory can include main memory 740 and non-volatile memory 750, and provides instructions and data to the processor 710.

[0083] In step S1: Obtain the information set before the application deadline, and construct a time series feature tensor based on the information set; the information set includes the time history boundary condition information of the operating day, meteorological environment information related to the load-side output, lagged statistics and lagged price information of historical market operation data, and time location encoding information; In step S2: The temporal feature tensor is encoded using a dual-path temporal encoder, which includes a long-range path for extracting the global temporal structure and a local path for extracting local dynamic features. The long-range path and the local path are fused to obtain a fused output sequence. In step S3: a retrieval clue is constructed based on the fused output sequence. After pattern separation of the retrieval clue, a master retrieval key is constructed, and a driving key is constructed by extracting the boundary driving quantity. The master retrieval key and the driving key are used to perform a double-key retrieval to obtain an initial retrieval trajectory. The initial retrieval trajectory is fed back to update the retrieval clue and iterative retrieval is performed to obtain the iteration termination trajectory. The gating coefficient is calculated based on the consistency between the retrieval clue and the iteration termination trajectory to obtain the weighted gating trajectory. In step S4: the fused output sequence and the weighted gated trajectory are decoupled into a morphological component and a horizontal component. The morphological component is formed by superimposing the initial morphology of the fused output sequence and the residual of the weighted gated trajectory. The horizontal component is estimated by the time-averaged features of the fused output sequence through a learnable method and anchored to the horizontal average value of the weighted gated trajectory. When the estimated horizontal value exceeds the historical memory range, it is corrected by a learnable extrapolation term. The corrected horizontal component is superimposed with the morphological component and projected onto the feasible region of the declared price to obtain the decision-type output quotation sequence.

[0084] In step S5, the decision-making and end-to-end training of the micro-report can be carried out, as well as the inference and application document generation.

[0085] Figure 1 It can be applied to processor 710, or implemented by processor 710. The processor may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit in the processor or by instructions in the form of software. The processor mentioned above can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this application can be directly embodied as being executed by a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0086] Of course, in addition to software implementation, the electronic device of this application does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. In other words, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0087] The description of the relevant parts of the load-side day-ahead reporting decision-making device based on context memory enhancement and convex hull extrapolation provided in Embodiment 3 of this application can be found in the detailed description of the corresponding parts of the load-side day-ahead reporting decision-making method based on context memory enhancement and convex hull extrapolation provided in Embodiment 1 of this application, and will not be repeated here.

[0088] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that the elements inherent in a process, method, article, or apparatus that includes a list of elements are included. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. Additionally, portions of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0089] While specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art can make other modifications or variations based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A load-side pre-reporting decision-making method based on contextual memory enhancement and convex hull extrapolation, characterized in that, Includes the following steps: Obtain the information set prior to the application deadline, and construct a temporal feature tensor based on the information set; The information set includes time-history boundary condition information for the operating day, meteorological environment information related to load-side output, lagged statistics and lagged price information of historical market operation data, and time location coding information. The temporal feature tensor is encoded using a dual-path temporal encoder, which includes a long-range path for extracting global temporal structure and a local path for extracting local dynamic features. The long-range path and the local path are fused to obtain a fused output sequence. Based on the fusion output sequence, a retrieval clue is constructed. After pattern separation of the retrieval clue, a master retrieval key is constructed. Boundary driving variables are extracted to construct a driving key. The master retrieval key and the driving key are used to perform a double-key retrieval to obtain an initial retrieval trajectory. The initial retrieval trajectory is fed back to update the retrieval clue and iterative retrieval is performed to obtain the iteration termination trajectory. The gating coefficient is calculated based on the consistency between the retrieval clue and the iteration termination trajectory to obtain the weighted gating trajectory. The fused output sequence and the weighted gated trajectory are decoupled into morphological components and horizontal components. The morphological component is formed by superimposing the initial morphology of the fused output sequence and the residual of the weighted gated trajectory. The horizontal component is estimated by learning through the time-averaged features of the fused output sequence and anchored to the horizontal average value of the weighted gated trajectory. When the estimated horizontal value exceeds the historical memory range, it is corrected by learning extrapolation terms. The corrected horizontal component is superimposed with the morphological component and projected onto the feasible region of the declared price to obtain the decision-oriented output quotation sequence.

2. The method according to claim 1, characterized in that, The method further includes: The decision-type output price sequence is used as the bid price. The winning bid event is approximated by a differentiable sigmoid function with a temperature parameter. The discrete indication of the winning bid event is replaced by a continuous soft winning bid probability. The expected settlement is calculated based on the soft winning probability, the declared electricity volume, and the difference between the day-ahead price and the real-time price at each time point; a coupled loss function is constructed, consisting of a weighted sum of a revenue loss term and a prediction error term, where the revenue loss term is the negative value of the expected settlement revenue, and the prediction error term is the weighted sum of squares of the errors between the declared price and the real-time price. The parameters of dual-path encoding, retrieval, decoupling, and differentiable decision-making are jointly optimized end-to-end under the coupling loss function.

3. The method according to claim 1, characterized in that, The long-range path for extracting the global temporal structure is as follows: anchor points are selected from the temporal feature tensor at fixed intervals, the anchor point sequence is used to establish global dependencies through multi-head self-attention calculation, the anchor point output is backfilled to all time points through linear interpolation, and then the long-range representation is output after residual connection and layer normalization.

4. The method according to claim 1, characterized in that, Local pathways for extracting local dynamic features are specifically as follows: For each time step, attention scores are calculated only between the query vector at that time step and the key vectors of the previous set number of historical time steps, including the time step itself. Negative infinity masks are applied to positions where attention is not allowed, so that the attention weights corresponding to the masked positions are normalized to zero. The calculated attention weights are summed with the value vectors at the corresponding time points to obtain the local representation at that time point.

5. The method according to claim 1, characterized in that, Retrieval clues are constructed based on the fused output sequence, and pattern separation is performed on the retrieval clues, specifically as follows: Search clues via linear projection After high-dimensional nonlinear activation, a fixed proportion of activation components are retained through a competition mechanism to obtain a sparse separation characterization: ； in, Characterization of sparse separation; This is a contention mask used to retain a fixed percentage of the active value, while setting the remaining components to zero. The linear projection matrix for pattern separation; To retrieve clues; It is a non-linear activation function; This represents element-wise product.

6. The method according to claim 5, characterized in that, Construct the primary retrieval key and extract the boundary driving variables to construct the driving key. Use the primary retrieval key and the driving key to perform a two-key retrieval to obtain the initial retrieval trajectory, specifically as follows: Based on the sparse separation characterization Generate primary search key The boundary driving quantity is extracted from the time history boundary condition information as the driving key. The similarity between the primary search key and the primary keys of each entry in the search memory, and the similarity between the driving key and the driving keys of each entry in the search memory are calculated separately. These similarities are then weighted and summed to obtain the search score for each entry. ； in, For the first The search score for each item; The inverse temperature of the main bond; For driving key weights; Use as the primary search key; For the first in the memory bank The primary search key for each entry; The dimension of the key vector; For driving keys; For the first in the memory bank The driving key for each entry; To retrieve the weight vector; For the first in the memory bank Each entry stores the historical real-time price trajectory; This is the initial search trajectory; For the first The retrieval weight of memory entries.

7. The method according to claim 6, characterized in that, The initial search trajectory is fed back to update the search clues, and the search is iterated to obtain the iteration termination trajectory. The gating coefficient is calculated based on the consistency between the search clues and the iteration termination trajectory to obtain the weighted gating trajectory, specifically: To fix The recursive iteration of the steps implements pattern completion, superimposing the current search trajectory onto the search clues after linear transformation to update the search clues: ； in, Use as the primary search key; For the feedback projection matrix; For the first The retrieval trajectory of each step; It is a dimensionless normalization constant; For layer normalization; Indicates based on primary search key and drive key Perform a double-key retrieval and output the weighted trajectory; Output the termination trajectory after one iteration. Use a fixed number of steps instead of a convergence criterion to eliminate non-differentiable branches; The gating coefficient is: ； in, This is the gating coefficient, with a value between 0 and 1; It is a multilayer perceptron; It is the sigmoid function; This indicates vector concatenation; To retrieve the projection of the clue in the comparison space; The projection of the iteration termination trajectory into the comparison space; The comparison space projection matrix for retrieval clues; The comparison space projection matrix of the trajectory; Represents element-wise product; To retrieve clues; The trajectory is the termination trajectory of the iteration; The gating coefficients are applied to the iteration termination trajectory to obtain the weighted gating trajectory and its gating coefficients.

8. The method according to claim 1, characterized in that, The morphological component is composed of the superposition of the initial morphology of the fused output sequence and the residual of the weighted gated trajectory, specifically expressed as: ； in, For fusion output sequence; The total number of moments in the temporal feature tensor. This is a combined operation of pooling and mapping. Based on shape, The gating coefficient, Inject the matrix into the residual. For the termination trajectory of the iteration, To calculate the average water value, To remember shape residuals, It is a vector whose elements are all 1s; The level component is estimated by using the time-averaged features of the fused output sequence to obtain the level value through learnable level estimation, and is jointly given by the level average of the iteration termination trajectory as the recall anchor term, specifically expressed as follows: ； in, For horizontal components, For the learnable level estimation function, To fuse the time-averaged features of the output sequence, For anchoring items of memory; For anchoring weight coefficients, The water average value of the iterative termination trajectory.

9. The method according to claim 8, characterized in that, When the estimated level value exceeds the maximum level value of historical trajectories in the historical memory bank, a correction is made by proportionally amplifying the anchoring through learnable extrapolation terms. Specifically: ； in, For exceeding the limit, To stop the gradient operation, This represents the maximum value of the historical trajectory level in the historical memory bank. The normalization constant is For learnable nonnegative extrapolation magnitude, It is a learnable scalar; It is a smooth positive function; The corrected horizontal component is superimposed on the morphological component to obtain the synthesized trajectory: ； in, The output price sequence is for decision-making purposes. This is an operation that truncates each element to the feasible domain of the declared price.

10. A load-side pre-reporting decision-making device based on contextual memory enhancement and convex hull extrapolation, comprising at least one processor and a memory, wherein the memory stores a computer program, characterized in that, When the computer program is executed by the at least one processor, it implements the load-side pre-reporting decision method based on contextual memory enhancement and convex hull extrapolation as described in any one of claims 1 to 9.

Images