A Source-Network-Load-Storage Scheduling Control Method and System Based on Deep Reinforcement Learning

By employing a deep reinforcement learning-based scheduling and control method, the problem of supply-demand imbalance caused by fluctuations in renewable energy output and load changes has been solved. This has enabled intelligent scheduling and safe and compliant control within the park, improving energy consumption efficiency and the adaptive capabilities of scheduling.

CN123073883APending Publication Date: 2026-11-10HUNAN ZHONGKECHU NEW ENERGY TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611259274.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-19
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing dispatching methods are unable to effectively cope with the random fluctuations in renewable energy output and the dynamic changes in load demand, resulting in power supply and demand discrepancies. Furthermore, they lack safety and compliance guarantees for dispatching actions, which may exceed physical constraint boundaries and make it difficult to identify and correct repeated violations.

Method used

A scheduling control method based on deep reinforcement learning is adopted. Through multi-timescale modeling, error mutation feature recognition, feature fusion, non-uniform dimensionality reduction and composite reward design, a scheduling policy network is constructed to perform scheduling action detection and safety correction, and generate compliant scheduling control instructions.

Benefits of technology

It enables intelligent collaborative scheduling of the park's power generation, grid, load, and storage operation status, improves the level of renewable energy consumption, ensures the safety and compliance of scheduling instructions, avoids power and load exceeding limits, and improves the adaptive capability and reliability of scheduling.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123073883A_ABST
    Figure CN123073883A_ABST
Patent Text Reader

Abstract

This invention discloses a source-grid-load-storage scheduling control method and system based on deep reinforcement learning. It constructs an error-aware state representation by identifying error mutations and fusing features across multiple time scales in multi-energy prediction data of a power park. Non-uniform dimensionality reduction is implemented based on the proactive response capability to error mutations according to each feature dimension, and priority weights are assigned to high-fluctuation operating intervals. A composite reward function is constructed using the quantification of curtailment margin as the core, combined with differentiated incentive weights during critical absorption periods, to drive gradient training of the scheduling strategy network. Power and load-based limit exceedance detection is performed on the scheduling actions output by the strategy network. A compliant instruction sequence is generated by matching preventative correction vectors with dynamic constraint boundaries. Scheduling control instructions are output through experience replay sampling of limit exceedance priorities and attribution-oriented selective soft updates, improving the absorption level of renewable energy in the power park while ensuring scheduling safety and compliance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of power dispatching technology, and in particular to a source-grid-load-storage dispatching and control method and system based on deep reinforcement learning. Background Technology

[0002] As the proportion of photovoltaic and wind power in the park's power structure continues to increase, the operating status of each link of source, grid, load and storage exhibits high uncertainty and strong coupling characteristics. Random fluctuations in renewable energy output and dynamic changes in load demand continue to cause power supply and demand deviations. The dual constraints of energy storage charging and discharging boundaries and tie line transmission capacity further narrow the dispatchable domain. Achieving coordinated and efficient source, grid, load and storage dispatch under multiple physical constraints is the core challenge currently facing energy management in the park.

[0003] Existing scheduling methods are mainly divided into two categories: heuristic rule-based methods and model optimization methods. The former has limited adaptability to complex operating scenarios and is difficult to effectively cope with renewable energy access scenarios with strong randomness. The latter relies on accurate modeling of the grid operating status, and the reliability of the optimization results is difficult to guarantee when there is a deviation between power prediction and actual output. More importantly, both types of methods lack adaptive guarantee mechanisms for the safety and compliance of scheduling actions. In actual execution, scheduling instructions may break through the physical constraint boundaries of power transmission or energy storage charging, and it is difficult to identify and correct repeated constraint violations. Summary of the Invention

[0004] This invention discloses a source-grid-load-storage scheduling control method and system based on deep reinforcement learning. It aims to achieve intelligent collaborative scheduling of the source-grid-load-storage operation status of the park through error-aware multi-scale feature construction, composite reward design oriented towards consumption, safety correction of scheduling actions, and attribution-oriented strategy-oriented optimization. This will improve the level of renewable energy consumption while ensuring the safety and compliance of scheduling instructions.

[0005] The first aspect of this invention proposes a source-network-load-storage scheduling and control method based on deep reinforcement learning, comprising: Collect source-grid-load-storage operation data of the park and model it according to multiple time scales to form a prediction set. Based on the prediction set, identify error mutation features to form error mutation markers. Perform feature fusion on the prediction set and the error mutation markers to form a state space. Non-uniform dimensionality reduction is performed on the state space and the error mutation marker to form a state feature matrix, and the high fluctuation range of the state feature matrix is ​​located to form a priority weight matrix; The state feature matrix and the priority weight matrix are used to quantify the curtailment margin to form a composite reward function. A scheduling policy network is established by performing deep policy gradient training using the state feature matrix and the composite reward function. For the scheduling policy network, scheduling action over-limit detection is performed to determine power over-limit items and charge over-limit items. The constraint deviation of the power over-limit items is analyzed to form a correction vector. Dynamic constraint boundary matching is performed on the charge over-limit items and the error mutation marker to determine compliant actions. Safety correction is performed according to the correction vector and the compliant actions to generate an instruction sequence. Based on the instruction sequence, the scheduling policy network is sampled by replaying the experience of exceeding the limit priority to form a gradient update sequence. The attribution characteristics of the gradient update sequence are analyzed to form an exceeding limit attribution mapping. Based on the exceeding limit attribution mapping, the scheduling policy network is selectively soft-updated to output scheduling control instructions.

[0006] A second aspect of this invention proposes a source-grid-load-storage scheduling and control system based on deep reinforcement learning, comprising: The state construction unit is used to collect source-grid-load-storage operation data of the park and model it according to multiple time scales to form a prediction set. Based on the prediction set, error mutation features are identified to form error mutation markers. Feature fusion is performed on the prediction set and the error mutation markers to form a state space. The weight priority unit is used to perform non-uniform dimensionality reduction on the state space and the error mutation marker to form a state feature matrix, and to locate the high fluctuation range of the state feature matrix to form a priority weight matrix. The policy training unit is used to quantify the curtailment margin using the state feature matrix and the priority weight matrix to form a composite reward function, and to establish a scheduling policy network through deep policy gradient training using the state feature matrix and the composite reward function. The action correction unit is used to detect scheduling action over-limits in the scheduling policy network to determine power over-limit items and charge over-limit items, analyze constraint deviations of the power over-limit items to form a correction vector, perform dynamic constraint boundary matching between the charge over-limit items and the error mutation marker to determine compliant actions, and perform safety correction to generate an instruction sequence according to the correction vector and the compliant actions. The instruction output unit is used to perform over-limit priority experience replay sampling on the scheduling policy network based on the instruction sequence to form a gradient update sequence, analyze the attribution characteristics of the gradient update sequence to form an over-limit attribution mapping, and perform selective soft updates on the scheduling policy network based on the over-limit attribution mapping to output scheduling control instructions.

[0007] The beneficial effects of this invention are reflected in the following points: First, for multi-energy operation data in the park, by using multi-timescale prediction modeling and identifying prediction error mutation events, the temporal distribution information of errors is integrated into the construction process of state features, enabling the state representation to have targeted perception capabilities for high-fluctuation operation intervals. The non-uniform dimensionality reduction strategy allocates compression weights differently based on the anticipatory response capability of each feature dimension to error mutations, retaining the feature information with the most early warning value for renewable energy output disturbances while compressing the state dimensions, and then using the priority labeling of high-fluctuation intervals to provide scenario-sensitive weight guidance for subsequent training sampling. Second, a composite reward function oriented towards consumption is designed with the quantification of curtailment margin as the core, identifying the critical period close to full consumption and assigning differentiated incentive weights, guiding the scheduling strategy network to prioritize learning the action selection pattern of high consumption value periods; the joint drive of consumption incentives and constraint compliance penalties enables the strategy network to establish a basic constraint awareness of safety boundaries while pursuing renewable energy consumption benefits, and the two types of objectives form a dynamic balance during the training process. Finally, before outputting the scheduling action, power-related and charge-related over-limit behaviors are independently detected. A preventive correction vector is generated by combining frequent over-limit pattern recognition and forward-looking bias prediction. Compliant scheduling actions are determined through dynamic constraint boundary matching, thus forming a safety correction guarantee in the instruction issuance stage. In the policy update stage, a priority experience playback mechanism is constructed based on the historical distribution of over-limit events. Gradient attribution analysis is used to locate the parameter dimension that contributes the most to over-limit behavior in the policy network. Selective soft update is used to implement targeted correction for high attribution dimensions, continuously improving the policy network's ability to comply with constraint boundaries while retaining the learned compliant action patterns. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the source-grid-load-storage scheduling and control method based on deep reinforcement learning according to the present invention.

[0009] Figure 2 This is a block diagram of the source-grid-load-storage scheduling and control system based on deep reinforcement learning, as described in this invention. Detailed Implementation

[0010] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0011] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0012] The technical solutions of the embodiments of this application will be described below.

[0013] like Figure 1 As shown, this embodiment of the invention provides a source-network-load-storage scheduling and control method based on deep reinforcement learning, including the following steps S11-S15: Step S11: Collect source-grid-load-storage operation data of the park and model it according to multiple time scales to form a prediction set. Based on the prediction set, identify error mutation characteristics to form error mutation labels. Perform feature fusion on the prediction set and error mutation labels to form a state space.

[0014] Specifically, operational data of the energy source, grid, load, and storage system in the industrial park are collected and modeled at multiple time scales to form a prediction set. The operational data of the energy source, grid, load, and storage system are categorized according to their physical attributes and accessed through the data acquisition channels. These channels simultaneously include static configuration items such as tie-line transmission parameter ledgers (including the allowable power adjustment range and upper and lower limits of transmission capacity), rated charging and discharging power and upper and lower limits of state of charge of energy storage devices, as well as measured time-series channels for each energy category. The acquisition granularity is set to 15 minutes for short-term time scales, 1 hour for medium-term time scales, and daily for long-term time scales. The prediction time domains for these three scales cover 4 hours, 24 hours, and 7 days, respectively. Each scale is modeled independently to adapt to the differences in the dominant mechanisms of prediction errors within different time domains. The short-term model focuses on the response to rapid disturbances, while the long-term model focuses on the overall grasp of trend changes. Photovoltaic power output forecasting uses meteorological irradiance and numerical weather prediction as driving inputs, wind power output forecasting uses near-surface wind speed, wind direction, and pressure gradient as driving inputs, load forecasting uses historical consumption patterns and real-time demand response activation status as characteristic inputs, and energy storage state of charge forecasting is calculated recursively through charge integrals, with the calculation accuracy decreasing over time. This level includes a confidence interval for the recursive cumulative error. For each energy category, forecasting models are established at three time scales and categorized into the corresponding photovoltaic, wind power, load, and energy storage levels in the forecast set. The forecast set uses the time step as the main index, aligning and organizing the forecast sequences from each data source and time scale with the corresponding measured sequences. The four levels of photovoltaic, wind power, load, and energy storage form a nested structure within the forecast set. The measured channels provide a benchmark for subsequent residual calculations, and the confidence interval for the cumulative error of the energy storage level shows a monotonically increasing trend on the time axis.

[0015] Error mutation markers are formed based on the characteristics of error mutations identified in the prediction set. Historical residuals are obtained by combining each prediction sequence in the prediction set with the aligned measured sequence. The mean residual within a window is calculated using a sliding window of 6 time steps. When the jump in the mean residual of adjacent windows exceeds 3 times the sliding standard deviation (the sliding standard deviation is estimated using the residuals within the same 6-step window as samples), it is identified as an error mutation event. The occurrence time and duration are included as location attributes in the marker sequence, and the peak amplitude of the mutation is recorded as a multiple of the standard deviation. Residual mutations in prediction sequences of different energy categories are asynchronous in time. Error mutation markers are detected separately for each category and then aggregated to form a mutation location matrix with energy category and time step as dual indices. The mutation direction is distinguished by positive and negative signs: positive values ​​indicate a sudden increase in the deviation between predicted and measured values, and negative values ​​indicate a sudden increase in the deviation between predicted and measured values. The sharp drop in photovoltaic output caused by strong summer convection forms a typical single-step amplitude jump in the short-term photovoltaic residual sequence, which, together with the increase in air conditioning load prediction deviation caused by cooling during the same period, constitutes the main source of high-density mutation intervals. Multi-source coordinated mutation is executed as an independent detection channel in parallel with the single-source 3-times-standard-deviation criterion. The criterion is that the number of energy categories with mutations in the same time step is no less than two and the mutation amplitude of each category exceeds twice the moving standard deviation of the category. The criterion of this channel is more lenient than that of the single-source channel. It is used to capture coordinated misalignment events that occur simultaneously but where no single source reaches 3 times the standard deviation. Error mutation labels separately mark the multi-source coordinated attributes for such moments. The mutation amplitude is measured by the joint mean of the mutation amplitudes of each sequence. Single-source mutations and multi-source coordinated mutations are stored in the mutation location matrix by distinguishing them with attribute labels.

[0016] In some embodiments, the step of performing feature fusion on the prediction set and the error mutation marker to form a state space includes: performing energy category parsing on the prediction set to generate multi-source prediction components; performing cross-source power balance conflict detection on the multi-source prediction components to generate a conflict component set; performing conflict intensity calibration on the conflict component set and the error mutation marker to generate an aligned feature sequence; and performing time-series-category dual-dimensional fusion on the aligned feature sequence and the multi-source prediction components to construct a state space.

[0017] Energy category analysis is performed on the forecast set to generate multi-source forecast components. Based on differences in energy conversion mechanisms, the forecast set is mapped to independent physical components: solar irradiation-driven photovoltaic power output sequences and wind speed field-driven wind power output sequences belong to different physical mechanisms and are each classified into independent renewable energy source components; the demand-side load forecast sequence is dominated by energy consumption behavior patterns, and its fluctuation rhythm differs from that of the weather-driven power source component, thus constituting a separate demand-side component; the energy storage state of charge (SOC) forecast, which physically represents the energy storage level, differs fundamentally in dimension from the power flow forecast and is classified into the energy storage SOC component; the photovoltaic and wind power components are collectively referred to as renewable energy source components, and the demand-side component is also called the load component when power calculations are involved. While classifying the multi-source forecast components, the rated charge / discharge power parameters of each energy storage SOC component element are retained for subsequent conversion of SOC variation space into power regulation margin. After grouping by energy category, the photovoltaic component exhibits an inverted U-shaped output curve around noon, while the wind power component maintains high output from early morning to dawn. The fluctuation rhythms of the two renewable energy components are asynchronous, and this temporal misalignment is difficult to extract directly from a single composite sequence, but becomes clearly apparent under the classification structure. For example, in summer evenings, photovoltaic output rapidly declines with sunset, while load remains high due to cooling demand, resulting in only a slight net decrease in overall power in the composite power sequence. However, under the multi-source prediction component, the steep slope of the photovoltaic component and the horizontal continuation of the load component form an inverse slope combination, which can be extracted independently and provides clear category-level input for subsequent cross-source conflict detection. After energy category parsing, the time axis of each component is strictly aligned with the original time axis of the prediction set. Multi-scale prediction sequences under the same energy category are distinguished by scale labels, and the differences in time step granularity of predictions at different scales in the prediction set are unified to the same resolution through interpolation.

[0018] Cross-source power balance conflict detection is performed on the multi-source prediction components to generate a conflict component set. At each time step of the multi-source prediction components, the power difference between the renewable power source component (the sum of photovoltaic and wind power components) and the load component is calculated. A conflict moment is identified when the absolute value of the difference exceeds the allowable power adjustment range of the grid interconnect (taken from the aforementioned static configuration item). The power difference, exceedance magnitude, and duration of the conflict moment are included in the conflict component set, forming a conflict intensity sequence indexed by the time step. In the evening, photovoltaic output rapidly declines with sunset while air conditioning cooling load remains high. During this period, the downward slope of the power source component and the maintenance slope of the load component in the multi-source prediction components form a continuously widening power difference. The conflict component set records a persistent strong conflict sequence within this type of continuous slope misalignment period. Continuous conflict periods lasting more than 4 hours correspond to a typical operating scenario highly dependent on energy storage for dispatch. When the renewable power source component prediction is lower than the load component prediction and both are higher than the load component prediction, it is a conflict situation of superimposed bidirectional prediction bias. The conflict component set labels such moments with bidirectional bias attributes to distinguish them from unilateral imbalance situations. The SOC prediction sequence of the energy storage state component in the multi-source prediction component is combined with the rated charge and discharge power parameters to obtain the available power adjustment margin for that time step. Conflict moments that can be mitigated by energy storage are adjusted in intensity level within the conflict component set, while the remaining conflict moments exceeding the energy storage adjustment capacity retain their original intensity values. The conflict component set characterizes the scheduling difficulty of each time period by the joint distribution of conflict intensity and duration. Periods with continuous conflict duration exceeding 4 hours and peak conflict intensity exceeding 50% of the allowable adjustment range of the tie line are separately designated as deep conflict periods. The weighted average of conflict intensity at each time step within a deep conflict period serves as a comprehensive quantitative indicator of the scheduling pressure level for that period.

[0019] Alignment feature sequences are generated based on conflict component sets and error mutation markers to calibrate conflict intensity. Each conflict moment in the conflict component set is aligned with the error mutation marker by time step. The original conflict intensity at overlapping moments is multiplied by an amplification factor to obtain the calibrated conflict intensity. The amplification factor is obtained by linearly mapping the standard deviation of the mutation amplitude: 2 times the standard deviation corresponds to 1x amplification, 6 times the standard deviation corresponds to 3x amplification, and so on, with linear interpolation used in between. Amplification beyond 6x is capped at 3x. The original intensity remains unchanged for non-overlapping conflict moments. The adjusted conflict intensity at each moment after calibration is included in the conflict intensity column of the alignment feature sequence. During severe convective weather, the rapid drop in irradiance caused by thunderstorm clouds simultaneously leads to a sharp inaccuracy in photovoltaic power output prediction. The conflict component set already records a high power deficit intensity during this period, and the amplification effect of the error mutation markers further amplifies the calibrated conflict intensity to the highest level. Therefore, the alignment feature sequence can fully encode the dual risk state of prediction distortion and power conflict simultaneously, rather than simply reflecting the magnitude of the power difference. After calibration, the adjusted conflict intensity and the mutation direction attribute of the error mutation marker at each time step are concatenated according to time step to form a two-dimensional feature structure with time step as the row, i.e., the aligned feature sequence. The mutation positions in the error mutation marker that do not overlap with any conflict time are added as additional information to the corresponding time step of the aligned feature sequence. This type of information indicates that the prediction bias has deteriorated in the period without obvious power conflict. The column structure of the aligned feature sequence is fixed as the calibration conflict intensity column and the corresponding mutation direction column for each energy category. The mutation direction column at pure conflict time is zero, the conflict intensity column at pure mutation time is zero, and both columns at double superposition time are non-zero values.

[0020] The state space is constructed by performing a time- and category-based dual-dimensional fusion of aligned feature sequences and multi-source prediction components. Using each time step as an alignment benchmark, the calibration conflict intensity and abrupt change direction of the aligned feature sequence at each time step are injected into the prediction value vector of the corresponding time step in the multi-source prediction component, forming a conflict-aware comprehensive feature slice for that time step. Category-level fusion is performed within the comprehensive feature slice. The prediction deviations between energy categories are rearranged under the weighting of the calibration conflict intensity, transforming the independent prediction sequences between categories into a conflict-aware joint feature representation. The relative contribution relationship between energy categories is explicitly encoded in this process. During the morning load ramp-up phase in the industrial park, photovoltaic output gradually increases from zero while high-power industrial energy consumption starts simultaneously. The slopes and amplitudes of both components change rapidly. The aligned feature sequence has a high calibration result for the conflict intensity during this period. The fused state space exhibits high-dimensional cross-category feature differences within the corresponding time step. This highly non-stationary feature section is precisely the scheduling scenario that deep reinforcement learning policy networks need to identify. The fusion result of the temporal and categorical dimensions is organized into a two-dimensional feature tensor in the state space, with time step as the first dimension and conflict-aware extended features as the second dimension. The row cross-section of the state space corresponds to the cross-category conflict-aware features at a certain time step, and the column cross-section corresponds to the evolution sequence of a certain energy category in the entire time domain. Both the aligned feature sequence and the multi-source prediction components retain all original dimensions during the fusion process. The state space fully retains all feature information from both input sources, and the feature dimension of the state space is equal to the sum of the feature dimensions of the multi-source prediction components and the feature dimensions of the aligned feature sequence.

[0021] Step S12: Perform non-uniform dimensionality reduction on the state space and error mutation markers to form a state feature matrix, and locate the high fluctuation range of the state feature matrix to form a priority weight matrix.

[0022] In some embodiments, performing non-uniform dimensionality reduction on the state space and the error mutation marker to form a state feature matrix includes: grouping the state space according to feature dimensions to generate a dimensional grouping set; performing error lead responsivity evaluation on the dimensional grouping set and the error mutation marker to generate a lead responsivity ranking; locating high responsivity dimension segments through the lead responsivity ranking to generate a compressed weight distribution; and performing non-uniform dimensionality reduction on the state space based on the compressed weight distribution to form a state feature matrix.

[0023] The state space is grouped and arranged according to feature dimensions to generate a dimensional grouping set. Based on the physical source and information type, the feature axes of the state space map each dimension to five groups: photovoltaic, wind power, load, energy storage status, and conflict sensing. Feature dimensions with similar physical meanings are grouped together. Each group unit corresponds to a semantically similar feature subspace in the state space, and the boundaries between groups are strictly defined by differences in energy mechanisms. The product of the number of time scales and the number of energy categories determines the total number of feature dimensions in the state space. Some dimensions exhibit high linear correlation. If a uniform compression rate is applied to the entire dimensional state space, high-information-density dimensional segments suffer the same information loss as low-information-density dimensional segments, thus destroying the originally important local feature structures. Photovoltaic prediction features are grouped into the photovoltaic group within renewable energy sources, wind power prediction features are grouped separately, and load features, energy storage status features, and conflict sensing features each constitute independent groups. The dimensional grouping set is organized into a structured set using group labels as indexes. The number of dimensions in each group depends on the total number of feature dimensions of the corresponding category in the state space. The photovoltaic group has the most dimensions due to its three-timescale modeling, while the conflict-aware feature group has the second most due to its integration of cross-source conflict intensity and mutation direction information. After grouping, the dimension index and physical meaning label within each group are retained. The consistency of the physical meaning of each dimension within a group is verified and confirmed by data statistics. After each group is determined, a pairwise Pearson correlation coefficient is calculated for all dimensions within the group. Dimension pairs with an absolute correlation coefficient value exceeding 0.85 are marked as high redundancy pairs. The ratio of the number of high redundancy pairs to the total number of dimensions in the group constitutes the internal redundancy rate of the group. The internal redundancy rate serves as a statistical description of the degree of feature independence within the group and, together with the subsequent compression weights, determines the number of principal components retained in the weighted principal component decomposition of the group. Groups with a high internal redundancy rate can appropriately reduce the number of principal components retained, provided that the cumulative variance explained rate meets the target.

[0024] For example, the step of evaluating the error lead response rate and generating a lead response rate ranking for the dimensional grouping set and the error mutation marker includes: performing intra-group autocorrelation decay rate analysis on the dimensional grouping set to generate a decay rate distribution sequence; performing a decay inflection point lead test on the decay rate distribution sequence and the error mutation marker to generate a lead response set; identifying short memory response dimensions based on the lead response set to generate a response intensity ranking; and applying threshold filtering to the response intensity ranking and the dimensional grouping set to generate a lead response rate ranking.

[0025] Within-group autocorrelation decay rate analysis is performed on the dimensional grouped sets to generate decay rate distribution sequences. For each dimension of the time series within the dimensional grouped sets, the autocorrelation coefficient corresponding to each lag step is calculated using the standard sample autocorrelation function in time series analysis (the standard autocorrelation function is an existing technique in time series analysis, and its calculation details are not repeated in this specification). The decay rate of each dimension is defined as the reciprocal of the number of lag steps (unit: steps) corresponding to the first time the autocorrelation coefficient drops below 0.1, with the dimension being steps. -1 A larger value indicates a shorter autocorrelation memory and greater sensitivity to recent disturbances in that dimension. The decay rate distribution sequence groups the decay rate values ​​of each dimension into a statistical distribution vector. For photovoltaic feature dimensions, the autocorrelation decay rate typically has a lag of more than 8 steps under stable irradiance on sunny days, corresponding to a decay rate of less than 0.125 steps. -1 The time lag during frequent cloud cover events is reduced to less than 2 steps, and the attenuation rate exceeds 0.5 steps. -1 The single-dimensional attenuation rate thus spans nearly four orders of magnitude. Load-related characteristic dimensions experience accelerated attenuation during industrial shift changes and slow attenuation during stable nighttime periods. The attenuation rate distribution sequence records the mean and dispersion of attenuation rates for each group. Conflict-aware characteristic groups, due to the integration of multi-energy information, typically exhibit higher attenuation rates, forming the statistical vector with the largest dispersion in the distribution sequence. After obtaining the attenuation rate distribution vectors for each group, a seasonal stability test of the attenuation rate is performed on each group within the dimension group set. The mean difference of attenuation rates for each dimension in different historical quarters is calculated. Dimensions with inter-quarter mean differences exceeding 30% of the overall mean within the group are marked as condition-dependent attenuation dimensions. The attenuation rate of such dimensions exhibits systematic drift with seasonal climate changes. In subsequent attenuation inflection point lead tests, independent quarterly fitting is required to eliminate the impact of seasonal bias on the lead time statistics.

[0026] A leading response set is generated based on the decay rate distribution sequence and error mutation markers, using a decay inflection point lead test. After calculating the first difference of the decay curves in each dimension of the decay rate distribution sequence, the location of local maxima is detected as the decay inflection point. The time difference between the occurrence time of the decay inflection point in each dimension and the time of the mutation event of the same energy category in the error mutation markers is calculated pairwise. Dimensions with a positive time difference (inflection point precedes mutation) and a historical sample size of at least 10 are included in the leading response set, while retaining the median and interquartile range of the lead time. Conflict-aware groups do not have a corresponding single energy category, and their alignment with the error mutation markers is specially handled: their decay inflection points are directly paired with the time of multi-source collaborative mutation events, rather than compared with single-source mutation events. The lead test examines whether the occurrence time of the decay inflection point in each dimension is earlier than the mutation time of the corresponding energy category in the error mutation markers. If the decay inflection point consistently occurs before the mutation time and the time difference is statistically significant, then the dimension is considered to have the ability to lead the error mutation response. Historically, the inflection points of photovoltaic characteristic dimensions tend to occur 2 to 4 time steps before abrupt changes in irradiance. The lead response set records the median lead duration for such dimensions. In error mutation labeling, the timing of multi-source coordinated mutations is independently counted in the test. Dimension groups that can simultaneously provide lead responses to multi-source coordinated mutations are labeled as multi-source coordinated response priority dimensions. The inclusion conclusions of each dimension group are verified by the Wilcoxon signed-rank test to confirm the statistical significance of the lead time difference. Dimension groups with p-values ​​below 0.05 are labeled as statistically significant lead response dimensions, while dimension groups with p-values ​​between 0.05 and 0.1 are retained and labeled with a low confidence level. The lead response set simultaneously records the number of statistically significant lead response events and the total number of mutation events for each dimension group.

[0027] Based on the advance response set, short-memory response dimensions are identified to generate a response intensity ranking. The median advance duration T_i (in steps) of each dimension group in the advance response set is used as the duration factor for the response intensity score. The response occurrence rate F_i (dimensionless proportion) is directly converted from the ratio of the number of statistically significant advance response events retained in the advance response set to the total number of mutation events. The response intensity score is calculated according to S_i = T_i × F_i, where the dimension of S_i is steps. The S_i value of the dimension group marked as having a multi-source coordinated response priority in the advance response set is additionally multiplied by a dimensionless weighting coefficient of 1.5. The advance response weight of multi-source coordinated mutations is higher than that of single-source mutations. The response intensity ranking is obtained by sorting each dimension group in descending order of S_i score. Dimension groups marked as prioritizing multi-source collaborative responses receive a 1.5-fold weighting in the scoring, resulting in higher S_i values ​​than dimension groups with similar parameters but lower confidence in the lead response set. The top-ranked dimension groups possess both long lead duration and high coverage, consistently ranking highly in response intensity rankings. Kernel density estimation is performed on the score sequences used for response intensity ranking to identify natural breakpoints in the score distribution. The score locations corresponding to local minima of the kernel density in the continuous score distribution are the breakpoints. The score intervals on either side of the breakpoint exhibit a clear separation in their distribution patterns. The breakpoint coordinates are added as an additional attribute to the response intensity ranking results. The S_i of the top-ranked dimension group is typically more than three times that of the bottom-ranked dimension group. This difference reflects the significant differentiation in the error lead perception capabilities of feature dimensions from different physical sources in the state space.

[0028] A threshold screening process is applied to the response intensity ranking and dimension grouping sets to generate an advanced response ranking. The partitioning mechanism employs a two-layer cascaded structure: the first layer uses the natural breakpoint of the score distribution in the response intensity ranking as the initial screening threshold. Dimension groups with scores higher than the breakpoint are determined to be responsive dimensions and enter the second layer of grading; dimension groups with scores lower than the breakpoint are directly determined to be non-responsive dimensions and labeled as non-responsive. The second layer further subdivides the responsive dimensions that pass the initial screening into three levels: high response, medium response, and low response, based on the three-thirds of their score subsets (corresponding to the high / medium and medium / low boundaries). Groups in the dimension grouping sets that fail the initial screening, although located at the end in the advanced response ranking, still retain their complete grouping identifiers and physical meaning labels. The weight allocation step directly locates the corresponding group in the dimension grouping set using the grouping labels from the advanced response ranking. This two-layer cascaded partitioning ensures that the advanced response ranking ultimately includes four discrete levels: high response, medium response, low response, and non-response. All groups in the dimension grouping sets have a corresponding level position in the advanced response ranking. High-response hierarchical dimension grouping typically concentrates meteorological and photovoltaic feature dimensions directly related to weather-driven disturbances. This pattern is particularly stable in transitional season samples with variable climate. When the weather pattern is monotonous and there are insufficient historical abrupt events, the scores of each dimension group in the response intensity ranking have low discriminative power and unstable breakpoint positions. In this case, the historical statistical window is expanded to the two most recent full quarters for supplementary estimation. The stability of the hierarchical division is based on the consistency rate of the hierarchical division between adjacent quarters being no less than 70%. If this condition is not met, the statistical window is further expanded until it is satisfied.

[0029] This scheme uses advanced responsiveness ranking to locate high-response dimensional segments and generate compressed weight distributions. The compressed weight represents the variance amplification factor obtained by the group in weighted principal component decomposition; the larger the weight, the more information is retained. In the advanced responsiveness ranking, high-response-level dimensional groups are defined as high-response-level dimensional segments, with a compressed weight set to 0.9 (dominant variance contribution), 0.7 for medium-response levels, 0.5 for low-response levels, and 0.3 for non-response levels. These four weights are weighted according to the number of dimensions in each level of grouping and then normalized to ensure that the overall mean of the compressed weight distribution is consistent with the preset target dimensionality reduction rate. The bottom quartile of dimensional groups corresponds to low-weight segments, and the compressed weights of the middle groups are linearly interpolated between high and low weights based on the advanced responsiveness ranking score. Within the same group, the compression weights of each dimension remain consistent. Non-uniform processing is performed on a group-by-group basis. The compression weight distribution uses the order of leading response as the primary key, mapping the weight values ​​of each dimension group to the corresponding feature dimension index. In historical periods with frequent weather-driven abrupt changes, meteorological feature groups rank higher, and their compression weights are in the high range of the distribution. The final compression weight distribution is a weight vector of the same length as the original feature dimension of the state space. The vector decreases in a step-like manner along the direction of leading response (from high response to non-response). The ratio of the average weight of the high response segment to the average weight of the non-response segment is not less than 2. This ratio constraint serves as an acceptance condition for whether the compression weight distribution can effectively distinguish responses. If the ratio is not met, the weights of the high response segment are increased, and the weights of the non-response segment are decreased until the condition is met. The weight adjustment range is limited to not changing the step level to which each dimension group belongs.

[0030] Non-uniform dimensionality reduction is performed on the state space based on a compressed weight distribution to form a state feature matrix. Each eigenvalue of the state space is multiplied by its corresponding weight value in the compressed weight distribution, and then weighted principal component decomposition is performed. The weight differences amplify the variance contribution of high-response dimensions and compress the contribution of low-response dimensions during decomposition. The number of principal components retained after dimensionality reduction is set to one-third to one-half of the original number of dimensions in the state space, with the specific ratio dynamically determined by the ratio of the area of ​​high- and low-weight segments in the compressed weight distribution. Weighted principal component decomposition is performed independently within each dimensional group. The principal components retained in each group inherit the energy category label of that group (PV / Wind / Load / Status of Energy Storage / Conflict Awareness), ensuring that the columns of the state feature matrix are mathematically principal components and physically traceable to a single energy category, allowing for subsequent statistical analysis grouped by category. The redundancy of multi-timescale prediction features increases with the proximity of timescales. Features at adjacent timescales share principal component directions in the state feature matrix. The larger the step amplitude of the compressed weight distribution, the more significant the difference in variance contribution between high-response and low-response segments, and the more concentrated the retained principal component directions are in the variation space of high-response dimension segments. The number of rows in the state feature matrix is ​​consistent with the number of time steps in the state space, and the number of columns is equal to the number of principal components retained after non-uniform dimensionality reduction. Dimension pairs that were originally highly correlated in the state space are merged and expressed in the state feature matrix. After the state feature matrix is ​​constructed, the variance explanation rate is calculated for each retained principal component. It is required that the cumulative variance explanation rate of the first k principal components is not less than 85%. If this is not met, the number of retained principal components is gradually increased until the target is met. The mean absolute value of the loadings of high-response dimension segments in each principal component loading vector must be higher than that of low-response segments.

[0031] In some embodiments, the step of locating the high-fluctuation intervals of the state feature matrix to form a priority weight matrix includes: performing time-step fluctuation second-order difference calculation on the state feature matrix to generate a fluctuation acceleration sequence; identifying multi-source synchronous acceleration transition segments of the fluctuation acceleration sequence to generate a synchronous fluctuation candidate set; performing multi-source acceleration joint intensity scoring on the synchronous fluctuation candidate set and the fluctuation acceleration sequence to generate an interval priority ranking; and performing weight labeling based on the interval priority ranking and the state feature matrix to form a priority weight matrix.

[0032] A fluctuating acceleration sequence is generated by performing time-step fluctuating second-order difference calculation on the state feature matrix. The eigenvalues ​​of each column of the state feature matrix are first Z-score standardized to eliminate dimensional differences. Then, adjacent time-step differences are calculated for each standardized column to obtain a first-order difference sequence. This adjacent difference calculation is then repeated for the first-order difference sequence to obtain a second-order difference sequence. The absolute values ​​of the second-order differences of each column are taken over each time step, and the mean is aggregated to form a comprehensive acceleration index. The fluctuating acceleration sequence is structured with time steps as rows and the absolute values ​​of the second-order differences of each column (inherited from the energy category labels of the state feature matrix) and the comprehensive acceleration index forming columns. This structure preserves the details of acceleration along each dimension while providing a single-value aggregation perspective. Each column of the state feature matrix corresponds to the evolution curve of a feature dimension across the entire time domain. Large-scale energy-consuming equipment in the park starts and stops synchronously during shift changes. The first-order differences of multiple load-type feature dimensions surge synchronously within the same time step, while the second-order differences exhibit multi-dimensional acceleration in the same direction at this time step. The fluctuation acceleration sequence forms a sharp, multi-source synchronous acceleration peak at this location, which is the core target signal for locating the transition zone. The comprehensive acceleration index reflects the degree of acceleration in the overall state evolution of the state feature matrix. A dynamic threshold (mean plus 1.5 times the standard deviation, with the same dimensions as the original sequence) is calculated based on the historical full-time domain mean and standard deviation. Time steps exceeding the dynamic threshold are marked as significantly accelerated time steps. These markers serve as the initial screening set of candidate peak time steps in subsequent multi-source synchronous acceleration transition zone location. The distribution density of significantly accelerated time steps across the entire time domain varies depending on the season and operational patterns. The number of significantly accelerated time steps during the spring and autumn transition season is usually higher than in the winter and summer seasons, reflecting the frequency of supply and demand state transitions in the park during the climate transition period.

[0033] For example, the step of identifying the multi-source synchronous acceleration transition segment of the fluctuating acceleration sequence and generating a synchronous fluctuating candidate set includes: performing sliding window low-amplitude quiescent statistics on the fluctuating acceleration sequence to generate a quiescent amplitude distribution map; determining the multi-source synchronous deep quiescent superposition interval through the quiescent amplitude distribution map to generate a quiescent candidate set; performing a quiescent mutation response evaluation on the quiescent candidate set to generate a mutation response score; and performing segment labeling on the mutation response score and the fluctuating acceleration sequence to generate a synchronous fluctuating candidate set.

[0034] A sliding window method is used to statistically analyze low-amplitude silence in fluctuating acceleration sequences to generate a silence amplitude distribution map. The sliding window moves across the fluctuating acceleration sequence with a fixed step size. When the 75th percentile of the absolute value of acceleration in all feature dimensions within the window is lower than a preset silence threshold (the silence threshold is the 25th percentile of the absolute value of all accelerations in the fluctuating acceleration sequence throughout the historical time domain), the window is identified as a low-amplitude silence interval. The silence degree score is calculated by combining the proportion of dimensions participating in silence within the window with the mean of the absolute value of acceleration in each dimension. This score is assigned to the silence amplitude distribution map according to the corresponding time step position. Frequent short-term silences during the daytime in summer in fluctuating acceleration sequences usually correspond to intermittent periods of cloud cover. These silences are short in duration and are followed by local photovoltaic fluctuations rather than large-scale transitions. The silence amplitude distribution map distinguishes the distribution width of short-term silences and long-term deep silences. The silence amplitude distribution map describes the silence intensity distribution across the entire time domain with time step as the horizontal axis and silence level score as the vertical axis. Periods with extremely low comprehensive acceleration scores correspond to peak regions in the silence amplitude distribution map. These peak regions mark the periods of calmest acceleration activity and are also the most likely precursors to transition bursts. The score sequence of the silence amplitude distribution map is smoothed by a 4-step moving average and used for peak detection. Consecutive periods with smoothed scores exceeding the 80th percentile of the historical range are recorded as candidate silence windows. Short-term silence (lasting less than 4 steps) and long-term deep silence (lasting more than 8 steps) are counted separately in the statistics of candidate silence windows. The ratio of the two types of silence quantifies the composition ratio of systematic accumulating silence and ordinary low-activity intervals in the current historical period.

[0035] A candidate set of silent data is generated by identifying multi-source synchronous deep silent overlapping intervals using the silent amplitude distribution map. The silent degree scores for each feature dimension in the silent amplitude distribution map are grouped by energy category (inherited from the energy category labels in the state feature matrix columns). A multi-source synchronous deep silent interval is defined as a continuous period where the silent score for each energy category group is simultaneously above the 80th percentile of its historical score and lasts for more than four time steps. Continuous periods meeting these conditions constitute silent candidate units. The silent candidate set stores the start and end times, duration, and the number of energy categories participating in synchronous silent for each candidate unit. During the stable winter weather period in the silent amplitude distribution map, the duration of multi-energy category synchronous silent data is relatively long, and the silent candidate set contains the candidate units with the widest coverage during this period. In summer, with frequent climate fluctuations, the duration of multi-source synchronous intervals in the silent amplitude distribution map is shorter, and the duration of candidate units is generally lower. The synchronous silencing depth of each candidate unit is measured by the geometric mean of the scores of each dimension of silencing. The geometric mean is more sensitive to extremely low-scoring dimensions than the arithmetic mean, and can distinguish between balanced silencing of each energy category and pseudo-multi-source silencing dominated by a single category. Candidate units are sorted into the silencing candidate set in descending order of geometric mean. The joint distribution of duration and geometric mean in the silencing candidate set describes the silencing characteristics of each candidate unit in two dimensions. The candidate units in the high duration and high depth region in the upper right corner are the most typical transition precursor silencing in history.

[0036] A mutation response evaluation is performed on the silent candidate set to generate a mutation response score. An evaluation window is set 8 time steps forward from the end time of each candidate unit in the silent candidate set. The final mutation response score is calculated as M_c = w_1 × n_c × q_c + w_2 × G_c, where M_c is the mutation response score of candidate unit c (a scalar score before normalization, subsequently normalized to the [0,1] interval for threshold division), n_c is the number of dimensions exceeding the dynamic threshold of each dimension within the evaluation window (dimensionless count), q_c is the sum of the threshold exceedance amplitudes of each exceeding dimension (dimension consistent with the second-order difference after Z-score standardization of the state feature matrix, i.e., dimensionless), the product of n_c and q_c constitutes the original mutation response score, G_c is the geometric mean of the silencing depth of the candidate unit (dimensionless), and w_1 and w_2 are weighting coefficients (dimensionless, w_1 + w_2 = 1, initial w_1 = 0.7, w_2 = 0.3). Candidate units with higher levels of deep quiescence typically exhibit larger mutation amplitudes after the quiescence ends. The positive correlation between the two is reflected through a joint weighted average of quiescence depth and mutation amplitude. A typical example of this positive correlation is the strong mutation triggered by the simultaneous deep quiescence of multiple sources in the park before the arrival of a cold wave, followed by a surge in heating load and a sudden drop in photovoltaic power. Candidate units for which no significant accelerated response was observed within the evaluation window received scores close to zero. After normalizing the mutation response scores to the [0,1] interval, candidate units with scores below 0.1 were determined as invalid precursor quiescence and removed from the quiescence candidate set. Candidate units with scores above 0.6 were retained to form a high-confidence precursor quiescence subset. The coverage of transition events corresponding to this subset in historical samples typically exceeds 80% of all transition events. This coverage rate serves as an acceptance criterion for the quality of the mutation response scores.

[0037] A synchronous wave candidate set is generated by segment labeling based on mutation response scores and wave acceleration sequences. Candidate units with mutation response scores exceeding the 60th percentile dynamic threshold are identified as valid precursor silences. Starting from the end of the valid precursor silence, the first comprehensive acceleration peak time step is retrieved from the wave acceleration sequence. The segment between the end of the valid precursor silence and this peak time step is extracted as a transition activation segment and included in the synchronous wave candidate set. Spontaneous acceleration events in the wave acceleration sequence that have not been verified by valid precursor silences are not included in the candidate set. Before the arrival of extreme cold waves, multiple meteorological-related feature dimensions enter a period of sustained silence. After the silence period ends, a sharp drop in photovoltaic power and a surge in heating load occur almost simultaneously. During this period, the wave acceleration sequence exhibits strong acceleration in the same direction across multiple dimensions. The synchronous wave candidate set labels this segment as a high-intensity transition candidate. This screening mechanism ensures that the candidate segments have a traceable historical recurrence regularity. Each transition activation segment in the synchronous fluctuation candidate set stores three attributes: the mutation response score of the corresponding effective precursor quiescence, the number of dimensions participating in synchronous acceleration within the segment, and the average peak acceleration. The Pearson correlation coefficient between the mutation response score of the effective precursor quiescence and the average peak acceleration within the transition activation segment serves as a verification indicator of the overall quality of the candidate set. A correlation coefficient of not less than 0.6 indicates that the scoring system's predictive ability for transition intensity is acceptable. If the coefficient is less than 0.6, the evaluation window length parameter of the mutation response score needs to be backtracked and re-estimated.

[0038] A multi-source acceleration joint intensity score is generated to rank the candidate sets of synchronous waves and the wave acceleration sequences to obtain interval priority. For each candidate segment in the synchronous wave candidate set, the mean peak acceleration value A_c (A_c is the arithmetic mean of the peak acceleration values ​​of each participating dimension in the corresponding time period, and its dimension is consistent with the second difference after standardization of the state feature matrix Z-score, i.e., dimensionless) and the number of steps D_c (unit: steps) for the duration of exceeding the threshold are extracted from the wave acceleration sequence. The joint intensity score is calculated according to J_c = A_c × D_c × k_c. The directional consistency coefficient k_c is a dimensionless correction term, which is 1.5 when the acceleration directions of all participating dimensions are the same, and 1.0 otherwise. The dimension of J_c is steps. The candidate segments are sorted in descending order of J_c to obtain the interval priority ranking. Strongly synchronous transition segments consistently rank high in interval priority ranking due to their numerous participating dimensions and high directional consistency. During summer evenings, photovoltaic output rapidly declines with sunset while air conditioning load remains high. The multi-dimensional, unidirectional acceleration characteristic of the fluctuation acceleration sequence is most typical during this period, and the corresponding candidate segments typically rank highly in J_c scores. The top-ranked candidate segments represent the time intervals with the most concentrated scheduling pressure and the highest risk of multi-energy coordination imbalance within the park. In the interval priority ranking, J_c scores are divided into four priority levels based on quartiles. The first priority corresponds to candidate segments with J_c scores exceeding the 75th percentile. Boundary values ​​for each level are attached to the ranking attribute fields. The ratio of J_c scores between the first and fourth priority segments reflects the degree of differentiation in scheduling pressure within the synchronous fluctuation candidate set. This ratio is typically higher during the summer-autumn transition season than the winter-spring transition season, indicating a more significant pressure gap between extreme and ordinary transition segments during periods of rapid climate change.

[0039] A priority weight matrix is ​​formed by weighting based on interval priority sorting and state feature matrix. In interval priority sorting, the baseline weight is uniformly set to 1 (dimensionless). The time step weight of the first priority segment is set to 3 to 5 times the baseline weight (between 3 and 5), the second priority segment to 2 to 3 times the baseline weight (between 2 and 3), and time steps not ranked in the top tier retain the baseline weight value of 1. The weight multiple of each priority segment is determined by linear mapping based on the J_c scoring range. Weight values ​​are assigned row-by-row according to the interval time range mapped to the corresponding row index in the state feature matrix, resulting in a weight column vector with the same number of rows as the state feature matrix. This weight column vector is referred to as the priority weight matrix in this specification for consistency with the terminology in the claims. The row dimension of the state feature matrix corresponds one-to-one with the time step. Weighting is performed directly on the row index, mapping the interval time range determined by the interval priority sorting to the corresponding row interval in the state feature matrix. Each row within this row interval is assigned the same segment weight value. Each row of the priority weight matrix stores a weight scalar. The peak position of the weight distribution is precisely aligned with the time position of the synchronization transition segment that ranks first in the interval priority ranking. During periods of high-frequency weather changes, the priority weight matrix of the park exhibits a characteristic distribution of multiple consecutive high-weight intervals alternating. The ratio of the global weight mean of the priority weight matrix to the baseline weight quantitatively describes the density of high-pressure dispatch scenarios within the current historical period. A ratio exceeding 1.5 indicates a significant proportion of high-pressure scenario time steps. After normalization, the sum of all elements in the priority weight matrix equals 1. The difference between the sum of the probability quality of time steps at each priority level and their proportion in the total number of time steps in the entire time domain constitutes the weight skewness index. The larger the skewness index, the more significant the deviation between the sampling distribution of the priority weight matrix and the uniform distribution.

[0040] Step S13: Quantify the power curtailment margin using the state feature matrix and priority weight matrix to form a composite reward function, and establish a scheduling policy network by performing deep policy gradient training using the state feature matrix and composite reward function.

[0041] In some embodiments, the step of quantifying the curtailment margin using the state feature matrix and the priority weight matrix to form a composite reward function includes: performing multi-time-period slicing on the state feature matrix to generate time-period operation components; calculating margin sensitivity based on the time-period operation components and the priority weight matrix to generate a margin distribution map; identifying critical periods approaching full consumption based on the margin distribution map to generate a full consumption critical set; and performing consumption-oriented reward weight synthesis based on the full consumption critical set and the time-period operation components to form a composite reward function.

[0042] The state feature matrix is ​​sliced ​​into multiple time periods to generate time period operation components. The time axis of the state feature matrix is ​​divided into four typical time periods according to the intraday power operation pattern: morning peak (6-9 am), midday off-peak (10-2 pm), evening peak (3-8 pm), and nighttime trough (9-5 am). The row feature vectors of the state feature matrix within the same time period are extracted as a subset of the operation features of that time period, and organized into a block matrix structure with time period labels as indexes, i.e., time period operation components. While segmenting the time-period operation component, based on the attribution relationship of each energy category label to the principal components of the state feature matrix (inherited from the energy category labels of the principal components retained in each group during dimensionality reduction), the power sequence of the corresponding renewable power source group and load group is traced back to the original prediction set in reverse order to obtain the renewable power generation sequence and load power sequence for that time period. The SOC prediction sequence of the energy storage state group is combined with the rated charge and discharge power parameters retained in the multi-source prediction component stage. The capacity space from the current SOC to the constraint boundary is divided by the unit scheduling time period length to obtain the energy storage available power margin sequence, which is recorded separately in the charging and discharging directions, corresponding to the energy storage available charging power margin and the energy storage available discharging power margin. In the state feature matrix, photovoltaic output is concentrated during the midday period in spring and autumn, while industrial load may be at its lowest point during the midday break. The time-period operation component groups this type of supply-demand imbalance separately, so that the calculation of curtailment margin is performed separately within each group. After the block division is completed, the homogeneity test of the feature distribution within each time block is performed. The Ljung-Box statistic is used to test the stationarity of the mean of the feature vector at each time step within the block. The time block with a p value lower than 0.05 is further subdivided into two sub-blocks, which are half morning and half evening. The sub-blocks must pass the homogeneity test again before they can participate in the subsequent margin calculation.

[0043] Based on the time-period operation components and priority weight matrix, margin sensitivity calculation is performed to generate a margin distribution map. For each time period, the renewable power generation sequence, load power sequence, and energy storage available power margin sequence are extracted in blocks. First, the supply-demand positive power difference is calculated (when supply exceeds demand, the difference between renewable power generation and load is taken; otherwise, zero is taken). Then, the basic curtailment margin (unit: MW) for each time step is calculated using "basic curtailment margin = energy storage available charging power margin + tie line power output margin - supply-demand positive power difference". The tie line power output margin is defined as the difference between the upper limit of the tie line's allowable power adjustment range and the current tie line power output at that time step (unit: MW, parameter source: tie line transmission parameter ledger described in S11 section 1). The basic curtailment margin reflects the renewable incremental capacity that can still be received at that time step. The basic curtailment margin is multiplied hourly by the corresponding weight value in the priority weight matrix to obtain the pressure-weighted margin sequence. The weighted margin mean and standard deviation for each period are summarized into a two-dimensional statistical structure with the period label as the row index to form the margin distribution map. The priority weight matrix has a larger weight in periods with high volatility, which can appropriately suppress the margin mean of that period after weighting. The weighted margin mean of each period is sorted in ascending order and then classified by quantile. The mean <0.3 times the total period mean is marked as high risk, the mean between 0.3 and 0.7 times the total period mean is marked as medium risk, and the mean >0.7 times the total period mean is marked as low risk. The three-level risk labels and the standard deviation of the margin fluctuation for each period are stored together in the period attribute field of the margin distribution map. High-risk periods with large standard deviations are given an unstable attribute to distinguish them from stable critical periods with stable fluctuations.

[0044] Based on the margin distribution map, critical periods approaching full absorption are identified, generating a full absorption critical set. Full absorption criticality refers to the operational state where the system's absorption capacity is nearly exhausted, and further acceptance of renewable energy would trigger power curtailment. In the margin distribution map, when the average margin value for each period is less than 30% of the overall mean (i.e., <0.3 × overall mean) and the standard deviation of margin fluctuation is less than 50% of the average margin value (i.e., <0.5 × average margin value), this period is identified as a stable critical period and included in the full absorption critical set. Periods where the average margin value is also less than 30% of the overall mean but the standard deviation of fluctuation is not less than 50% of the average margin value are identified as unstable critical periods, included in the full absorption critical set along with the stable critical periods, but with a separate secondary weight to distinguish them. The priority coefficient is calculated using max(min(1-mean margin / overall mean, 1), 0) and takes values ​​in the interval [0, 1] (dimensionless). A lower bound truncation ensures the priority coefficient is 0 when the mean margin is greater than the overall mean, and an upper bound truncation ensures the priority coefficient is 1 when the mean margin is negative (i.e., the period is actually in an overloaded operating state). The lower the mean margin, the higher the priority coefficient. During the midday solar full-power generation period in spring and autumn, coinciding with industrial lunch breaks, the energy storage margin is usually largely consumed by other scheduling actions. The margin distribution map records the lowest mean margin during this period, and the full absorption critical set is recorded most frequently during this period, representing the most concentrated period for obtaining priority absorption incentives. The full absorption critical set stores the mean margin and corresponding priority coefficient for each critical period. Unstable critical periods are assigned a secondary priority coefficient (60% of the primary coefficient), distinguishing them from the complete incentive amplification effect of stable critical periods. The priority coefficient distribution of stable critical periods shifts upward as the historical renewable energy penetration rate increases. The full absorption critical set is refreshed once before the start of each scheduling day based on the latest margin distribution map. Periods with priority coefficient changes exceeding 0.1 are marked separately in the update record.

[0045] A composite reward function is formed by synthesizing reward weights oriented towards consumption based on the full consumption critical set and the time-period operation components. The priority coefficient of each critical time period in the full consumption critical set is used as the time-period amplification coefficient for consumption incentives. It is multiplied by the actual renewable energy consumption amount (unit: MWh, normalized to the dimensionless interval [0,1] after the maximum consumption amount of the whole time period) in the time-period operation component to obtain the weighted consumption benefit C_weighted. The constraint penalty term P_penalty is obtained by weighting the power limit exceedance and the charge limit exceedance separately and summing them, and normalizing them to the interval [0,1] after their respective historical peak values. The power limit exceedance weight is higher than the charge limit exceedance weight to prioritize grid security. The time-step reward value of the composite reward function is calculated as R=α×C_weighted-β×P_penalty, where α and β are both positive dimensionless coefficients, representing the relative priority of consumption incentives and constraint penalties, respectively. The initial ratio is set to 3:1. The two normalization processes ensure that the ratio of α and β only reflects the priority and does not involve dimension conversion. The absorption incentives for non-critical periods are calculated using the basic weights, while the absorption incentives for critical periods are amplified using the priority coefficients of the full absorption critical set. The period execution component provides the actual absorption amount for the corresponding period as the basis for calculating C_weighted. The constraint violation rate is defined as the proportion of time steps with power or charge exceeding limits in recent training rounds out of the total number of time steps. When this proportion exceeds 5%, α / β is narrowed to 2:1 to strengthen the safety constraint penalty weights. When it is below 1% and the absorption rate stagnates, the ratio is restored to 3:1. Each adjustment does not exceed 0.5 to prevent abrupt changes in the reward function structure from affecting gradient stability. The above α and β coefficients, together with the calculation rules for C_weighted and P_penalty, constitute the composite reward function. The composite reward function outputs a scalar reward value R for the state-action combination at each time step and saves it in the time step sequence for use during the policy gradient training phase.

[0046] A scheduling policy network is established through deep policy gradient training using a state feature matrix and a composite reward function. The scheduling policy network employs a three-layer fully connected deep neural network. The input layer dimension equals the number of columns in the state feature matrix, and the output layer, activated by Tanh, forms a Gaussian policy in a continuous action space. Training is performed using an Actor-Critic algorithm (typically the PPO algorithm) based on policy gradients. The discount factor γ for the cumulative discount reward is set to 0.95 (dimensionless), and exploration is naturally achieved through random sampling of the action distribution. The Actor-Critic algorithm is chosen because subsequent policy updates require importance sampling of priority experience for exceeding limits, necessitating a supporting algorithmic framework for off-policy correction. The row vectors of the state feature matrix at each time step serve as the environmental state input to the network. The network outputs a joint probability distribution of energy storage charging / discharging power, demand response activation, and connection line power adjustment. After the agent performs sampled actions, the immediate reward is calculated using the composite reward function. Policy parameters are updated along the gradient direction of the cumulative discount reward, with each iteration increasing the probability of high-scoring actions and decreasing the probability of low-scoring actions. The incentive amplification effect of the composite reward function during the critical period of full absorption leads the network to develop a specific optimization tendency for near-full absorption scenarios. The strongest incentive signal is obtained during the midday peak of photovoltaic power generation in spring and autumn, which overlaps with the industrial lunch break. In such scenarios, the entropy value of the policy action distribution narrows continuously to a stable low level as the training rounds progress, which reflects that the network's preference for action selection during high-value periods tends to be certain. Convergence is determined based on the simultaneous satisfaction of two conditions: the moving standard deviation of the mean absorption rate is less than 0.5% and the constraint violation rate is less than 2% in the past 200 training rounds. Entropy narrowing is only used as an auxiliary monitoring. After convergence, a generalization test is performed on seasonal climate scenarios not covered by the training set. If the deviation of the absorption rate and constraint violation rate in the test period exceeds 10% compared with the training period, it indicates the risk of overfitting, and the training data needs to be expanded and retrained.

[0047] Step S14: For the scheduling strategy network, perform scheduling action limit detection to determine power limit exceedance items and charge limit exceedance items. Analyze the constraint deviation of the power limit exceedance items to form a correction vector. Perform dynamic constraint boundary matching on the charge limit exceedance items and error mutation markers to determine compliant actions. Implement safety correction according to the correction vector and compliant actions to generate an instruction sequence.

[0048] Specifically, the scheduling policy network is used to detect and identify power and charge limit violations in the scheduling actions. The action vector sampled from the joint probability distribution at the current time step consists of three components: energy storage charging / discharging power, demand response activation, and tie-line power adjustment. This vector is referred to as the scheduling action vector in this step. This vector is compared sequentially with the physical constraint boundaries. Action components exceeding the upper and lower limits of tie-line transmission power or the maximum value of energy storage charging / discharging power are extracted as power limit violations, and the violation component identifier (represented by a combination of the action vector component index and dimension name) and the magnitude of the exceedance are recorded. Actions that are expected to cause the state of charge to exceed the upper and lower limits after the execution of charging / discharging power actions are extracted as charge limit violations, and the magnitude of the command value and the deviation of the state of charge from the constraint boundaries are retained. The physical ratings of power constraints are derived from the equipment parameter ledger. Charge constraints include the upper and lower limits of the state of charge (SOC) of the energy storage equipment (the specific values ​​for both types of constraints are taken from the aforementioned static configuration items). If the charging / discharging power actions output by the scheduling strategy network cause the SOC to exceed the constraint range, the corresponding actions are extracted as charge exceedance items. Exceedance detection is performed dimension-by-dimensionally on the scheduling action vector. During peak winter periods, after multiple rounds of continuous discharge by the energy storage equipment, the SOC approaches the lower limit. If the scheduling strategy network still outputs a large discharge command, a charge exceedance item is generated. During peak photovoltaic periods, if the interconnection line continues to increase its output power while operating at full load, a power exceedance item is generated. Both types of exceedance items can coexist within the same time step. Exceedance detection results are categorized into three types based on the number of power exceedance dimensions: single-dimensional exceedance, two-dimensional exceedance, and multi-dimensional exceedance. Multi-dimensional exceedance time steps synchronously trigger alarm flags, which are then included in the exceedance severity field of the command sequence.

[0049] In some embodiments, the step of forming a correction vector for the analytical constraint deviation of the power limit violation term includes: determining the limit violation direction of the power limit violation term to generate a limit violation component set; identifying recurring limit violation patterns through the limit violation component set to generate a frequent limit violation marker; performing forward-looking deviation prediction on the frequent limit violation marker to generate a correction priority sequence; and performing preventive action bias correction on the correction priority sequence and the limit violation component set to form a correction vector.

[0050] Power limit violations are assessed for directional violation, generating a limit violation component set. The original value of each violation component within a power limit violation item is subtracted from its corresponding constraint boundary value to obtain the limit violation deviation. A positive difference is marked with a positive limit violation symbol, and a negative difference with a negative limit violation symbol. The limit violation component set stores one record for each violation component, including its identifier, limit violation deviation value, and direction symbol, and is sorted in descending order of relative limit violation magnitude (the ratio of limit violation deviation to constraint range). Limit violations involving tie-line power exceeding the upper limit occur frequently during large-scale photovoltaic grid connection periods. In these cases, the correction direction for reduced purchased electricity is negative. The number of records for tie-line overload violations is significantly higher during peak photovoltaic seasons than in other seasons. Simultaneously, if the energy storage charging and discharging power does not exceed the limit, there is no corresponding record in the limit violation component set; only the constraint dimension of the power limit violation is included. The number of limit violation components is highest during periods of simultaneous fluctuations in multiple energy sources, and the number of records in the limit violation component set for these periods is significantly higher than for periods of single-energy fluctuations. When storing the out-of-limit deviation values, the out-of-limit component set simultaneously calculates the relative out-of-limit amplitude of each component. The larger the relative out-of-limit amplitude, the higher the urgency of the out-of-limit for that component. After all records are included, consistency statistics are performed on the out-of-limit direction. When the number of positive out-of-limit records and the number of negative out-of-limit records are both non-zero, a bidirectional conflict flag is marked. A bidirectional conflict indicates that there are both excessively large and excessively small components in the scheduling action vector within the same time step. The positive and negative sides need to be biased in opposite directions and cannot be merged. The bidirectional conflict flag and the maximum relative out-of-limit amplitude of each side are stored in the global attribute field of the out-of-limit component set.

[0051] Frequent over-limit markers are generated by identifying recurring over-limit patterns through over-limit component sets. The frequency of over-limit records for each constraint dimension in the over-limit component set is statistically analyzed over the most recent 30 time steps. If the number of over-limit occurrences for the same constraint dimension exceeds 10, it is identified as a frequent over-limit dimension. The frequent over-limit marker records its identifier, frequency direction, and recent average over-limit amplitude. Energy storage charging and discharging power frequently bottoms out during continuous peak periods in winter. Such frequent over-limits are represented by high-frequency repetitions of negative over-limits in the over-limit component set statistics. The frequent over-limit marker indicates a negative frequent over-limit for this dimension and records its average over-limit amplitude. The frequent over-limit marker is contrasted with low-frequency over-limit records that appear only once or twice in the over-limit component set. Low-frequency over-limits are not included in the frequent over-limit marker and only undergo single-time post-event correction. The distinction between frequent and low-frequency over-limits reflects a differentiated handling strategy for structural deviations and occasional disturbances. The number of dimensions covered by the frequent over-limit marker is usually far less than the total number of records in the over-limit component set. After the frequent occurrence marking is completed, the co-occurrence rate of each frequent occurrence dimension is calculated for each time step of the limit violation. The historical frequency of any two frequent occurrence dimensions violating the limit at the same time step is calculated. Dimension groups with a co-occurrence rate of more than 60% are marked as associated frequent occurrence groups. The average limit violation amplitude of each dimension in the associated frequent occurrence group is represented by the maximum value in the group. The correction priority of members in the group is increased by additional weighting in the subsequent correction priority sequence calculation. The identifier and co-occurrence rate of the associated frequent occurrence group are stored in the group-level attribute field of the frequent limit violation marking. The more associated frequent occurrence groups there are, the stronger the cross-dimensional synergy of the current limit violation mode, which is usually related to the systemic reason that multiple types of constraints are under pressure at the same time under a specific operating pattern.

[0052] A forward-looking bias prediction is conducted to generate a correction priority sequence for frequent limit violation markers. The historical limit violation magnitude time series of each frequent limit violation dimension in the frequent limit violation markers is fitted with a trend line by linear regression. The trend line is extrapolated to the next 3 to 5 time steps. The area integral between the extrapolated segment and the constraint boundary (the dimension corresponding to the limit violation dimension for power type is MW·step, and the dimension corresponding to the limit violation dimension for charge type is percentage of state of charge·step) is normalized to the dimensionless interval [0,1] by the maximum value of the area integral of all historical frequent violation dimensions, and is used as the forward-looking risk score E_i (dimensionless) for that dimension. Each frequent limit violation dimension is arranged in descending order of E_i to obtain the correction priority sequence. The E_i of the associated frequent violation members is multiplied by an additional association weighting coefficient of 1.3 before normalization, so that the relatively independent frequent violation dimensions are moved forward in the ranking. In the frequent limit-over-limit markers, a positive slope in the trend line for positive frequent limit-over-limit dimensions (such as continuous exceedance of tie-line power limits) indicates that the deviation is still widening, and the forward-looking risk score is high in the correction priority sequence. Similarly, for negative frequent limit-over-limit dimensions (such as continuous bottoming out of energy storage state of charge), trend extrapolation is performed to assess the continued limitation of subsequent discharge capacity. With the continuous improvement of photovoltaic output in the spring, the extent of tie-line power exceeding the upper limit may show a gradual widening trend. The E_i of this dimension in the correction priority sequence increases with the increase of the trend slope, reflecting the cumulative pressure on tie-line constraints under the background of continuous growth in photovoltaic output. The priority sequence for correction is updated on a rolling basis according to the scheduling time step. Each new record of exceeding the limit is added to the historical amplitude time series and then linear regression is re-executed. Dimensions with a negative trend line slope and whose limit exceedance amplitude has decreased for three consecutive steps in the recent period are automatically marked with a convergence flag and moved to the end of the sequence. The convergence flag is retained in the sequence, but the corresponding E_i value is set to zero when calculating the bias amplitude, ensuring that the constrained dimensions that have become stable no longer occupy correction resources. The proportion of the number of dimensions holding the convergence flag to all frequently exceeding the limit dimensions reflects the overall degree of improvement in the current limit exceedance pattern.

[0053] Preventive action bias corrections are applied to the correction priority sequence and the out-of-limit component set to form a correction vector. The bias correction magnitude for each out-of-limit dimension in the correction priority sequence is calculated as B_i = 0.6 × E_i + 0.4 × O_i, where E_i is the forward-looking risk score of that dimension normalized to [0,1] in the correction priority sequence, and O_i is the normalized value of the current out-of-limit deviation of that dimension in the out-of-limit component set (with the constraint range as the denominator, taking values ​​[0,1]). The coefficients 0.6 and 0.4 respectively reflect the relative contribution weights of forward-looking risk and current deviation to the bias magnitude. The dimensions of B_i are consistent with those of O_i (dimensionless proportion, which is subsequently multiplied by the constraint range to restore the power value). The bias direction is strictly opposite to the direction of frequent out-of-limit occurrence for each dimension. The correction vector stores the corresponding bias amount with each scheduling action dimension as the index, and the corresponding position of the dimension not selected for correction is set to zero. Dimensions ranked lower in the correction priority sequence receive smaller biases. If there are no over-limit records for that dimension in the over-limit component set at the current time step, then O_i is zero. The bias magnitude is determined solely by the forward-looking risk score. During periods of full photovoltaic power generation, tie-line power frequently exceeds the upper limit. The current over-limit deviation recorded in the over-limit component set and the trend risk score of the correction priority sequence jointly determine the bias amount for tie-line power reduction. The negative correction amount for the tie-line dimension in the correction vector guides the scheduling action to proactively preserve power margin. After the correction vector is calculated, a feasibility verification is performed. The combined value after adding the bias to the original action is resubmitted into all physical constraints for checking. If there are still over-limits, the bias is increased proportionally until the combined value falls within the 2% margin inside the constraint boundary (2% margin refers to 2% of the constraint range corresponding to the over-limit dimension. For example, when the tie line power constraint range is ±100MW, 2% margin corresponds to a position 2MW away from the upper and lower boundaries). If increasing the bias causes a new over-limit, the dimension that causes the new over-limit is removed from the bias target of the correction vector and iterated again. The upper limit of the iteration is 3 times. If the feasibility verification still fails after exceeding the upper limit of the iteration, a correction failure flag is recorded and all correction vectors at that time step are set to zero, and the state of no preventive correction is rolled back.

[0054] In some embodiments, the step of performing dynamic constraint boundary matching to determine compliance for the charge violation item and the error mutation marker includes: performing charge level segmentation to generate a charge classification sequence for the charge violation item; performing a backtracking of the charge violation history to generate a charge violation history weight for the charge classification sequence and the error mutation marker; using the charge violation history weight to locate high-risk charge segments and generate a boundary adjustment set; and performing constraint boundary remapping to determine compliance for the boundary adjustment set and the charge violation item.

[0055] Charge level violations are segmented to generate a charge level sequence. The current state of charge value recorded in the charge level violation item is divided into three segments: high charge segment (above 80%), medium charge segment (40% to 80%), and low charge segment (below 40%). The threshold values ​​for these three segments are taken from the typical recommended values ​​in the energy storage equipment operation specifications for dividing the shallow charge and shallow discharge safety zone and the deep charge and deep discharge risk zone (80% is the upper limit of deep charge safety and 40% is the lower limit of deep discharge safety). The specific values ​​can be adjusted by the user according to the battery chemistry characteristics and life management strategy of the energy storage equipment. Each violation record is labeled with its segment level and violation direction (upper limit violation or lower limit violation). The charge level sequence is saved in groups according to the segment labels. Exceeding the limit in the high-charge segment is usually caused by excessive charging power, leading to exceeding the upper limit. During the midday summer, when photovoltaic power is excessive and energy storage needs to charge in large quantities, the charge state is already close to full charge. In this scenario, the high-charge upper limit exceeds the limit frequently. The charge classification sequence labels such records with the upper limit exceedance attribute. Exceeding the limit in the low-charge segment is usually caused by continuous discharge, leading to the lower limit exceeding the limit. The charge classification sequence labels such records with the lower limit exceedance attribute. After grouping each segment, the frequency and average magnitude of exceeding the limit within each group are statistically analyzed. The ratio of the upper limit exceedance frequency in the high-charge segment to the lower limit exceedance frequency in the low-charge segment is recorded as the charge imbalance index. When the imbalance index exceeds 2, it indicates an asymmetric exceedance characteristic (indicating that the charging strategy of the scheduling strategy network is aggressive while the discharging strategy is conservative); when it is below 0.5, the direction is reversed. The charge imbalance index and asymmetric exceedance characteristics are stored in the global attribute field of the charge classification sequence and serve as a priori reference for asymmetric weight allocation in the calculation of the historical exceedance weights.

[0056] Historical weights for exceeding limits are generated by backtracking the charge level sequence and error mutation markers. Using the charge limit exceeding time step as an anchor, a backtracking window of 3 steps is traced forward (a 3-step window corresponds to the typical time delay steps in which the impact of the error mutation event is transmitted to the charge state exceeding limit through energy storage charging and discharging actions; that is, the energy storage regulation response will be cumulatively reflected as a charge state deviation within approximately 2 to 3 steps after the error mutation occurs). If an event from the error mutation marker exists within the window, the exceeding record is classified into the corresponding error mutation type (positive single-source / negative single-source / multi-source collaborative). After aligning the exceeding records of each charge segment in the charge level sequence with the error mutation markers according to the above rules, the historical exceeding frequency under each combination is statistically analyzed using the charge segment label as the row index and the error mutation type as the column index. The resulting two-dimensional weight structure after row normalization (the sum of the weights of each error type under the same charge segment is 1) is the historical weight for exceeding limits. The historical weight for exceeding limits is organized using the charge segment label as the row index and the error mutation type as the column index. The timing of multi-source coordinated mutations in the error mutation markers is highly correlated with the over-limit records in the low-charge segment of the charge classification sequence. This is because inaccurate multi-source synchronous predictions are usually accompanied by significant deviations on both the supply and demand sides, forcing energy storage to reach deep discharge limits. The over-limit historical weights obtain the highest weight values ​​at the intersection of the low-charge segment and the multi-source coordinated mutation. When the winter heating load forecast is significantly underestimated, the over-limit records in the low-charge segment of the charge classification sequence are densely distributed in the negative mutation column of the error mutation markers. The weight of this column in the over-limit historical weights is significantly higher than that of the positive mutation column, forming a directional weight asymmetry. The difference in the frequency of over-limits for the same charge segment under different error scenarios is explicitly quantified in the over-limit historical weights, and the distribution patterns of the over-limit historical weights under positive and negative mutation scenarios are drastically different.

[0057] High-risk charged segments are located using historical weights for exceeding limits, generating boundary adjustment sets. When the weight value of each charged segment combined with an error mutation type in the historical weights for exceeding limits exceeds twice the mean of that row, it is identified as a high-risk combination, and the corresponding charged segment enters the boundary adjustment set. The boundary tightening amount is calculated as Δ_b = (w_h - w_h_threshold) / w_h_threshold × Δ_max, where w_h is the historical weight value for exceeding limits corresponding to the high-risk combination (dimensionless), w_h_threshold is twice the mean of that row (dimensionless), and Δ_max is 15% of the width of the charged state segment. Δ_b and Δ_max have the same dimensions (both are percentages of the charged state). The tightening direction is consistent with the historical exceeding-limit direction: the lower limit of low-charged segments tightens upwards, and the upper limit of high-charged segments tightens downwards. High-risk segments corresponding to multi-source collaborative mutation scenarios have the highest priority in the boundary adjustment set. Before the cold wave, multi-energy forecasts were synchronously inaccurate, and the historical weight of low-charge segments of energy storage devices exceeding limits reached the highest in the multi-source collaborative column. The boundary adjustment set tightens the lower limit of charge for this segment upwards to Δ_max. The boundary adjustment set stores four attributes for each high-risk segment: label, error scenario type, tightening direction, and tightening magnitude. After the tightening boundary values ​​for each high-risk section are determined, the feasible region area is verified. The feasible region area is defined as the product of the feasible range width of the energy storage state of charge (the difference between the upper and lower limits of the state of charge, in units of state of charge percentage) and the feasible range width of the corresponding charge and discharge power (the difference between the upper and lower limits of the charge and discharge power, in units of MW) (the dimension is MW·state of charge percentage). When the ratio of the sum of the feasible region areas of all energy storage devices after tightening to the original sum of the feasible region areas calculated by the same caliber before tightening is less than 50%, the protection mechanism is triggered. Δ_max is automatically reduced to the minimum value required to ensure that the feasible region area is not less than 50% of the original. The protection trigger state and the actual tightening range are stored in the global attribute field of the boundary adjustment set.

[0058] For boundary adjustment sets and charge limit exceedance items, constraint boundary remapping is performed to determine compliant actions. The tightening boundary of each high-risk segment in the boundary adjustment set is matched with the records of charge limits exceeding that segment. The constraint boundary of the matched record is replaced with the tightened value, and the exceedance deviation is recalculated under the tightened boundary. The charging and discharging power command is adjusted in the direction that brings the exceedance deviation to zero. Actions that meet the tightened constraint boundary after adjustment are considered compliant actions. When the current charge state does not fall into any high-risk segment, the constraint boundary remains unchanged, and the record is corrected with minimum magnitude based on the original deviation. When multiple energy storage devices simultaneously fall into different high-risk segments, boundary remapping is performed sequentially according to the priority order of each segment in the boundary adjustment set. The compliant actions of the preceding devices are determined before the actions of the subsequent devices are processed, avoiding constraint conflicts caused by simultaneous tightening of multiple devices. Before an extreme cold wave arrives, temperatures plummet and load forecasts are significantly underestimated. Charge exceeding limits occur frequently in low-charge sections. The boundary adjustment set tightens the lower limit of low charge upwards. The discharge power of corresponding energy storage devices in compliant actions is limited to a more conservative level than the original actions. The difference between compliant and original actions is greatest during cold wave scenarios. After a compliant action is determined, a secondary verification of the original constraint space is performed. The compliant action is substituted into the untightened original constraint space to confirm that it still does not violate the original constraints. Compliant actions that pass the secondary verification are marked with a double compliance flag. Actions whose values ​​are very close to the original constraint boundary (less than 1% of the constraint range from the boundary) are marked with a boundary proximity warning. The time step of the boundary proximity warning is multiplied by an additional 1.2 times weight during experience replay sampling to strengthen the policy network's focus on optimizing the current compliant but insufficient buffer margin state.

[0059] The system generates a sequence of instructions based on safety corrections performed according to the correction vector and compliance actions. The bias values ​​of each power dimension in the correction vector are superimposed on the corresponding components of the original actions in the scheduling strategy network. The adjusted charging and discharging power of each energy storage device in the compliance actions replaces the corresponding components in the original actions. When the two dimensions overlap (typically, the energy storage charging and discharging power dimension is simultaneously subject to power and charge constraints), the correction vector is first applied to that dimension, and then the biased value is truncated a second time using the tightening boundary of the compliance actions, ensuring the final value satisfies both types of constraints. After these two steps, final constraint verification is performed. The verified scheduling action vector, along with the timestamp and the over-limit correction flag, is stored in the instruction sequence. In a scenario where the park experiences both tie line overload and energy storage bottoming out simultaneously in the winter evening, the correction vector reduces the tie line power to within the constraints, and the compliance actions limit the energy storage discharge to above the tightened lower charge limit. The combined effect of these two corrections ensures that the instruction sequence can still output executable actions that balance safety and absorption even in a double over-limit scenario. The instruction sequence is arranged sequentially according to execution time. Each instruction contains a scheduling action vector for the corresponding time step and an over-limit correction flag. The correction vector and the correction trace of the compliant action are completely preserved in the instruction sequence through the flag. The scheduling action vector that fails the final constraint verification triggers the emergency conservative action rollback mechanism, which sets the energy storage charging and discharging power to zero and adjusts the tie line power to 50% of the rated value as a safety default action to override the instruction at that time step. When the rollback is triggered for three consecutive time steps, an automatic alarm is triggered and the action sampling of the policy network is suspended. Instead, rule-based scheduling (a backup mechanism for scheduling energy storage and tie lines according to preset conservative rules, typically where energy storage performs constant power small-scale charging and discharging according to the current state of charge, and tie line power is maintained at about 50% of the rated value) continues to operate until the policy network resumes outputting compliant actions.

[0060] Step S15: Based on the instruction sequence, perform over-limit priority experience replay sampling on the scheduling policy network to form a gradient update sequence. Analyze the attribution characteristics of the gradient update sequence to form an over-limit attribution mapping. Based on the over-limit attribution mapping, perform selective soft updates on the scheduling policy network and output scheduling control instructions.

[0061] Specifically, a gradient update sequence is formed by sampling the over-limit priority experience of the scheduling policy network based on the instruction sequence. Each instruction in the instruction sequence, together with the state feature matrix row vector of the corresponding time step, the sampling action, and the immediate reward, forms an experience tuple, which is continuously added to the experience buffer with a capacity of N_buffer. In the instruction sequence, the over-limit correction flag serves as the base weight for sampling priority. The regular time step without correction records takes a base weight of 1, the power over-limit correction time step takes 2 times the base weight, the charge over-limit correction time step takes 2.5 times the base weight, the double over-limit time step takes 4 times the base weight, the multi-dimensional over-limit time step is multiplied by an additional 1.5 on top of the double over-limit weight, and the boundary approach warning time step is multiplied by 1.2 times on top of the above weights to strengthen the optimization focus on the insufficient buffer margin state. The empirical buffer is sampled according to the above weight probabilities to obtain the normalized sampling probability P_i (the sampling probability is normalized by the power of α_PER of the priority weight of each time step, where α_PER is the priority power exponent and is 0.6. The larger α_PER is, the higher the probability of high priority time steps being sampled. The sum of all P_i is 1). The empirical tuples obtained from each sampling form the gradient update sequence of this round. Importance sampling weights are calculated as w_i = [1 / (N_buffer×P_i)]^β_IS, where N_buffer is the buffer capacity (number of entries, dimensionless count), and β_IS is the annealing coefficient (linearly increasing from 0.4 to 1.0). w_i is multiplied into the gradient calculation of each tuple in the gradient update sequence to prevent gradient variance inflation due to oversampling in high-priority time steps. Historical time steps with large fluctuations in supply and demand during winter are frequently extracted in the gradient update sequence, allowing the scheduling strategy network to achieve denser gradient optimization in selecting compliant actions under such high-pressure scenarios. When the proportion of time steps exceeding the limit is less than 5% in the last 50 rounds of sampling, the sampling configuration automatically regresses to uniform sampling. α_PER decreases linearly by 0.05 in each subsequent training round until it drops to 0 (when α_PER is 0, the sampling weights of all time steps are equal, which is equivalent to uniform sampling), and β_IS is simultaneously accelerated to 1.0 to stabilize gradient estimation in the later stages of training.

[0062] An over-limit attribution mapping is formed based on the attribution characteristics of gradient update sequences. The update tuples of the over-limit time steps in the gradient update sequence are grouped into power over-limit and charge over-limit categories. The attribution granularity is divided according to the input layer feature modules, hidden layer, and output layer action modules of the scheduling strategy network. The input layer feature modules are divided into five feature modules based on the energy category labels of the state feature matrix: photovoltaic, wind power, load, energy storage status, and conflict sensing. The output layer action modules are divided into three action dimensions: energy storage charging and discharging power, demand response activation, and tie-line power adjustment. The mean L2 norm of the gradient vector of the parameters under each module is calculated as the gradient magnitude of that module (with dimensions consistent with the original parameter gradient). An attribution matrix is ​​constructed with over-limit type as the column and module as the row, serving as the core structure of the over-limit attribution mapping. During the summer photovoltaic over-generation period, the attribution value of the meteorological irradiance feature input module in the power over-limit column is high, while that of the energy storage status module is low. The over-limit attribution mapping clearly identifies the irradiance category as the main source of tie-line overload during this period. When the attribution values ​​are highly concentrated in a few modules, the cause of the limit violation is clear and can be located. During the peak period of frequent limit violations in winter, the sample size of the gradient update sequence is the largest, and the concentration of attribution in the limit violation attribution mapping is most significant during this period. After quantile normalization, the attribution values ​​of each module are labeled as high attribution modules above the 75th percentile, medium attribution modules between the 25th and 75th percentiles, and low attribution modules below the 25th percentile. The Jaccard similarity between the high attribution module sets of the power limit violation column and the charge limit violation column is used as the similarity index of the two types of limit violation attribution structures. When the similarity is not less than 0.3, the attribution level of each module is uniformly determined by the largest attribution level after taking the union. When it is less than 0.3, the two columns are labeled independently. For the same module, the larger one is taken according to the rule that "any column with high attribution is considered high attribution". The final attribution level label and the attribution values ​​of each module are recorded together in the limit violation attribution mapping to ensure that the attribution differences between the two types of limit violations are not diluted by each other.

[0063] Based on the out-of-limit attribution mapping, the scheduling strategy network is selectively soft-updated to output scheduling control commands. The attribution level label of each module in the out-of-limit attribution mapping directly determines the intensity of this round of soft update—high attribution modules need to be updated in a focused manner to weaken error tendencies, while low attribution modules need to remain stable to retain learned compliant action patterns; the soft update coefficient τ of high attribution modules is 0.8 to 0.9, medium attribution modules are linearly interpolated according to the attribution magnitude between 0.3 and 0.8, and low attribution modules are 0.1 to 0.3. Each parameter module completes differentiated updates according to θ_new=τ×θ_grad+(1-τ)×θ_old, where θ_new is the updated parameter value, θ_grad is the target parameter value after gradient descent of the current batch, θ_old is the parameter value before the update, and τ is the dimensionless soft update coefficient (∈[0,1]). After the soft update is completed, the scheduling strategy network performs forward inference with the row vector of the state feature matrix of the current time step as input, and the sampled actions are converted into scheduling control commands after safety correction and issued to each execution device in the park. During the period of sudden increase in heating load in the evening during winter, the attribution value of the energy storage status input module in the over-limit attribution mapping display remained high. This round of soft update applied a large τ value correction to the module, and after multiple rounds of training, the over-limit rate of the charge of the scheduling control command continued to decrease. After each soft update, attribution drift detection is performed. The current round is compared with the set of high attribution modules from the previous 5 rounds. When the Jaccard similarity is higher than 0.8, the attribution structure is considered stable. In the stable state, the upper limit of τ for high attribution layers is automatically reduced by 0.05, and the upper limit of medium and low attribution layers is reduced proportionally to prevent the converged parameters from oscillating due to continuous large step updates. When the Jaccard similarity is lower than 0.5, the attribution structure is considered to have drifted, and the upper limit of τ is restored to the initial value of 0.9 to ensure that newly emerging high attribution modules are fully updated. When the attribution structure is stable for 10 consecutive rounds and the violation rate is lower than 1%, the soft update is considered to have converged. In the initial training phase, the convergence of the strategy network is determined by a joint criterion of the absorption rate sliding standard deviation and the violation rate. In the online operation phase, the convergence of the soft update is determined by a joint criterion of the stability of the attribution structure and the violation rate. The scheduling strategy network must first satisfy the former before it can be put into online operation and the latter can be enabled.

[0064] To implement the source-grid-load-storage scheduling and control method based on deep reinforcement learning corresponding to the above method embodiments, and to achieve the corresponding functions and technical effects, see [link to documentation]. Figure 2 , Figure 2 This paper presents a structural block diagram of a source-grid-load-storage scheduling and control system 200 based on deep reinforcement learning, as provided in an embodiment of this application, including: The state construction unit 201 is used to collect source-grid-load-storage operation data of the park and model it according to multiple time scales to form a prediction set, identify error mutation features based on the prediction set to form error mutation markers, and perform feature fusion on the prediction set and the error mutation markers to form a state space. The weight priority unit 202 is used to perform non-uniform dimensionality reduction on the state space and the error mutation marker to form a state feature matrix, and to locate the high fluctuation range of the state feature matrix to form a priority weight matrix. The strategy training unit 203 is used to quantify the curtailment margin using the state feature matrix and the priority weight matrix to form a composite reward function, and to establish a scheduling strategy network by performing deep strategy gradient training using the state feature matrix and the composite reward function. Action correction unit 204 is used to detect scheduling action over-limits for the scheduling policy network to determine power over-limit items and charge over-limit items, analyze constraint deviations of the power over-limit items to form a correction vector, perform dynamic constraint boundary matching between the charge over-limit items and the error mutation marker to determine compliant actions, and perform safety correction to generate an instruction sequence according to the correction vector and the compliant actions. The instruction output unit 205 is used to perform out-of-limit priority experience replay sampling on the scheduling policy network based on the instruction sequence to form a gradient update sequence, analyze the attribution characteristics of the gradient update sequence to form an out-of-limit attribution mapping, and perform selective soft update on the scheduling policy network based on the out-of-limit attribution mapping to output scheduling control instructions.

[0065] The aforementioned source-grid-load-storage scheduling and control system 200 based on deep reinforcement learning can implement the source-grid-load-storage scheduling and control method based on deep reinforcement learning in the above method embodiments. The options in the above method embodiments are also applicable to this embodiment, and will not be detailed here. The remaining content of this application's embodiments can be referred to the content of the above method embodiments, and will not be repeated in this embodiment.

[0066] The purpose of the above embodiments is to reproduce and derive the technical solution of the present invention by way of example, and to fully describe the technical solution, purpose and effect of the present invention. The purpose is to enable the public to have a more thorough and comprehensive understanding of the disclosure of the present invention, and not to limit the scope of protection of the present invention.

Claims

1. A source-network-load-storage scheduling and control method based on deep reinforcement learning, characterized in that, include: Collect source-grid-load-storage operation data of the park and model it according to multiple time scales to form a prediction set. Based on the prediction set, identify error mutation features to form error mutation markers. Perform feature fusion on the prediction set and the error mutation markers to form a state space. Non-uniform dimensionality reduction is performed on the state space and the error mutation marker to form a state feature matrix, and the high fluctuation range of the state feature matrix is ​​located to form a priority weight matrix; The state feature matrix and the priority weight matrix are used to quantify the curtailment margin to form a composite reward function. A scheduling policy network is established by performing deep policy gradient training using the state feature matrix and the composite reward function. For the scheduling policy network, scheduling action over-limit detection is performed to determine power over-limit items and charge over-limit items. The constraint deviation of the power over-limit items is analyzed to form a correction vector. Dynamic constraint boundary matching is performed on the charge over-limit items and the error mutation marker to determine compliant actions. Safety correction is performed according to the correction vector and the compliant actions to generate an instruction sequence. Based on the instruction sequence, the scheduling policy network is sampled by replaying the experience of exceeding the limit priority to form a gradient update sequence. The attribution characteristics of the gradient update sequence are analyzed to form an exceeding limit attribution mapping. Based on the exceeding limit attribution mapping, the scheduling policy network is selectively soft-updated to output scheduling control instructions.

2. The method according to claim 1, characterized in that, The step of fusing features between the prediction set and the error mutation markers to form a state space includes: Energy category analysis is performed on the prediction set to generate multi-source prediction components; Perform cross-source power balance conflict detection on the multi-source prediction components to generate a conflict component set; Based on the conflict component set and the error mutation marker, an alignment feature sequence is generated by calibrating the conflict intensity. The state space is constructed by performing time-series and category dual-dimensional fusion based on the aligned feature sequence and the multi-source prediction components.

3. The method according to claim 1, characterized in that, The step of performing non-uniform dimensionality reduction on the state space and the error mutation marker to form a state feature matrix includes: The state space is grouped and arranged according to feature dimensions to generate a dimension grouping set; For the dimensional grouping set and the error mutation marker, an error lead response rate assessment is performed to generate a lead response rate ranking; The high-response-dimension segments are located by the advanced response-ranking to generate a compressed weight distribution; The state space is subjected to non-uniform dimensionality reduction based on the compressed weight distribution to form a state feature matrix.

4. The method according to claim 1, characterized in that, The step of quantifying the curtailment margin using the state feature matrix and the priority weight matrix to form a composite reward function includes: The state feature matrix is ​​sliced ​​into multiple time periods to generate time period operation components; Based on the time period operation components and the priority weight matrix, a margin sensitivity calculation is performed to generate a margin distribution map; Based on the margin distribution map, identify the critical period approaching full absorption and generate the full absorption critical set; A composite reward function is formed by synthesizing reward weights oriented towards consumption based on the full absorption critical set and the time period operation components.

5. The method according to claim 1, characterized in that, The high-fluctuation range of the state feature matrix is ​​used to form a priority weight matrix, including: Perform time-step fluctuation second-order difference calculation on the state feature matrix to generate a fluctuation acceleration sequence; Identify the multi-source synchronous acceleration transition segments of the wave acceleration sequence to generate a synchronous wave candidate set; The candidate set of synchronous fluctuations and the fluctuation acceleration sequence are sorted by priority ordering of the multi-source acceleration joint intensity score generation interval. A priority weight matrix is ​​formed by weighting based on the interval priority sorting and the state feature matrix.

6. The method according to claim 1, characterized in that, The step of forming a correction vector for the analytical constraint deviation of the power limit exceeding term includes: The power over-limit term is used to determine the over-limit direction and generate an over-limit component set; The frequent over-limit marker is generated by identifying repeated over-limit patterns using the over-limit component set; For the frequently exceeding the limit marker, a forward-looking bias prediction is performed to generate a corrected priority sequence; A preventative action bias correction is applied to the correction priority sequence and the over-limit component set to form a correction vector.

7. The method according to claim 1, characterized in that, The step of performing dynamic constraint boundary matching to determine compliance for the charge over-limit term and the error mutation marker includes: Perform charge level segmentation to generate a charge level sequence for the charge exceeding the limit term; For the charge classification sequence and the error mutation marker, perform out-of-limit history backtracking to generate out-of-limit history weights; The boundary adjustment set is generated by using the historical weights for exceeding limits to locate high-risk charged sections; Perform constraint boundary remapping on the boundary adjustment set and the charge over-limit term to determine compliance actions.

8. The method according to claim 3, characterized in that, The step of evaluating the error lead response rate and generating a lead response rate ranking based on the dimensional grouping set and the error mutation marker includes: Perform intra-group autocorrelation decay rate analysis on the dimensional grouping set to generate a decay rate distribution sequence; Based on the decay rate distribution sequence and the error mutation marker, a decay inflection point advance test is performed to generate an advance response set; Based on the aforementioned advanced response set, short memory response dimensions are identified to generate a response intensity ranking; Threshold filtering is applied to the response intensity ranking and the dimension grouping set to generate a leading response ranking.

9. The method according to claim 5, characterized in that, The process of identifying the multi-source synchronous acceleration transition segments of the wave acceleration sequence and generating a synchronous wave candidate set includes: A silent amplitude distribution map is generated by performing sliding window low-amplitude silent statistics on the wave acceleration sequence; The silence amplitude distribution map is used to determine the multi-source synchronous deep silence superposition interval and generate a silence candidate set. Perform a post-silencing mutation response evaluation on the silent candidate set to generate a mutation response score; Segment labeling is performed on the mutation response score and the fluctuation acceleration sequence to generate a synchronous fluctuation candidate set.

10. A source-grid-load-storage scheduling and control system based on deep reinforcement learning, characterized in that, include: The state construction unit is used to collect source-grid-load-storage operation data of the park and model it according to multiple time scales to form a prediction set. Based on the prediction set, error mutation features are identified to form error mutation markers. Feature fusion is performed on the prediction set and the error mutation markers to form a state space. The weight priority unit is used to perform non-uniform dimensionality reduction on the state space and the error mutation marker to form a state feature matrix, and to locate the high fluctuation range of the state feature matrix to form a priority weight matrix. The policy training unit is used to quantify the curtailment margin using the state feature matrix and the priority weight matrix to form a composite reward function, and to establish a scheduling policy network through deep policy gradient training using the state feature matrix and the composite reward function. The action correction unit is used to detect scheduling action over-limits in the scheduling policy network to determine power over-limit items and charge over-limit items, analyze constraint deviations of the power over-limit items to form a correction vector, perform dynamic constraint boundary matching between the charge over-limit items and the error mutation marker to determine compliant actions, and perform safety correction to generate an instruction sequence according to the correction vector and the compliant actions. The instruction output unit is used to perform over-limit priority experience replay sampling on the scheduling policy network based on the instruction sequence to form a gradient update sequence, analyze the attribution characteristics of the gradient update sequence to form an over-limit attribution mapping, and perform selective soft updates on the scheduling policy network based on the over-limit attribution mapping to output scheduling control instructions.