An electric quantity allocation data processing method based on ensemble learning

Abstract

The application provides a power allocation data processing method based on ensemble learning, and relates to the technical field of power system calculation and data processing. The method is used for constructing a multi-source heterogeneous feature matrix containing a system supply-demand balance index for obtaining historical data; an ensemble learning model is guided by the index to split nodes, and a time-period prediction electricity price data sequence is output; an isolated forest algorithm is used to process historical spot data to generate spot clearing benchmark data curves; a dynamic difference between the prediction sequence and the benchmark curve is mapped to gravity field strength data through an asymmetric value function; a repulsive damping coefficient changing exponentially is calculated by extracting bottom physical support boundary parameters; the gravity field strength data and the repulsive damping coefficient are superimposed to obtain optimal power allocation data processing results. The application solves the problems of low prediction accuracy under nonlinear working conditions, and excessive adjustment and numerical oscillation caused by lack of asymmetric adjustment and flexible constraints in resource allocation.

Description

Technical Field

[0001] This invention relates to the field of power system calculation and data processing technology, and in particular to a power allocation data processing method based on ensemble learning. Background Technology

[0002] As the electricity market system under full-volume bidding matures, formulating scientific and precise monthly medium- and long-term electricity allocation and regulation strategies has become a core means of maintaining system resource stability and mitigating the risk of drastic fluctuations in spot market prices. However, existing electricity allocation decision-making data processing technologies still face many challenges in practical applications, mainly in the following three aspects: First, the accuracy of predictions based on underlying data sequences is limited. Traditional prediction methods often employ statistical methods such as the Autoregressive Moving Average (ARIMA) model and exponential smoothing, or simple multiple linear regression models. These methods are typically based on linear assumptions and struggle to effectively capture the complex nonlinear relationships between multi-source heterogeneous data in the power system operating environment. Particularly for regional power grids with a high proportion of clean energy, system clearing prices are heavily influenced by multiple dynamic factors, including weather changes, water inflow, and the physical balance of supply and demand, exhibiting strong volatility and randomness. Traditional models lack in-depth analysis of physical marginal characteristics, often resulting in significant prediction errors during operational transitions, failing to meet the demands of refined regulation.

[0003] Second, the control model lacks a dynamic asymmetric response mechanism. Most existing resource allocation optimization models are linear deterministic models, assuming absolutely accurate input data when executing strategy optimization. However, environmental noise is inevitable in actual operation. When the prediction system has a small data deviation, the linear deterministic model may still perform a large number of resource actions based on that deviation; while facing extreme downside risks, it often responds slowly. Existing methods lack an asymmetric driving mechanism that can adaptively adjust according to the direction and magnitude of the deviation, which easily leads to over-adjustment and high-frequency ineffective actions in the system.

[0004] Third, there is insufficient flexible constraint control over the underlying physical security and compliance boundaries. In actual power grid dispatching operations, power allocation must strictly adhere to the underlying physical support capacity boundaries of generator units and the bottom line of external rule assessments. Existing methods often use rigid "if-else" logic thresholds in program code to truncate these boundary conditions. This rigid boundary handling not only disrupts the continuity of the optimization algorithm in the solution space but also easily triggers severe numerical oscillations in the system when the power allocation ratio approaches the limit boundary, causing the solution process to fail to converge smoothly and even outputting erroneous execution instructions that exceed the limits.

[0005] In summary, how to utilize advanced machine learning algorithms to deeply mine multi-source heterogeneous data to improve prediction accuracy under nonlinear operating conditions, and on this basis, construct a power distribution data processing method that takes into account both adaptive response adjustment and flexible physical boundary control, is a technical problem that urgently needs to be solved in related fields. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, this invention provides a power allocation data processing method based on ensemble learning. This method solves the problems of low prediction accuracy due to the inability to accurately capture nonlinear operating conditions, and the tendency for over-adjustment and edge numerical oscillations in resource optimization allocation due to the lack of asymmetric adjustment mechanisms and flexible boundary constraints.

[0007] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a power allocation data processing method based on ensemble learning, comprising: S1: Obtain historical power grid operation status data and market bidding interaction data to construct a multi-source heterogeneous feature matrix; the multi-source heterogeneous feature matrix includes system supply and demand balance indicators; S2: Using the multi-source heterogeneous feature matrix as input, an ensemble learning prediction model guided by physical marginal features is constructed. In the node splitting stage of the decision tree in the ensemble learning prediction model, the system supply and demand balance index is used to guide the ensemble learning prediction model to divide the feature space and obtain the time-segmented predicted electricity price data sequence. S3: Based on historical spot market operation data, the isolated forest algorithm is used to identify and remove outlier data points, and a spot clearing benchmark data curve is generated as a zero-deviation reference benchmark. S4: Construct a piecewise asymmetric value function to map the dynamic difference signal between the predicted electricity price data sequence and the spot clearing benchmark data curve into adjustment potential energy. Through calculation, obtain the gravitational field strength data that characterizes the direction and intensity of adjustment. S5: Extract the physical support boundary parameters of the generator set's bottom layer, and calculate the repulsive damping coefficient that changes exponentially as the target power distribution approaches the physical support boundary. S6: Calculate the potential field by superimposing the gravitational field strength data and the repulsive damping coefficient to obtain the optimal power allocation data processing result that satisfies the underlying physical security constraints, thus completing the power allocation data processing.

[0008] The beneficial effects of this invention are as follows: This invention provides a power allocation data processing method based on ensemble learning, which has the following beneficial effects: (1) This invention constructs a multi-source heterogeneous feature matrix by acquiring power grid operation and market interaction data. This matrix not only includes basic supply and demand load data but also innovatively introduces time-series price structure entropy and system supply and demand balance index. In the ensemble learning prediction model construction stage, the system supply and demand balance index is used as a physical marginal feature to guide the model in feature space partitioning, and the isolated forest algorithm is combined to purify the spot clearing benchmark data curve. This mechanism effectively overcomes the shortcomings of traditional statistical models in handling sudden changes in system operating state, deeply explores the underlying nonlinear influence laws, and significantly improves the accuracy and robustness of time-segmented predicted electricity price data sequences.

[0009] (2) This invention overcomes the limitation of symmetric response in traditional linear models by constructing a piecewise asymmetric value function to convert dynamic difference signals into dimensionless gravitational field strength data. This function exhibits saturated compression characteristics in the positive difference interval, effectively limiting over-configuration commands caused by small fluctuations; and polarization amplification characteristics in the negative difference interval, enabling rapid generation of defensive reverse callback actions. This nonlinear mapping mechanism allows the system to intelligently identify the true adjustment value of environmental signals, enhancing the system's optimization stability in complex fluctuating environments.

[0010] (3) This invention abandons the traditional hard threshold truncation method, extracts the underlying physical support boundary parameters of the generator set, and innovatively calculates and generates a repulsive damping coefficient that changes exponentially as the target allocated power approaches the physical support boundary. By superimposing the gravitational field strength data with this repulsive damping coefficient, the system can automatically generate a gradually increasing reverse virtual repulsive force when the virtual allocation approaches the underlying physical safety limit. This not only completely eliminates the solution space oscillation caused by hard truncation, but also gives the system the ability to smoothly approach the optimal power allocation data processing result within the safety assessment red line, achieving dual protection of compliance and underlying physical safety. Attached Figure Description

[0011] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein: Figure 1 This is an exemplary flowchart of a power allocation data processing method based on ensemble learning, as shown in some embodiments of this specification. Detailed Implementation

[0012] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0013] Example Figure 1 This is an exemplary flowchart illustrating a power allocation data processing method based on ensemble learning, according to some embodiments of this specification. Figure 1 As shown, the process includes the following steps. In some embodiments, the process may be executed by a processor.

[0014] S1: Obtain historical power grid operation status data and market bidding interaction data to construct a multi-source heterogeneous feature matrix; among which, the multi-source heterogeneous feature matrix includes the system supply and demand balance index.

[0015] Historical power grid operation status data is a collection of data reflecting the actual operating conditions of the real power grid physical system during various dispatch cycles over a past period. For example, historical power grid operation status data may include statistical values ​​of actual power generation of various types of generating units, such as hydropower units, wind power units, photovoltaic units, and thermal power units in each past natural month; time-series label data for different hydrological cycles, such as the high-water season, normal-water season, and low-water season; and basic parameters reflecting the physical supply and demand situation, such as the actual load demand generated by the whole society during the corresponding period.

[0016] In some embodiments, the processor can extract generator telemetry data records and power grid gateway metering data within the corresponding historical time window by calling the historical database at the bottom layer of the power grid dispatching system. Then, the extracted multi-source heterogeneous data is time-scale aligned and cleaned and denoised. Finally, the processed physical quantities of each dimension are summarized to obtain historical power grid operation status data.

[0017] Market bidding interaction data reflects price-related interaction signals generated by various market participants in the power resource allocation process within the power grid operating environment. For example, market bidding interaction data may include the actual average transaction price data of the previous settlement period, the confirmed annual resource allocation benchmark price data, and derived calculation indicators that reflect the non-steady-state fluctuation characteristics of price distribution during the replacement of marginal units (such as hydropower units and thermal power units) in the system, used to characterize the price inertia and fluctuation state during the resource allocation process.

[0018] In some embodiments, the processor can access the settlement information system of the power trading center, read the settlement vouchers and contract records of each transaction within a set period, calculate the weighted average price for each time period, extract the long-term basic price record for the corresponding year, and obtain market bidding interaction data after data format conversion and structured encapsulation.

[0019] A multi-source heterogeneous feature matrix is ​​a high-dimensional structured data set used to comprehensively characterize the complexity of the physical supply and demand state and resource allocation environment of a power system. For example, a multi-source heterogeneous feature matrix can include actual power generation data such as hydropower, wind power, and photovoltaic power extracted from the supply side; centrally dispatched load data extracted from the demand side; seasonal label data reflecting time periods; and feature variables of various dimensions and data types, such as time-series price structure entropy and system supply and demand balance indicators that deeply characterize the nonlinear and abrupt changes in the system.

[0020] In some embodiments, the processor can acquire historical power grid operation status data and market bidding interaction data separately, normalize the data from different sources to eliminate dimensional differences, and then splice and combine them according to a unified time series dimension. The calculated time series price structure entropy and system supply-demand balance indicators, among other derived indicators, are added as independent feature columns and fused together to obtain a multi-source heterogeneous feature matrix. Table 1 shows a subset of the model input feature parameters, which specifically illustrates the data dimensions collected when constructing the multi-source heterogeneous feature matrix. On the supply side, hydropower and wind / solar power generation reflecting the physical output scale of the entire grid are extracted; on the demand side, data reflecting the predicted electricity load of the entire society are extracted; the time series dimension marks the wet / dry hydrological cycle; and on the bidding behavior dimension, the basic reference price is extracted, and the structure entropy and supply-demand balance indicators characterizing the nonlinear fluctuations of the system are calculated. These multi-source data collectively constitute the underlying feature matrix driving the prediction model.

[0021] Table 1. Subset of Model Input Feature Parameters

[0022] In some embodiments, the processor can extract clearing price sequence data from historical power grid operation status data, calculate the ratio of its standard deviation to mean, and obtain the time-series price structure entropy; acquire the load forecast data of the entire network and the forecast output limit data of the dispatchable units of the entire network, calculate the ratio of the two, and obtain the system supply and demand balance index; and fuse the basic supply and demand load data, market bidding interaction data, time-series price structure entropy and system supply and demand balance index to construct a multi-source heterogeneous feature matrix.

[0023] Clearing price sequence data is a set of unified settlement price data at specific time points, determined after underlying physical supply and demand matching calculations in a spot market environment. For example, clearing price sequence data may include unified marginal electricity price values ​​recorded every 15 minutes or hourly over multiple past scheduling cycles. This sequence data reflects the dynamic evolution of system resource value over time due to real-time changes in supply and demand and marginal unit switching.

[0024] In some embodiments, the processor can extract the node marginal price records generated when the system solves the optimal power flow calculation under the premise of meeting load demand and physical security constraints by parsing the operation log of the historical power grid dispatch automation system, and splice and reassemble these discrete price records in chronological order to obtain clearing price sequence data arranged in time sequence.

[0025] Time-series price structure entropy is a characteristic indicator used to quantify the degree of non-steady-state fluctuations in the distribution of electricity market price data as system operating conditions change. For example, time-series price structure entropy can include a value calculated based on the dispersion of clearing price samples within a certain historical period. The larger the value of this indicator, the more drastic the jumps in resource allocation value caused by uncertainties in external inputs, and the stronger the randomness of the system's operating state.

[0026] In some embodiments, the processor can select historical clearing electricity price sequence data from the same period as input samples, first calculate the standard deviation of the electricity price sequence data within a set time window to measure the absolute dispersion of the data; then calculate the arithmetic mean of the electricity price sequence data within the same time window; finally, divide the calculated standard deviation by the arithmetic mean, and use the result of the relative dispersion calculation as the time series price structure entropy.

[0027] In some embodiments, to deeply characterize the nonlinear abrupt changes in the operating conditions of the electricity market, the calculation expression for the time-series price structure entropy is as follows: ; in, Represents the entropy of the time-series price structure. This represents the function for calculating standard deviation. This represents the function for calculating the arithmetic mean. This represents the historical unified clearing electricity price series data for the same period in the spot market. This indicator is used to characterize the degree of relative non-steady-state fluctuation in the distribution of market price data caused by input-side uncertainties during the system's marginal unit replacement process.

[0028] The total grid load forecast data is an estimate of the total electricity demand of the whole society within a specific future time period, calculated in advance based on the historical operating patterns of the power grid and external factors such as weather. For example, the total grid load forecast data may include monthly or hourly power grid load demand curve data. This data reflects the total load target that the physical power grid needs to meet at future moments and is a core demand-side benchmark parameter for subsequent power grid resource allocation optimization and physical supply-demand balance calculation.

[0029] In some embodiments, the processor can acquire weather forecast data, historical actual load data for the same period, and holiday calendar information, and then input this information into a pre-trained time series prediction model. Through forward propagation calculation of the model, the processor outputs the predicted value of the total active power demand of each node of the power grid in a future specified period, thereby obtaining the load prediction data of the entire network.

[0030] Predicted power output ceiling data characterizes the maximum power generation capacity that various generating units in the power grid can achieve under specific future operating conditions, limited by natural resources or equipment physical conditions. For example, predicted power output ceiling data may include the estimated maximum power generation of hydropower units limited by hydrological forecasts, the maximum output of thermal power units limited by equipment maintenance and fuel constraints, and the maximum available output value of new energy units such as wind and solar power units based on meteorological conditions, for all dispatchable units in the entire network.

[0031] In some embodiments, the processor can call the physical operating parameter library of various types of units and the data of the external environment forecasting system respectively, combine the water inflow forecast data of the hydroelectric domain with the reservoir scheduling equation for calculation, and combine the wind speed and solar radiation forecast data with the output characteristic curve of the new energy units for calculation. Then, the theoretical maximum power generation capacity calculated by each unit is summed and aggregated to obtain the upper limit data of the predicted output for the corresponding time period.

[0032] The system supply-demand balance index is a physical parameter that reflects the tension between the load demand and the total physical supply capacity of the underlying generating units in the current or future period. For example, the system supply-demand balance index may include a dimensionless ratio data, which is used to identify the physical critical point at which marginal generating units in the system switch over. When the index exceeds a certain threshold, it means that the system pricing logic may undergo a nonlinear jump from one type of generating unit to another.

[0033] In some embodiments, the processor can synchronously acquire the network load forecast data at the corresponding time node, as well as the forecasted upper limit data of the network's dispatchable hydropower, thermal power and new energy units, accumulate the forecasted upper limit data of each type of unit to obtain the system's total physical supply capacity base, and then divide the network load forecast data by the total physical supply capacity base. The resulting division ratio is the system's supply and demand balance index.

[0034] In some embodiments, to reflect the tension between the current grid load demand and the physical supply capacity of generating units, the system supply-demand balance index is calculated using the following expression: ; in, Indicates the current time period The system supply and demand balance index, Indicates the current time period Network-wide load forecasting data obtained from power grid dispatching data; , and These represent the current time period extracted based on the physical operating boundary parameters of each unit. Data on the predicted upper limit of power output of hydropower units, thermal power units, and new energy units that can be dispatched throughout the network.

[0035] Basic supply and demand load data is a set of data that constitutes the original supply capacity and demand status of the underlying physical operation of the power grid system. For example, basic supply and demand load data can include power generation records of hydropower units in various river basins of the entire network, output data records of wind farms and photovoltaic power stations, and actual load power consumption data of the entire network or sub-regions. These data constitute the most basic set of parameters describing the balance between energy inflow and outflow in the power grid system.

[0036] In some embodiments, the processor can directly read the historical measurement data of each physical measurement node stored in the database of the power grid energy management system, perform time-series alignment operations on the collected generator active power output readings and substation load outgoing power readings, remove abnormal reading values ​​caused by communication failures, classify, reorganize and store them according to the supply side and demand side, and finally obtain basic supply and demand load data.

[0037] S2: Using the multi-source heterogeneous feature matrix as input, an ensemble learning prediction model guided by physical marginal features is constructed. In the node splitting stage of the decision tree in the ensemble learning prediction model, the system supply and demand balance index is used to guide the ensemble learning prediction model to divide the feature space and obtain the time-segmented predicted electricity price data sequence.

[0038] Physical marginal characteristics are data that characterize the critical physical properties of a power system when its operating state undergoes nonlinear abrupt changes (such as a switch in the type of generating unit that dominates pricing). For example, physical marginal characteristics can include the system supply and demand balance index, which reflects the tension state when the system load approaches the upper limit of the total generating unit capacity. It embodies the boundary conditions when the system's physical operating conditions transition from relaxed to tense, and is a key guiding signal that drives predictive models to accurately capture nonlinear pricing logic.

[0039] In some embodiments, the processor can extract feature values ​​reflecting the relative tightness of supply and demand by calculating the ratio between the extracted basic load forecast data and the generator output upper limit data, and then combine and extract these values ​​and the first derivative difference calculation results reflecting the changing trend as structured variables describing the critical state of the system, thereby obtaining high-dimensional physical marginal feature data that can be used for model training.

[0040] An ensemble learning prediction model is a machine learning network structure composed of multiple basic predictors, capable of deeply mining features from multi-source heterogeneous data to fit complex nonlinear relationships. For example, an ensemble learning prediction model may include a model structure built on the random forest algorithm, whose input layer receives multi-source heterogeneous feature matrices; the internal intermediate processing layer contains multiple independently growing decision trees, and the feature space is partitioned by forcing the system supply and demand balance index as the priority condition during the decision tree node splitting stage; the output layer is used to extract and integrate the leaf node distribution features of each decision tree to output the final predicted value.

[0041] In some embodiments, the processor can extract feature matrices from historical datasets using hierarchical sampling techniques to construct a training sample set, and then initialize multiple decision tree model structures. During the parameter training phase, the node splitting criterion logic is modified, and the system supply and demand balance is used as the priority splitting feature input to the model for tree structure growth calculation. After training and error convergence are completed, the network node parameters of all decision trees are saved, thus obtaining the constructed ensemble learning prediction model.

[0042] The predicted electricity price data series is a time series set output by an ensemble learning prediction model, reflecting the reference value of electricity resource allocation over a future period. For example, the predicted electricity price data series can include the predicted average price values ​​for each consecutive period, divided by day or hour, for a future target month, along with their corresponding confidence intervals. This series of data objectively removes interference from human commercial intentions and serves as the core basic data input for calculating the dynamic adjustment potential energy.

[0043] In some embodiments, the processor can send a data stream containing a multi-source heterogeneous feature matrix of the period to be predicted into an ensemble learning prediction model with loaded network weights, triggering forward inference logic calculation of multiple decision trees within the model; then read the numerical distribution state of all decision tree calculations falling into the terminal leaf nodes, perform weighted average statistical analysis on these numerical states, and splice the statistical results in chronological order to generate a predicted electricity price data sequence.

[0044] In some embodiments, the processor may employ hierarchical sampling techniques to construct a sample set based on a multi-source heterogeneous feature matrix, thereby obtaining a training sample set. During the model training phase, the system supply and demand balance index is used as a splitting feature and the training sample set as inputs to the ensemble learning prediction model. This guides the decision tree to prioritize capturing the critical points of system pricing logic jumps and perform feature space partitioning, thereby obtaining leaf node distribution features. Using the leaf node distribution features of the decision tree in the ensemble learning prediction model, a time-segmented predicted electricity price data sequence containing the predicted average price and its confidence interval is generated.

[0045] The training sample set is a dataset containing known input feature data and corresponding target label data, used to drive the parameter updates and internal logic fitting of the machine learning model. For example, the training sample set may include multi-source heterogeneous feature matrices extracted based on historical time periods as feature variables, and actual clearing electricity price sequences corresponding to historical periods as target labels. The data distribution within this set covers typical environmental states of different underlying physical conditions.

[0046] In some embodiments, the processor can acquire the original historical power grid operating status characteristic data and historical clearing price records, execute stratified sampling calculation logic, divide the total data matrix into rows according to the preset hydrological cycle environmental labels, and then randomly extract data feature and label data pairs from each of the divided data sub-blocks according to a set ratio, and reorganize and integrate them to obtain a training sample set that can cover all operating conditions.

[0047] Leaf node distribution features are the set of statistical attributes of an input sample in an ensemble learning model, representing the data that appears when the input sample finally falls into a terminal node after the top-down path judgment calculation of multiple decision trees. For example, leaf node distribution features may include the arithmetic mean, discrete variance, and numerical density distribution range of the numerical sets associated with the input target sample in all decision tree terminal nodes. This feature comprehensively reflects the degree of convergence of the model network's judgments on the current input state.

[0048] In some embodiments, the processor can monitor and record the complete path index of the target input data features as they traverse and compute through hundreds of independent decision tree structures when controlling the ensemble learning model to perform forward inference tasks; when each path computes to the terminal structure, it reads and extracts the associated mean value stored in its leaf node register, summarizes the values ​​in all independent output terminal registers and performs statistical induction operations to obtain the leaf node distribution features.

[0049] S3: Based on historical spot market operation data, the isolated forest algorithm is used to identify and remove outlier data points, generating a spot clearing benchmark data curve as a zero-bias reference.

[0050] Historical spot market operation data refers to data records of the underlying physical operation status and transaction results of the power grid with the same natural attributes as the target time period to be predicted, on a past time axis. For example, historical spot market operation data may include records of nodal marginal electricity prices for the same month in history, data on the physical start-up status of generator units, data on line power flow congestion, and actual dispatch load data, etc. These data constitute the data source for mining the current system zero-deviation reference benchmark.

[0051] In some embodiments, the processor can extract the time-stamp features to be predicted from the current system input, such as the month field, and then traverse backwards to access the power grid historical database server. It can perform conditional retrieval according to the annual cycle step size, extract all data record files from the previous few years that match the time-stamp field, and extract and assemble the retrieved cross-year underlying scheduling records and clearing price data to obtain historical spot market operation data for the same period.

[0052] The spot clearing benchmark data curve is a smoothed curve reflecting the time-series fluctuations in the typical resource allocation value of the system under a specific physical environment after removing abnormal scheduling instructions or random physical disturbances. For example, the spot clearing benchmark data curve may include a reference continuous numerical vector that changes over time, representing the stable benchmark state data surface that the system should objectively exhibit under the current hydrological or seasonal conditions.

[0053] In some embodiments, the processor can input the historical spot market operation data containing noise into the isolated forest algorithm model as the initial sample. The algorithm calculates and extracts the abnormal scores of each data point, performs filtering operations to delete data points with abnormal scores exceeding the standard to purify the sample space. Subsequently, the remaining normal numerical samples after purification are clustered according to the same time index node and the average value is calculated. The calculated average values ​​of each time period are spliced ​​together according to the time dimension to generate the spot clearing benchmark data curve.

[0054] In some embodiments, the processor can retrieve initial historical price sample data with the same operating characteristics from historical spot market operation data based on the similarity of power grid supply and demand; use the isolated forest algorithm to identify and remove abnormal data points caused by non-steady-state instructions in the initial historical price sample data to obtain a purified sample space; extract time-series features from the purified sample space to generate a spot clearing benchmark data curve that reflects the fluctuation pattern.

[0055] Grid supply-demand similarity is a numerical characteristic used to measure the similarity of the underlying physical energy supply and demand relationship of a power grid system at different historical time sections. For example, grid supply-demand similarity can include specific quantitative values ​​obtained by calculating the feature vectors of the total load and the total physical supply capacity of various generating units in the current target time period, and the Euclidean or cosine distance between them and the feature vectors of each time period in the historical database. This is used to accurately locate historical data samples with reference value.

[0056] In some embodiments, the processor can construct a multidimensional data vector by extracting the total network load forecast data and the total physical supply capacity base of the current forecast period and a specific historical period, respectively, and call the similarity calculation module, such as executing the cosine similarity calculation function, using the multidimensional data vector as the function input parameter to perform algebraic dot multiplication and modulus division operations, and using the calculated spatial angle feature mapping value as the power grid supply and demand similarity output.

[0057] The initial historical price sample data is a rough selection set of price values ​​initially recalled from the historical database based on environmental similarity characteristics without outlier detection and removal. For example, the initial historical price sample data may include all actual historical clearing price records recalled based on the calculated power grid supply and demand similarity score, which inevitably includes extreme noise data points that deviate from the normal pattern due to temporary physical fault trips or extreme manual dispatch interventions.

[0058] In some embodiments, the processor can read the grid supply and demand similarity score data for all historical periods output by the front-end computing module, execute condition judgment logic, and filter out all target historical time indices with similarity scores greater than a preset empirical judgment threshold; then, according to these extracted time indices, perform a full query and read from the historical clearing database, merge and summarize all discrete electricity price value records retrieved by the query, and obtain the initial historical price sample data.

[0059] Outlier data points are isolated data records in a specified dataset whose numerical feature vectors deviate significantly from the dense distribution area of ​​the main data, representing isolated data records caused by unsteady physical operations or extreme random disturbances in the system. For example, outlier data points may include extremely high spikes in the initial recalled electricity price samples due to low-probability events such as local grid congestion or sudden equipment maintenance. These isolated data do not represent the normal physical patterns of the system.

[0060] In some embodiments, the processor can input the acquired initial historical price sample data into the isolated forest algorithm program, which constructs multiple isolated trees by randomly selecting features and randomly dividing the cutting plane; then it tracks the segmentation path depth required for each data point sample to be isolated from the root node to the leaf node in the isolated tree; and filters out sample data with extremely short average segmentation paths and isolation scores greater than the judgment warning threshold, identifies them, and extracts them as abnormal data points.

[0061] The purified sample space is a high-quality set of numerical values ​​that reflects the stable endogenous laws of the system, remaining after removing extreme perturbation data points from the initial dataset that do not conform to the normal physical operation. For example, the purified sample space may include a relatively continuous and compact matrix of price numerical samples remaining after accurately removing isolated outlier data points identified by the Isolation Forest algorithm from the initial historical price sample data containing noise.

[0062] In some embodiments, the processor can receive the row index identifier sequence of all abnormal data points output from the anomaly detection module; perform row deletion traversal operation on the original matrix storing the initial historical price sample data; erase or remove the data records at the corresponding positions one by one according to the above index identifier sequence; reorganize the index order of the remaining unmarked normal data matrix; and finally obtain the purified sample space after removing interference features.

[0063] S4: Construct a piecewise asymmetric value function to map the dynamic difference signal between the predicted electricity price data sequence and the spot clearing benchmark data curve into adjustment potential energy. Through calculation, obtain the gravitational field strength data that characterizes the direction and intensity of adjustment.

[0064] An asymmetric value function is a computational data structure logic that maps the dynamic difference signal of an input system into an output adjustment parameter with a clear directionality and differentiated sensitivity. For example, an asymmetric value function may include built-in piecewise mapping algorithm logic, which specifies that when a positive deviation signal is received, a concave function characteristic operation is performed to smooth and limit the output value; when a negative deviation signal is received, a convex function characteristic operation is performed to exponentially polarize and amplify the output response value.

[0065] In some embodiments, the processor can obtain the sign characteristics of the input normalized price difference signal by establishing a segmented discrimination tree logic within the control program: when the sign is positive, a power operation subroutine with a sensitivity decreasing exponent parameter is called to limit the expansion of the calculation result; when the sign is negative, a power operation subroutine with a negative deviation polarization response parameter is called to amplify the calculation result, and the code module integrating the above dual-track segmented calculation execution path is encapsulated to obtain the asymmetric value function.

[0066] The dynamic difference signal reflects the real-time physical magnitude of the deviation between the predicted electricity price status data and the expected baseline status data at different time points. For example, the dynamic difference signal can include the absolute difference result obtained by subtracting the predicted electricity price value sequence of each specific time period from the corresponding time point value in the spot clearing benchmark data curve, which serves as a stable reference surface.

[0067] In some embodiments, the processor can perform time series alignment operations on the array of predicted electricity price data sequences loaded into the memory buffer and the array of spot clearing benchmark data curves; using the vector operation module, the processor performs subtraction algebra operations on the values ​​in the predicted array and the corresponding values ​​in the benchmark array under the same time index, and after iterating through all time indexes, the resulting array of difference results reflecting the state offset is used as a dynamic difference signal.

[0068] Adjustment potential energy is a dimensionless quantity derived from the original physical data of price difference deviations through internal nonlinear system mapping operations. It is used to characterize the strength of the system's endogenous tendency to adjust resource allocation. For example, adjustment potential energy can include the output value of the normalized price difference signal processed by the system using an asymmetric value function. The sign and magnitude of this value no longer represent a simple price, but rather map to the mathematical driving energy accumulated in the system that attempts to break the current allocation state.

[0069] In some embodiments, the processor can use the normalized price difference signal calculated by the front-end preprocessing as the input parameter of the independent variable and send it into the encapsulated asymmetric value function execution logic; trigger a specific exponential level power and coefficient scaling operation inside the function; after the operation is completed, extract the value of the calculated mapping transformation result, output the value of the abstract driving force characteristic value that is deviated from the original dimension property and assign and save it to obtain the regulating potential energy.

[0070] Gravitational field strength data is the transformation of an unbounded distribution of internal abstract potential energy into finite force characteristic values ​​that directly guide the magnitude and direction of actions in the solution space through normalization constraint calculations. For example, gravitational field strength data can include continuous values ​​that have been processed by a specific smoothing mapping function and forcibly compressed and constrained to a fixed range of real numbers from -1 to +1. When these values ​​approach the boundary values, it indicates that the system needs to perform extreme intensity adjustment operations, while near zero, it indicates that the system maintains a dead zone stable state.

[0071] In some embodiments, the processor can extract the value of the adjustment potential energy characteristic generated by the system front-end module; then call the processor's underlying mathematical function operation library to start the hyperbolic tangent nonlinear transformation function module; pass the value of the adjustment potential energy characteristic as a parameter into the hyperbolic tangent function for nonlinear spatial compression transformation calculation, and extract and save the smooth result data that is limited to the specified numerical domain after calculation as gravitational field strength data.

[0072] In some embodiments, the processor can combine the predicted electricity price data sequence, the spot clearing benchmark data curve, and the time-series price structure entropy in the multi-source heterogeneous feature matrix to calculate the dynamic decision threshold; use the difference between the predicted electricity price data sequence, the spot clearing benchmark data curve, and the dynamic decision threshold to calculate the normalized price difference signal; input the normalized price difference signal into a piecewise asymmetric value function, limit the signal response amplitude with a gradual increase in the positive deviation interval, and amplify the signal response amplitude with a steep decrease in the negative deviation interval, outputting a dimensionless adjustment potential energy; use the hyperbolic tangent function to map the dimensionless adjustment potential energy into a field strength value within a preset numerical range to obtain gravitational field strength data.

[0073] Dynamic decision thresholds are used to assess the current level of objective background noise in the environment and to construct dynamic boundary data to judge whether system state deviation signals have effective penetration and adjustment value. For example, dynamic decision thresholds may include a comprehensive amplified numerical parameter calculated by combining the average system prediction deviation and superimposing the time-series price structure entropy index, which reflects the inherent volatility of the current physical operation, to filter out weak and meaningless frequent disturbance signals.

[0074] In some embodiments, the processor can obtain a basic judgment base by reading the difference data set between the predicted electricity price data and the spot clearing benchmark data over a period of time, calculating the arithmetic mean of the absolute value of the difference set; at the same time, it can extract the time series price structure entropy feature index data under the corresponding time window; execute multiplication logic to multiply the basic judgment base with a dynamic expansion ratio coefficient containing the feature index and scale it up, and finally calculate and generate a dynamic decision threshold.

[0075] The normalized price difference signal is a relative characteristic significance index obtained by standardizing the ratio of the original system state deviation difference (containing absolute physical units) relative to the current environmental noise floor level boundary. For example, the normalized price difference signal can include dimensionless proportional data generated by dividing the original dynamic difference signal by the dynamic decision threshold. This signal value reflects the ratio of the system's true deviation to the environmental tolerance noise floor.

[0076] In some embodiments, the processor can extract the absolute value of the dynamic difference signal generated at the current moment as the numerator parameter of the dividend by calling the logical division operation unit, and extract the value of the dynamic decision threshold generated under the current operating condition as the denominator parameter of the divisor; execute the division ratio operation logic that eliminates the physical unit dimension, and extract the quotient data features obtained by the operation as the normalized price difference signal output reflecting the significant deviation of the system's relative state.

[0077] In some embodiments, the expression for the dynamic decision threshold is: ; The expression for the normalized spread signal is: ; The expression for the dimensionless regulating potential energy is: ; The expression for gravitational field strength data is: ; in, Indicates the dynamic decision threshold. This represents the data for the corresponding time period in the time-period predicted electricity price data sequence. This represents the data for the corresponding time period in the spot clearing benchmark data curve. This indicates the total number of time periods set based on the observation period. This represents the temporal price structure entropy in a multi-source heterogeneous feature matrix. This represents the normalized spread signal. This represents the dimensionless regulating potential energy. and These represent the sensitivity reduction index, which is preset based on historical data and experience. This represents the negative bias polarization response coefficient preset based on the system's defense response requirements. This represents data on the strength of the gravitational field. Represents the hyperbolic tangent function. Indicates the current time period.

[0078] S5: Extract the physical support boundary parameters of the generator set's bottom layer, and through calculation, obtain the repulsive damping coefficient that changes exponentially as the target power distribution approaches the physical support boundary.

[0079] The underlying physical support boundary parameters of generator sets are mapped from the actual physical operational safety limits of power grid generating equipment and external rigid management rules. They serve as absolute bottom-line threshold data for constraining the limit coordinates of the system resource allocation solution space. For example, the underlying physical support boundary parameters of generator sets may include the maximum adjustable power generation absolute upper limit calculated based on the physical unit nameplate capacity, and the minimum allocation baseline data that must be guaranteed to prevent penalties by rule conversion, forming an absolute data red line that the algorithm cannot cross.

[0080] In some embodiments, the processor can send a read request to an external rules database to extract market management limit ratio and penalty bottom line tolerance data, and simultaneously obtain the real-time physical capacity available data of the generator set through the internal scheduling system; call the computing kernel to perform penalty boundary inversion calculation and multiplication conversion operation on the management limit ratio and the real-time physical capacity available data to obtain a set of absolute power values ​​representing the lower limit and upper limit of the safety allocation extreme value, and store them as the underlying physical support boundary parameters of the generator set.

[0081] The repulsive damping coefficient is a characteristic value dynamically generated by the system in the constructed system dynamic potential field model, used to exponentially offset the attenuation of internal regulating potential energy as the current distribution state approaches the dangerous limit boundary. For example, the repulsive damping coefficient can include a multiplier value floating between zero and one. When the system is in the safe core region, the coefficient takes a value close to one and does not produce a drag effect; once the system is about to break through the red line of the underlying physical support boundary parameter, the coefficient drops sharply to zero, thus forming a strong numerical damping effect.

[0082] In some embodiments, the processor can obtain the repulsive damping coefficient by reading the current system source load coupling saturation deviation characteristic value calculated by the front end; using the absolute value of the saturation value as the independent variable and substituting it into the mathematical equation preset by the system with a specified exponential penalty attenuation parameter; executing the exponential boundary resistance calculation logic based on spatial distance, extracting its attenuation output value and truncating and limiting it for storage.

[0083] In some embodiments, the processor can acquire the allocated power data at the current moment and extract the underlying physical support boundary parameters of the generator set; calculate the degree of closeness of the allocated power data to the underlying physical support boundary parameters to obtain the source-load coupling saturation characterizing the deviation of the data; construct a repulsive field based on the source-load coupling saturation to generate a resistance value that changes exponentially as the saturation approaches the safety constraint boundary, as the repulsive damping coefficient.

[0084] Allocated power data refers to the total physical power volume characteristics within the system that has been confirmed, locked, and assigned to be executed within a specific time slice or the initial state of a single optimization iteration of an algorithm. For example, allocated power data may include the baseline total megawatt-hours value recorded in existing medium- to long-term contracts that have been signed and are in effect, which is read by the system at the initial stage of the operation. This data serves as the starting point for all subsequent resource increase / decrease simulations and boundary distance calculations.

[0085] In some embodiments, the processor can initiate a batch retrieval request to the system transaction contract management database server, and based on the target calculation time interval identifier, retrieve all relevant contract record details whose current status is effective or in the existing execution period; perform cumulative summation operations on the electricity value fields registered in these contract details, and extract and cache the final collected physical total value as the allocated electricity data of the system for the current time period.

[0086] Source-load coupling saturation is a normalized distance characteristic parameter that quantifies the degree to which the current resource allocation obligations of a system approach its underlying objective allowable operating limit or the intensity of congestion. For example, source-load coupling saturation can include a percentage characteristic value generated by further discounting the absolute difference between the current allocated power data and the underlying physical limit capacity boundary relative to the baseline capacity; this value directly touches the sensitive bottom line of system state security.

[0087] In some embodiments, the processor can extract the current allocated power data features from the memory state variables, as well as the corresponding upper or lower limit reference values ​​of the underlying physical support boundary parameters of the generator set; calculate the absolute deviation distance between the two using arithmetic subtraction; then divide the absolute deviation distance value by the total system capacity or a preset spatial proportion constant to eliminate the scale dimension, and output the final relative approximation proportion feature as the source-load coupling saturation.

[0088] A repulsion field is a computational logic coverage area constructed at the edge of the system's data optimization solution space to generate a reverse data penalty and suppression effect. For example, a repulsion field may include a digital space defined by a set of exponentially decaying algorithm functions that are automatically triggered based on the degree to which the system state deviates from the center; as long as the data adjustment instructions for resource allocation cross this warning calculation range, they will immediately encounter strong nonlinear numerical reduction and correction, thereby avoiding physical limits.

[0089] In some embodiments, the processor can configure a threshold monitoring program in a memory-resident process of the control system to monitor the source-load coupling saturation state parameter of the system; when it is determined that the absolute value of the saturation parameter exceeds the preset warning interval start line, the background exponential boundary decay function module is awakened; the module is activated to forcibly intervene in each subsequent operation clock cycle to generate a decay coefficient to correct the total control output, and the entire intervention operation logic in effect is the repulsive field.

[0090] In some embodiments, the expression for the source-load coupling saturation is: ; The expression for the repulsive damping coefficient is: ; in, Indicates the source-load coupling saturation. This indicates the allocated power data at the current moment. This represents the boundary parameters of the underlying physical support of the generator set. This represents the normalized parameters based on the preset capacity of the power grid system. Indicates the repulsive damping coefficient. This represents the function that takes the minimum value. This represents the repulsion index based on the preset physical safety constraint strength. Indicates the current time period.

[0091] S6: Calculate the potential field by superimposing the gravitational field strength data and the repulsive damping coefficient to obtain the optimal power allocation data processing result that satisfies the underlying physical security constraints, thus completing the power allocation data processing.

[0092] Underlying physical safety constraints are a set of objective data boundary conditions set to ensure the normal and stable operation of power grid physical equipment, prevent physical failures and collapses, and meet the minimum requirements of external administrative assessments. For example, underlying physical safety constraints may include the maximum power generation limit characteristic value of hydroelectric generating units that must be restricted due to the need to reserve flood control reservoir capacity, and the minimum physical output guarantee limit value of the units that must be turned on to ensure grid frequency stability.

[0093] In some embodiments, the processor can establish a data interface with the power grid's underlying dispatch center's automated control system and safety anti-misoperation verification module; cyclically read the real-time thermal stability limit value table and minimum operating mode control value list of each unit equipment under the current operating conditions issued by the interface according to a predetermined cycle; convert the format of these immutable key physical limit data and write them into the system's read-only protected memory area, thereby forming effective underlying physical security constraints.

[0094] The optimal power allocation data processing result is a final execution instruction characteristic value generated after multi-dimensional dynamic potential field superposition calculation and multiple boundary resistance filters, which can achieve an adaptive balance between maintaining the bottom line of physical safety and obtaining the optimal solution for value adjustment. For example, the optimal power allocation data processing result may include the absolute value of the final contract adjustment amount, which includes a specific direction indicator and a corrected step size, after a series of virtual optimization gravity-driven and physical repulsive field resistance comprehensive calculations.

[0095] In some embodiments, the processor can read the absolute value of the gravitational field strength and directional characteristics output by the front-end optimization module, simultaneously extract the available spatial data generated by the spatial calculation module, and extract the current repulsive damping coefficient generated by the boundary control module; call the multiplication operation processor unit to comprehensively multiply the above-mentioned absolute value of gravitational strength, spatial margin value, drag coefficient, and directional sign to obtain the controlled correction step size; finally, add the correction step size to the initial state allocated power data to generate and save the final optimal power allocation data processing result. As shown in Table 2, this is the monthly contract transaction result table, which records in detail the specific process data of the algorithm performing potential field superposition calculation in 24 consecutive time periods. The table clearly shows how the system dynamically outputs oscillation-free contract decision adjustment quantities based on the difference space between the current existing contracts and the power generation capacity, combined with the calculated foreground gravitational and boundary repulsive coefficients, and finally obtains the optimal allocation processing result that satisfies the physical operating boundary of the unit.

[0096] Table 2 Monthly Contract Transaction Results

[0097] In some embodiments, the processor can determine the available space data under different adjustment directions based on the target allocation boundary parameters and the currently allocated power data; extract the adjustment ratio features and direction features corresponding to the gravitational field strength data; superimpose and multiply the available space data, adjustment ratio features, direction features and repulsive damping coefficient to calculate the final optimal power allocation data processing result, and complete the power allocation data processing.

[0098] Target allocation boundary parameters are the absolute control values ​​that the system determines, based on the characteristics of the corresponding direction, as not to be crossed when deciding whether to execute a resource action in a specific direction (positive expansion or negative contraction). For example, target allocation boundary parameters may include the "allowed upper limit physical parameter" (retrieved as the upper limit) when the system instruction characteristic indicates a request to increase resource holdings; and the "compliance assessment lower limit physical parameter" (retrieved as the lower limit) when the instruction indicates a sell-off contraction.

[0099] In some embodiments, the processor can obtain the current instruction direction feature sign bit extracted from the front-end gravitational field strength data; execute conditional branch matching retrieval logic: if the sign bit indicates a positive direction, retrieve the upper limit extreme value of the unit's available physical capacity from the underlying physical parameter library; if the sign bit indicates a negative direction, retrieve and calculate the minimum minimum guarantee performance baseline value to prevent penalties and violations from being imposed from the rule parameter library; and assign the retrieved and calculated unidirectional extreme value data to the target allocation boundary parameters.

[0100] Available space data reflects the physical tolerance value of the resource capacity that the system theoretically still has for free scheduling by the algorithm, starting from the currently confirmed resource configuration state base point and moving along the set operation adjustment direction until it touches the corresponding target limit boundary. For example, available space data can include the value of still idle and unused resource quota obtained by subtracting the currently locked allocated power data from the maximum target allocation boundary parameter value in the increase direction.

[0101] In some embodiments, the processor can determine the current logical processing direction based on the extracted directional features, and then call the subtraction operation logic module; if the directional features point to an expansion operation, the extracted positive target allocation boundary parameter value is used as the minuend, and the current allocated power data value is used as the subtrahend to perform a subtraction operation; if the directional features point to a contraction operation, the current allocated power data value is used as the minuend, and the negative target allocation boundary parameter is used as the subtrahend to perform a subtraction operation; the difference value representing the absolute safety margin is calculated and saved as available space data.

[0102] The adjustment ratio feature is extracted from the comprehensive gravitational field strength data sequence containing multidimensional information. It is an absolute numerical feature used purely to measure the strength of the system's core algorithm's willingness to consume and utilize the remaining available space resources. For example, the adjustment ratio feature may include the pure numerical feature of the gravitational field strength after removing the positive and negative direction indicators. This feature's value range is limited to between zero and one, and its magnitude determines the size of the action calculation step under ideal conditions without external obstruction.

[0103] In some embodiments, the processor can read the signed gravitational field strength feature numerical input stream generated by the previous mapping module; apply the processor's built-in mathematical absolute value calculation processing function instruction to the input stream data; and, by filtering out the positive and negative polarity sign bits in the input values, allocate the extracted pure scalar values ​​representing the intensity to a specified system memory space as adjustment ratio feature data output representing the desired adjustment range.

[0104] Directional features are extracted and separated from the control flow of comprehensive gravitational field strength data. They are purely discrete symbolic state flags specifically used to guide the execution trend of subsequent resource allocation commands. For example, directional features may include positive and negative flag states extracted from gravitational field values: positive flag states representing an increase in the amount of physical allocation held during expansion, or negative flag states representing a decrease in the amount of physical occupancy held during contraction, determining whether subsequent calculation operations are superposition or reduction.

[0105] In some embodiments, the processor can receive the raw calculated data of the gravitational field strength output by the front-end module; input it into the polarity state determination function (such as the symbol extraction function module in mathematics) defined by the system kernel to execute the conditional discrimination logic; when the input raw data is greater than zero, output a specific logic value to represent a positive identification state, and when the raw data is less than zero, output another specific logic value to represent a negative identification state, and record the discrimination logic state features of this discretized output as the direction features required for subsequent processing.

[0106] In some embodiments, the expression for available spatial data is: ; The expression for the change in allocation data in the optimal power allocation data processing results is as follows: ; in, Indicates available spatial data. This represents the upper limit parameter for allowed holdings extracted based on the target allocation boundary parameters. This indicates the currently allocated power data. This represents data on the strength of the gravitational field. This represents the lower limit parameter for allowable reduction of holdings extracted based on the target allocation boundary parameters. This indicates the distribution of change data. This represents the repulsive force damping coefficient at the current moment. The sign function indicates that the optimal power allocation data processing result is obtained by summing the currently allocated power data and the allocation change data. Indicates the current time period.

[0107] In some embodiments, to ensure the stable operation of the nonlinear value function and the artificial potential field damping mechanism and to avoid system divergence, the constant parameters involved in the above-mentioned dynamic calculation models must be preset based on the statistical regularities of historical power grid operating data. The specific parameter value ranges and setting criteria are as follows: Sensitivity Decreasing Index and When constructing a piecewise nonlinear value function, and The degree of compression in the positive profit range and the sensitivity in the negative loss range are controlled separately. This is to ensure the convexity and concavity of the mapping curve without causing numerical overflow. and The range of values ​​for all values ​​is limited to and In a preferred embodiment of the present invention, based on backtesting and verification of the smoothness of historical price difference signal response, it is recommended to set... .

[0108] Negative bias polarization response coefficient This coefficient is used to amplify the defensive signal for reducing holdings by a factor of two when a risk of loss is detected. Its value must be greater than 1 to reflect the polarization amplification effect; typically, the value is set within a range of... Between these values. Values ​​that are too small will not provide sufficient defense, while values ​​that are too large can easily trigger an overreaction in the system, causing oscillations in the solution set. In specific implementations of a highly volatile hydropower-dominated market, it is recommended to set... .

[0109] Normalized parameters When calculating the source-load coupling saturation, Used to convert absolute physical charge deviations into dimensionless proportions to match the exponential base requirements of damping functions. The value is usually set to a constant value between 5% and 15% of the total available installed capacity within the target dispatch area of ​​the power grid, which is used to smooth the transition zone of the boundary buffer.

[0110] Repulsion index This index determines the steepness of the increase in the intensity of the virtual repulsive field as the system state approaches the physical boundary. To form an effective dynamic soft wall in front of the safety boundary, It should be a positive number, and its range of values ​​is generally set to 1. The larger the value, the more drastic the increase in drag as the system approaches the boundary, and the more severe the penalty for exceeding the limit.

Claims

1. A power allocation data processing method based on ensemble learning, characterized in that, include: S1: Obtain historical power grid operation status data and market bidding interaction data to construct a multi-source heterogeneous feature matrix; the multi-source heterogeneous feature matrix includes system supply and demand balance indicators; S2: Using the multi-source heterogeneous feature matrix as input, an ensemble learning prediction model guided by physical marginal features is constructed. In the node splitting stage of the decision tree in the ensemble learning prediction model, the system supply and demand balance index is used to guide the ensemble learning prediction model to divide the feature space and obtain the time-segmented predicted electricity price data sequence. S3: Based on historical spot market operation data, the isolated forest algorithm is used to identify and remove outlier data points, and a spot clearing benchmark data curve is generated as a zero-deviation reference benchmark. S4: Construct a piecewise asymmetric value function to map the dynamic difference signal between the predicted electricity price data sequence and the spot clearing benchmark data curve into adjustment potential energy. Through calculation, obtain the gravitational field strength data that characterizes the direction and intensity of adjustment. S5: Extract the physical support boundary parameters of the generator set's bottom layer, and calculate the repulsive damping coefficient that changes exponentially as the target power distribution approaches the physical support boundary. S6: Calculate the potential field by superimposing the gravitational field strength data and the repulsive damping coefficient to obtain the optimal power allocation data processing result that satisfies the underlying physical security constraints, thus completing the power allocation data processing.

2. The power allocation data processing method based on ensemble learning according to claim 1, characterized in that, S1 includes: Extract the clearing price sequence data from the historical power grid operation status data, calculate the ratio of its standard deviation to mean, and obtain the time-series price structure entropy; Obtain the forecast data of the entire network load and the forecast upper limit data of the dispatchable units of the entire network, calculate the ratio of the two, and obtain the system supply and demand balance index. By fusing basic supply and demand load data, market bidding interaction data, time-series price structure entropy, and system supply and demand balance indicators, a multi-source heterogeneous feature matrix is ​​constructed.

3. The power allocation data processing method based on ensemble learning according to claim 1, characterized in that, S2 includes: A stratified sampling technique is used to construct a sample set based on a multi-source heterogeneous feature matrix, thus obtaining a training sample set. During the model training phase, the system supply and demand balance index is used as a splitting feature and the training sample set as input to the integrated learning prediction model. This guides the decision tree to prioritize capturing the critical points of system pricing logic jumps and perform feature space partitioning to obtain leaf node distribution features. By utilizing the leaf node distribution characteristics of the decision tree in the ensemble learning prediction model, a time-segmented predicted electricity price data sequence containing the predicted average price and its confidence interval is generated.

4. The power allocation data processing method based on ensemble learning according to claim 1, characterized in that, S3 includes: Based on the similarity between power grid supply and demand, initial historical price sample data with the same operating characteristics are extracted from the historical spot market operation data of the same period. The isolated forest algorithm is used to identify and remove outlier data points caused by non-steady-state commands in the initial historical price sample data, resulting in a purified sample space. Temporal features are extracted from the purified sample space to generate a spot clearing benchmark data curve that reflects the fluctuation pattern.

5. The power allocation data processing method based on ensemble learning according to claim 2, characterized in that, S4 includes: By combining the predicted electricity price data sequence, the spot clearing benchmark data curve, and the time-series price structure entropy in the multi-source heterogeneous feature matrix, the dynamic decision threshold is calculated. The normalized price difference signal is obtained by calculating the difference using the predicted electricity price data sequence, the spot clearing benchmark data curve, and the dynamic decision threshold; The normalized price difference signal is input into a piecewise asymmetric value function. In the positive deviation range, the signal response amplitude is limited by a gradual increase, while in the negative deviation range, the signal response amplitude is amplified by a steep decrease, and the dimensionless regulating potential energy is output. The dimensionless adjustment potential energy is mapped to the field strength value within a preset range using the hyperbolic tangent function, thus obtaining the gravitational field strength data.

6. The power allocation data processing method based on ensemble learning according to claim 5, characterized in that, The expression for the dynamic decision threshold is: ； The expression for the normalized spread signal is: ； The expression for the dimensionless regulating potential energy is: ； The expression for gravitational field strength data is: ； in, Indicates the dynamic decision threshold. This represents the data for the corresponding time period in the time-period predicted electricity price data sequence. This represents the data for the corresponding time period in the spot clearing benchmark data curve. This indicates the total number of time periods set based on the observation period. This represents the temporal price structure entropy in a multi-source heterogeneous feature matrix. This represents the normalized spread signal. This represents the dimensionless regulating potential energy. and These represent the sensitivity reduction index, which is adjusted based on historical data and pre-set based on experience. This represents the negative bias polarization response coefficient preset based on the system's defense response requirements. This represents data on the strength of the gravitational field. Represents the hyperbolic tangent function. Indicates the current time period.

7. The power allocation data processing method based on ensemble learning according to claim 1, characterized in that, S5 includes: Obtain the allocated power data at the current moment and extract the underlying physical support boundary parameters of the generator set; The approximation of the allocated power data to the underlying physical support boundary parameters is calculated to obtain the source-load coupling saturation, which characterizes the deviation of the data. A repulsive field is constructed based on the source-load coupling saturation, generating a drag value that changes exponentially as the saturation approaches the safety constraint boundary, which serves as the repulsive damping coefficient.

8. The power allocation data processing method based on ensemble learning according to claim 7, characterized in that, The expression for the source-load coupling saturation is: ； The expression for the repulsive damping coefficient is: ； in, Indicates the source-load coupling saturation. This indicates the allocated power data at the current moment. This represents the boundary parameters of the underlying physical support of the generator set. This represents the normalized parameters based on the preset capacity of the power grid system. Indicates the repulsive damping coefficient. This represents the function that takes the minimum value. This represents the repulsion index based on the preset physical safety constraint strength. Indicates the current time period.

9. The power allocation data processing method based on ensemble learning according to claim 1, characterized in that, S6 includes: Based on the target allocation boundary parameters and the current allocated power data, determine the available space data under different adjustment directions; Extract the adjustment ratio and directional features corresponding to the gravitational field strength data; The available spatial data, adjustment ratio characteristics, directional characteristics, and repulsive damping coefficient are superimposed and multiplied to calculate the final optimal power allocation data processing result, thus completing the power allocation data processing.

10. The power allocation data processing method based on ensemble learning according to claim 9, characterized in that, The expression for the available spatial data is: ； The expression for the change in allocation data in the optimal power allocation data processing results is as follows: ； in, Indicates available spatial data. This represents the upper limit parameter for allowed holdings extracted based on the target allocation boundary parameters. This indicates the allocated power data at the current moment. This represents data on the strength of the gravitational field. This represents the lower limit parameter for allowable reduction of holdings extracted based on the target allocation boundary parameters. This indicates the distribution of change data. This represents the repulsive force damping coefficient at the current moment. The sign function indicates that the optimal power allocation data processing result is obtained by summing the currently allocated power data and the allocation change data. Indicates the current time period.

Images