A micro-grid important equipment switching method and device based on big data analysis

Abstract

This invention discloses a method and apparatus for switching critical equipment in a microgrid based on big data analysis. The method involves collecting microgrid load demand, energy prices, and equipment efficiency data, and preprocessing this data to form an operating status parameter set. Load mean-variance analysis and load fluctuation trend extraction are performed on the operating status parameter set, and switching trigger condition parameters are determined through weighted fusion calculation. A switching probability distribution is generated through time-series nonlinear prediction, and low-probability switching segments are extracted to form switching suppression parameters, which are then fused with the switching trigger condition parameters to form a switching scheduling feature set. The method identifies the delayable switching margin for non-critical equipment and inter-cluster migration anomalies, constructing a switching decision matrix. Finally, convergent iterative correction is performed to determine the switching decision weights, and the switching decision matrix is ​​integrated to output equipment switching scheduling instructions, achieving adaptive scheduling of critical equipment switching in the microgrid driven by multi-dimensional operating data.

Description

Technical Field

[0001] This invention relates to the field of microgrid energy management technology, and in particular to a method and apparatus for switching important equipment in a microgrid based on big data analysis. Background Technology

[0002] Microgrids typically connect to multiple heterogeneous power sources, such as photovoltaics, energy storage, and diesel generators. The output of each power source is driven by weather conditions, load behavior, and market electricity prices. Switching operations of critical equipment are a key control measure to maintain the safe and economical operation of the system. The timing and selection of switching targets directly affect operating costs, equipment lifespan, and power supply continuity.

[0003] However, existing switching control methods mostly rely on fixed threshold rules or single signal triggering logic. Static thresholds cannot track dynamic changes in load patterns and electricity price structures, and are prone to false triggering or missed triggering under drastically different operating conditions such as peak electricity consumption and holiday shutdowns. They lack joint evaluation of the physical cost and economic benefits of switching operations and do not have the ability to identify and actively suppress low-value switching. They also lack the ability to automatically identify structural changes in the operating state when faced with sudden abnormal operating conditions, and are prone to scheduling lag in extreme weather or unexpected fault scenarios. Summary of the Invention

[0004] This invention discloses a method and apparatus for switching critical equipment in a microgrid based on big data analysis. It constructs a unified operational status description foundation by performing time-series synchronous preprocessing on load demand, energy price, and equipment efficiency data. Based on this, it analyzes load fluctuation patterns and price sensitivity characteristics to form quantitative criteria for switching trigger conditions. Furthermore, it combines time-series nonlinear prediction to identify and suppress low-value switching sections, generating switching scheduling features that balance economic benefits and operational costs. Finally, through state transition anomaly identification and convergence iterative correction, it outputs switching scheduling instructions for each device, achieving adaptive and refined management of critical equipment switching in the microgrid.

[0005] The first aspect of this invention proposes a method for switching critical equipment in a microgrid based on big data analysis, comprising the following steps: Collect load demand data, energy price data, and equipment efficiency data of the microgrid, and perform time-series synchronization preprocessing on the load demand data, energy price data, and equipment efficiency data to form a set of operating status parameters; A load mean-variance analysis is performed on the set of operating status parameters to generate a price-sensitive threshold range. Load fluctuation trend features are extracted from the set of operating status parameters to generate a switching prediction identifier. A weighted fusion calculation is performed based on the price-sensitive threshold range and the switching prediction identifier to determine the switching trigger condition parameters. A handover probability distribution is generated by performing time-series nonlinear prediction on the set of operating status parameters. Low-probability handover segments are extracted from the handover probability distribution to generate handover suppression parameters. The handover triggering condition parameters are then fused with the handover suppression parameters to form a handover scheduling feature set. The non-critical equipment in the switching scheduling feature set can be delayed to generate a scheduling buffer capacity. The inter-cluster migration event is identified in the switching scheduling feature set to generate an abnormal triggering identifier. The state transition features are extracted by fusing the scheduling buffer capacity and the abnormal triggering identifier to construct a switching decision matrix. The handover triggering condition parameters and the handover scheduling feature set are subjected to convergence iterative correction to determine the handover decision weights. The handover decision matrix is ​​then integrated with the handover decision weights to output the device handover scheduling command.

[0006] A second aspect of this invention proposes a switching device for critical equipment in a microgrid based on big data analysis, comprising: The data acquisition module is used to collect load demand data, energy price data and equipment efficiency data of the microgrid, and to perform time-series synchronous preprocessing on the load demand data, energy price data and equipment efficiency data to form a set of operating status parameters. The threshold decision module is used to perform load mean-variance analysis on the set of operating status parameters to generate a price-sensitive threshold range, extract load fluctuation trend features from the set of operating status parameters to generate a switching prediction identifier, and perform weighted fusion calculation based on the price-sensitive threshold range and the switching prediction identifier to determine the switching trigger condition parameters. The probability prediction module is used to perform time-series nonlinear prediction on the set of operating status parameters to generate a handover probability distribution, extract low-probability handover segments from the handover probability distribution to generate handover suppression parameters, and fuse the handover triggering condition parameters based on the handover suppression parameters to form a handover scheduling feature set. The matrix construction module is used to identify the non-critical equipment delayable switching margin of the switching scheduling feature set to generate a scheduling buffer capacity, perform inter-cluster migration event identification on the switching scheduling feature set to generate an abnormal triggering identifier, and extract state transition features based on the fusion of the scheduling buffer capacity and the abnormal triggering identifier to construct a switching decision matrix. The instruction output module is used to perform convergent iterative correction on the handover trigger condition parameters and the handover scheduling feature set to determine the handover decision weights, and to integrate the handover decision matrix through the handover decision weights to output the device handover scheduling instruction.

[0007] The beneficial effects of this invention are reflected in the following aspects: First, it performs time-series synchronous preprocessing of load demand, energy price, and equipment efficiency data to construct a unified foundation for describing operating conditions. Based on this, it jointly analyzes load fluctuation patterns and price sensitivity characteristics, establishing a weighted fusion calculation mechanism of economic signals and trend prediction signals. This allows the determination of switching trigger conditions to simultaneously consider both the external electricity price environment and the internal load evolution trend, overcoming the limitation of fixed threshold rules in tracking dynamic changes in operating conditions. Second, it generates a switching probability distribution by performing time-series nonlinear prediction on the set of operating state parameters, identifying low-probability segments where switching actions have historically been infrequent. Combined with a joint evaluation of the physical cost of switching operations in each segment, it forms a switching suppression parameter, which is then fused with the switching trigger condition parameter. This establishes an adversarial balance mechanism between trigger-driven and suppression-constrained actions, avoiding unnecessary wear and tear on equipment mechanical life caused by ineffective switching operations during low-yield periods. Finally, the delayable switching margin of non-critical equipment in the switching scheduling feature set is identified to determine the scheduling buffer capacity, and the load feature vectorization clustering of the switching scheduling feature set is performed to identify inter-cluster migration anomalies. After the two types of information are fused, a switching decision matrix reflecting the state transition law is constructed. Furthermore, the switching decision weight is determined through convergent iterative correction, so that the generation of scheduling instructions can adaptively optimize between economic drive and flexible scheduling, thereby improving the robustness of microgrid switching scheduling under complex operating conditions and sudden anomaly scenarios. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating a method for switching critical equipment in a microgrid based on big data analysis, according to the present invention.

[0009] Figure 2 This is a structural block diagram of a microgrid important equipment switching device based on big data analysis according to the present 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] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0012] 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.

[0013] The technical solutions of the embodiments of this application will be described below.

[0014] like Figure 1 As shown, this embodiment of the invention provides a method for switching critical equipment in a microgrid based on big data analysis, including the following steps S110-S150: Step S110: Collect load demand data, energy price data and equipment efficiency data of the microgrid, and perform time-series synchronization preprocessing on the load demand data, energy price data and equipment efficiency data to form a set of operating status parameters.

[0015] Specifically, the system collects load demand data, energy price data, and equipment efficiency data for the microgrid. Load demand data is collected in real-time by current transformers and power quality analyzers deployed at each access node of the microgrid, with a sampling interval of 1 minute, recording the active power, reactive power, and three-phase current readings at each node. Industrial load exhibits a high plateau curve during the daytime, while residential load forms a bi-peak structure in the morning and evening; the temporal distribution of load demand data directly reflects differences in electricity consumption behavior. Energy price data is retrieved from the real-time quotation interface of the power trading platform at a granularity of 1 minute, including three components: time-of-use pricing, peak-valley price difference, and renewable energy feed-in tariff. The price difference during peak and valley periods typically reaches 3 to 5 times. Peak-hour energy price data triggers equipment to prioritize switching to local energy storage power supply, while valley-hour energy price data drives energy storage charging. The time-of-use fluctuation pattern of energy price data is a core source of external economic signals. Equipment efficiency data is reported by each equipment controller at a 5-minute cycle, covering three indicators: photovoltaic inverter conversion efficiency, energy storage battery charging and discharging efficiency, and diesel generator fuel consumption rate. The efficiency of photovoltaic inverters dynamically varies between 88% and 97% depending on irradiance. High-temperature environments cause the charging and discharging efficiency of energy storage batteries to decrease by 2 to 4 percentage points. Real-time fluctuations in equipment efficiency data reflect the current energy conversion capacity of each device. Load demand data, energy price data, and equipment efficiency data are stored independently through their respective acquisition channels, and the original timestamps of the three types of data retain their original granularity without pruning.

[0016] Load demand data, energy price data, and equipment efficiency data undergo time-series synchronization preprocessing to form an operational status parameter set. Time-series synchronization uses a uniform alignment granularity of 1 minute. Energy price data, being 1-minute granularity, requires no interpolation. Equipment efficiency data is upsampled from a 5-minute granularity to a 1-minute granularity using linear interpolation to align with load demand data. After all three types of data are aligned to the same time axis, they enter the preprocessing flow. Missing values ​​are filled using nearest neighbor imputation. Periods with more than 10 consecutive minutes of missing data are also filled using nearest neighbor imputation, and a low-confidence label is added to the corresponding time in the operational status parameter set. Abnormal spikes in load demand data are detected using the 3σ criterion, and out-of-limit values ​​are replaced with the average of the three preceding and following times. Outliers in energy price data are repaired using the average of the same period on adjacent days. Records in equipment efficiency data with efficiency values ​​exceeding the physical upper limit are directly removed. After removal, null values ​​are filled using linear interpolation of the preceding and following averages to maintain time-series integrity. After quality processing, the three types of data are each normalized to zero mean and unit variance. Before normalization, copies of the original physical values ​​of the three types of data are retained for subsequent formula calculations. Normalization eliminates dimensional differences, allowing load power, electricity price, and efficiency percentage to participate in subsequent analysis within the same numerical space. The operating status parameter set merges the three types of data into a two-dimensional structure of time rows and feature columns. Each row corresponds to a time point, and the columns are, in order, the dimensions of each node of load demand data, the three components of energy price data, and the three indicators of equipment efficiency data. The operating status parameter set is organized using a sliding segmentation method with a 24-hour window length and a 1-hour step size. Each segment participates independently in the analysis and calculation of each step. The complete coverage period of the operating status parameter set is no less than 30 days to support the historical sample size required for trend analysis.

[0017] Step S120: Perform load mean-variance analysis on the set of operating status parameters to generate a price-sensitive threshold range, extract load fluctuation trend features from the set of operating status parameters to generate a switching prediction identifier, and perform weighted fusion calculation based on the price-sensitive threshold range and the switching prediction identifier to determine the switching trigger condition parameters.

[0018] In some embodiments, the step of performing load mean-variance analysis on the operating status parameter set to generate a price-sensitive threshold interval includes: processing the operating status parameter set by time period to generate a load mean sequence; performing variance statistics on the load mean sequence to generate a fluctuation intensity distribution; performing weighted coupling calculation on the fluctuation intensity distribution and energy price data in the operating status parameter set to generate a price fluctuation coupling parameter; and performing efficiency correction based on the price fluctuation coupling parameter and equipment efficiency data in the operating status parameter set to generate a price-sensitive threshold interval.

[0019] The operating status parameter set is processed by load averaging over time periods to generate a load averaging sequence. The time axis of the operating status parameter set is divided into 30-minute periods. All load power readings within a single period are aggregated using the arithmetic mean to obtain the representative load value for that period. The 24-hour analysis window is divided into 30-minute periods to generate 48 representative value elements. Load power at missing moments in the operating status parameter set is completed by linear interpolation of adjacent valid values ​​before participating in the average aggregation. If there are fewer than 10 valid sampling points in a single period, that period is marked as a low-quality period. The representative values ​​of low-quality periods are subject to confidence discounts in subsequent analyses to reduce their interference weight in judging the overall trend. The trend of the load averaging sequence depicts the basic electricity intensity profile of the microgrid at different times of the day. The load averaging sequence of the industrial plant microgrid is generally high from 8:00 to 18:00 during the day shift. During the night shift, with fewer personnel and non-critical equipment shutdown, the load drops to a low point. The peak-to-valley difference reflects the value difference of equipment switching at different times. The energy savings per kilowatt-hour from switching to an efficient power supply path during the high-platform period are much higher than those during the nighttime low-valley period. The difference between adjacent elements in the load mean sequence reflects the trend of load change at the time granularity. A period with a consistently positive difference indicates that electricity demand is in a continuous expansion phase, while a period with a consistently negative difference indicates that electricity activity is contracting. During holidays, the shutdown of industrial equipment leads to a generally low overall mean of the load mean sequence. The peak-valley distribution pattern of the load mean sequence directly affects the accuracy of subsequent variance statistics in identifying high-fluctuation periods.

[0020] A fluctuation intensity distribution is generated by variance statistics from the load mean sequence. The load mean sequence is progressively advanced using a sliding window of length 6 elements. The local variance of the six representative values ​​within each window reflects the dispersion of the load during that time period. The sliding step size is set to 1 time period to ensure high overlap between adjacent positions. The local variances of all positions in the load mean sequence are arranged chronologically to form the fluctuation intensity distribution. Positions with high fluctuation intensity distribution values ​​correspond to periods of rapid load changes. In a residential community microgrid, from 6:00 AM to 7:30 AM, households gradually wake up, and water heaters, microwave ovens, and rice cookers are turned on one after another. A large number of electrical appliances are put into use in a short period of time, causing the load mean sequence to rise rapidly during this period, and the high points of the fluctuation intensity distribution are concentrated here. After 10:00 PM, residential electricity consumption tends to stabilize, and televisions and a small amount of lighting remain at low levels, causing the fluctuation intensity distribution to fall back to the low value range. The fluctuation intensity distribution is globally normalized. After normalization, positions exceeding 0.7 are marked as high fluctuation periods, and positions below 0.3 are marked as low fluctuation periods. When high fluctuation periods are concentrated in fixed periods, it indicates that the load changes are periodic. The switching trigger conditions for high randomness periods need to be set with a wider tolerance margin to cope with unpredictable load changes. The intensity magnitude of the fluctuation intensity distribution provides a weighting basis for the weighted coupling calculation of energy price data.

[0021] Price fluctuation coupling parameters are generated by weighted coupling of the fluctuation intensity distribution with energy price data in the operating status parameter set. The time period markers of the fluctuation intensity distribution and the time-of-use electricity prices in the operating status parameter set are aligned with the same time period index. The comprehensive score of the price fluctuation coupling parameters incorporates peak-valley price difference correction, calculated as C = α × V_wave × P_norm + β × ΔP_norm, where C is the price fluctuation coupling parameter score, V_wave is the normalized value of the fluctuation intensity distribution, P_norm is the normalized value of the spot electricity price for the current time period, and ΔP_norm is the normalized value of the peak-valley price difference. Both P_norm and ΔP_norm are taken from the corresponding components of the normalized copies of the original physical values ​​in the operating status parameter set. α and β are the fluctuation weight and price difference weight, respectively, with values ​​of 0.7 and 0.3, and C ranges from 0 to 1. The introduction of the peak-valley price difference correction term suppresses the score of only periods with severe load fluctuations but extremely small peak-valley price differences, avoiding misjudging periods with weak price signals as high switching value intervals. During the peak startup period of the industrial park's microgrid, dozens of devices in the workshop are powered on simultaneously, resulting in significant load fluctuations. This, coupled with peak electricity prices being at their highest throughout the day, leads to a significantly higher price fluctuation coupling parameter score compared to the late-night period when devices operate at lower speeds and electricity prices are at their lowest. Switching to energy storage power at this time can simultaneously avoid demand charges and fully utilize pre-charged energy storage during off-peak hours. In the fluctuation intensity distribution, the price fluctuation coupling parameter score corresponding to periodically high-fluctuation periods remains stable across multiple historical analysis windows, while the score for random high-fluctuation periods fluctuates considerably across windows.

[0022] Efficiency correction is performed based on equipment efficiency data from the price fluctuation coupling parameter and operating status parameter set to generate a price-sensitive threshold range. The price fluctuation coupling parameter reflects the joint sensitivity of load and price, but it does not consider the reduction in switching economic benefits due to the actual conversion efficiency of the current equipment. High summer temperatures lead to increased internal resistance and decreased charging / discharging efficiency of energy storage batteries, significantly reducing the actual net switching benefit under the same price difference. Simply relying on the price fluctuation coupling parameter ranking will overestimate the switching benefits during such periods, necessitating efficiency correction to balance this. The weighted average of the photovoltaic inverter conversion efficiency and energy storage charging / discharging efficiency for each time period is extracted from the equipment efficiency data in the operating status parameter set as a comprehensive efficiency coefficient, with weights determined based on the rated capacity ratio of each device. The coupling strength after efficiency correction is obtained by multiplying the price fluctuation coupling parameter score for each time period by the corresponding comprehensive efficiency coefficient. The corrected sequence is then reordered, with the 90th percentile defining the upper limit of high sensitivity, the 50th percentile defining the lower limit of medium sensitivity, and the 20th percentile defining the low sensitivity benchmark. When the overall efficiency coefficient is below 0.85, the upper limit of the price sensitivity threshold range is automatically increased by 5%. This means that when equipment efficiency declines, stronger price and load signals are needed to cross the high sensitivity threshold and initiate a switchover, preventing the equipment from being frequently triggered in an inefficient state. The three threshold lines of the price sensitivity threshold range are refreshed step-by-step with the sliding window of the operating status parameter set. When the historical distribution of price fluctuation coupling parameters undergoes structural changes, the price sensitivity threshold range is automatically adjusted to match the latest distribution.

[0023] In some embodiments, the step of extracting load fluctuation trend features from the set of operating status parameters to generate a switching prediction identifier includes: performing trend direction analysis on the set of operating status parameters to generate a load trend vector; identifying trend acceleration segments from the load trend vector to generate load fluctuation trend features; performing trend continuation evaluation based on the load fluctuation trend features to generate a trend continuation parameter; and combining the trend continuation parameter with the load fluctuation trend features to generate a switching prediction identifier.

[0024] A load trend vector is generated by performing trend direction analysis on the set of operating status parameters. The load power time series in the operating status parameter set is analyzed by differential direction between adjacent time points. A positive differential value is assigned +1 to indicate an upward trend, a negative differential value is assigned -1 to indicate a downward trend, and a differential absolute value less than 2% of the historical average is assigned 0 to indicate stability. The differential direction sequence is statistically analyzed using a sliding window of length 5, which shows the proportion of positive and negative directions within the window. A positive direction proportion exceeding 0.6 indicates an upward trend, a negative direction proportion exceeding 0.6 indicates a downward trend, and all other cases are considered oscillations. The operating status parameter set is arranged according to the trend determination results of the sliding window throughout the entire time period to form the load trend vector. Each element of the load trend vector carries a direction label of upward, downward, or oscillating, as well as the estimated slope value of the corresponding window. During the morning shift start-up phase of an industrial plant, operators gradually activate each production line according to their clock-in order: air compressors are put into operation first, followed by welding units, and finally the painting line. The load power exhibits a continuous, multi-stage upward trend. The load trend vector is consistently upward for 14 consecutive moments during this period, achieving a trend continuity index of 14. The slope coefficient increases progressively at each level, indicating a sustained accumulation of trend momentum rather than random fluctuations. The longer the operating status parameter set covers the time period, the more comprehensive the statistical support of the load trend vector for various trend patterns. Short-cycle data leads to a lower trend continuity index, easily misinterpreting periodic oscillations as trend decay.

[0025] Load fluctuation trend features are generated by identifying trend acceleration segments from the load trend vector. The slope sequence of the load trend vector is used to detect the rate of change of the slope by the difference between adjacent slopes. When the rate of change of the slope exceeds 0.5 times the mean of the sequence, it is identified as an acceleration point. Three or more consecutive acceleration points constitute a trend acceleration segment. A single acceleration point is considered an instantaneous disturbance and is not included in the acceleration segment list. The starting position, duration, and direction of the trend acceleration segment are extracted as acceleration segment descriptors. In the commercial complex under the jurisdiction of the microgrid, during the lunch period, the restaurant's high-power kitchen equipment is fully operational, and the influx of customers causes the central air conditioning to increase its load and the elevators to operate frequently. Multiple high-power devices are loaded almost simultaneously within about 15 minutes. The load trend vector shows 8 consecutive acceleration points during this period, all in positive direction. The duration of the acceleration segment descriptor is recorded as 8, and the maximum slope occurs around the 10th minute of the lunch peak. The load fluctuation trend characteristics consist of three types of information: a list of acceleration segments, the maximum slope value of each segment, and the direction markers of each segment. The top three segments with the highest maximum slope values ​​are defined as the dominant acceleration segments. The time position and direction markers of the dominant acceleration segments serve as the core inputs for trend continuation assessment. The fluctuation trend characteristics are organized as a structure array, with the array length equal to the total number of identified acceleration segments. An excessive number of acceleration segments within a single analysis window indicates multiple activations of load momentum during that period, typically corresponding to complex operating scenarios where multiple production lines alternately start and stop or multiple peak electricity consumption periods overlap.

[0026] For example, the step of generating trend continuation parameters based on the load fluctuation trend characteristics includes: generating a trend change rate sequence by performing differential operations based on the load fluctuation trend characteristics; identifying sudden increase segments in the change rate from the trend change rate sequence to generate a sudden increase interval set; performing cross-time period consistency verification on the sudden increase interval set to generate a cross-time period effective acceleration segment; and generating trend continuation parameters based on the cross-time period effective acceleration segment.

[0027] A trend change rate sequence is generated by differential operations based on the load fluctuation trend characteristics. The slope values ​​of each acceleration segment in the load fluctuation trend characteristics are arranged in chronological order to form a slope time series. The slope time series is then differentially analyzed with adjacent elements at a step size of 1 to obtain the slope increment sequence, i.e., the trend change rate sequence. The sequence length is 1 less than that of the slope time series. A positive value in the trend change rate sequence indicates an increasing slope, i.e., an increasing trend momentum, while a negative value indicates a decreasing slope, i.e., a decreasing trend momentum. The larger the absolute value, the more drastic the momentum change. When the load fluctuation trend characteristics cover multiple acceleration segments, the slope time series of each segment generates a local trend change rate sequence. The local sequences are then spliced ​​together according to the starting time of each segment to form a complete sequence. During the evening peak rise phase of residential microgrids, residents' electricity consumption behavior is highly synchronous. Household appliances are turned on sequentially in a fixed order: dinner, shower, and television. The slope increases steadily throughout the rise phase, the trend change rate sequence is stable with a positive bias and a low standard deviation, and the momentum change is uniform. In contrast, industrial microgrids are affected by production scheduling. An emergency order for a particular batch causes multiple production lines to temporarily accelerate, resulting in a sharp jump in the slope within a short period. The trend change rate sequence exhibits isolated spikes, and the standard deviation is significantly higher. The difference in the trend change rate sequence morphology between the two scenarios directly determines the accuracy of identifying subsequent surge segments. At the splicing and connection points of various local sequences, because the preceding and following segments belong to different trend pulse cycles, the difference values ​​at the connection points are prone to false jumps. The difference values ​​at the splicing and connection points need to be marked as boundary transition values ​​and excluded in the subsequent identification of surge segments.

[0028] A surge interval set is generated by identifying surge segments in the rate of change sequence. Surge points are defined as moments in the rate of change sequence where the value exceeds the sequence mean plus two standard deviations. This threshold ensures that only acceleration events significantly deviating from the normal momentum change range are captured, filtering out low-amplitude random fluctuations. Three or more consecutive surge points constitute a surge segment; single or fewer than three consecutive surge points are considered isolated noise and excluded from the surge interval set. Two adjacent surge segments with an interval of less than two moments are merged into one segment. The start and end boundaries of the merged segment are taken as the outer boundaries of the two segments, ensuring that surge processes that physically belong to the same batch of startup events but are separated by a short interval are fully reconstructed. In an industrial microgrid undergoing equipment rotation scheduling, the first batch of production lines starts up first after shift change, followed by the second batch approximately two minutes later. These two batches of startup events are closely sequential; merging them forms a complete batch startup acceleration segment. The duration and peak intensity of this complete segment are higher than any single segment, more accurately reflecting the actual impact of this batch startup on load momentum. The surge interval set records three pieces of information for each surge segment: the start index, the end index, and the peak rate of change. If no surge point is identified in the trend rate of change sequence, the surge interval set is assigned an empty set, indicating that the current trend momentum change is gradual, and the load acceleration behavior has not formed an effective accumulation within the current analysis window, resulting in a correspondingly low confidence level for the switchover prediction. If the number of segments in the surge interval set exceeds three, it indicates that the load momentum has been activated abruptly multiple times within that window. This typically corresponds to complex operating scenarios with dense production shifts or multiple overlapping peak electricity consumption periods. Subsequent cross-time period consistency verification will further filter segments with historical regularity to eliminate interference from occasional activation events.

[0029] Cross-period consistency verification is performed on the burst interval set to generate effective acceleration segments across time periods. Each burst segment in the burst interval set is marked with a time period index. Cross-period consistency verification evaluates the co-occurrence rate of the burst interval set in the current window with the burst interval sets of the same historical time periods. The co-occurrence rate is determined by dividing the number of times the starting index of the current burst segment appears at the same position in multiple historical windows by the total number of historical windows. A high co-occurrence rate means that the acceleration pattern of this time period has been repeatedly verified in historical data, and has stable switching prediction support value. Burst segments with a co-occurrence rate exceeding 0.7 are marked as high overlap segments and directly included in the effective acceleration segments across time periods; burst segments with a co-occurrence rate between 0.3 and 0.7 are included in the effective acceleration segments across time periods after linear weighting based on the co-occurrence rate; those below 0.3 are marked as occasional segments and are not included in the effective acceleration segments across time periods. In a residential community's microgrid, a large number of residents return home from get off work each evening, turning on air conditioners and water heaters almost simultaneously. This results in a highly concentrated period of electricity consumption within approximately 20 minutes. This surge in electricity usage has consistently occurred within a 30-day historical timeframe and is identified as a high-overlapping segment, included in the cross-timeframe effective acceleration segment, providing high-confidence support for subsequent trend continuation assessments. Random events, such as sudden surges caused by a large number of devices simultaneously energizing after a temporary power outage in a residential building, have extremely low historical co-occurrence rates and are marked as occasional segments, excluded from the cross-timeframe effective acceleration segment to prevent unnecessary premature switching operations due to strong, occasional signals. When the co-occurrence rate of all surge segments within a surge interval is below 0.3, the cross-timeframe effective acceleration segment is assigned an empty set, indicating insufficient historical regularity in the current load acceleration pattern. The confidence level of the switching prediction indicator is correspondingly lowered, and the decision-making weight shifts to be dominated by the price-sensitive threshold range.

[0030] Trend continuation parameters are generated based on effective acceleration segments across time periods. The historical co-occurrence rate of each segment within the effective acceleration segments reflects the repeatability and reliability of load acceleration behavior during that period. A high co-occurrence rate means that the acceleration pattern during that period has been repeatedly verified in historical data. The two segments with the largest peak change rates are defined as the dominant effective acceleration segments, and the duration from the start to the end of the dominant effective acceleration segment is defined as the trend duration. The comprehensive score calculation formula for the trend continuation parameter is T_score = 0.4 × R_occur + 0.3 × K_norm + 0.3 × L_norm, where T_score is the comprehensive score for the trend continuation parameter, R_occur is the average historical co-occurrence rate of the dominant effective acceleration segments, K_norm is the normalized value of the peak change rate, and L_norm is the normalized value of the trend duration. All three are dimensionless, and T_score ranges from 0 to 1. At the start of each day's morning shift at a certain energy storage microgrid, operators, upon entering the site, confirm the status of each device according to safety procedures and sequentially energize them. This operational process is highly standardized under factory regulations. The dominant effective acceleration section exhibits a high historical co-occurrence rate, strong momentum peak, and stable duration, with all three scores approaching full marks. The comprehensive score for the trend continuity parameter is close to 0.9, indicating a high trend. Therefore, the prediction confidence for switching energy storage discharge to supplement the instantaneous peak at the start of the morning shift is sufficient during this period. When the effective acceleration section spans multiple time periods is an empty set, the comprehensive score for the trend continuity parameter is assigned a value of 0. A comprehensive score exceeding 0.7 indicates a high trend, 0.4 to 0.7 indicates a medium trend, and below 0.4 indicates a low trend. These three levels correspond one-to-one with the confidence level of the switching prediction indicator.

[0031] A switching prediction indicator is generated by combining trend continuation parameters and load fluctuation trend characteristics. The comprehensive score of the trend continuation parameters reflects the confidence strength of the trend continuation, while the directional marker of the dominant acceleration segment of the load fluctuation trend characteristics indicates the expected continuation direction. The combination of these two factors determines the directionality and confidence level of the switching prediction indicator. A positive switching prediction indicator is generated when the comprehensive score of the trend continuation parameters for an upward load fluctuation trend exceeds 0.7. This positive indicator indicates that the load is expected to continue to rise, and the switching path tends to connect high-capacity power supply units in advance, completing the switching deployment before the actual load reaches its peak, thus avoiding temporary power shortages due to equipment response lag during demand surges. A negative switching prediction indicator is generated when the comprehensive score of the trend continuation parameters for a downward load fluctuation trend exceeds 0.7. This negative indicator drives the switching path to a low-power standby mode, with high-capacity power supply equipment exiting operation early, avoiding long-term inefficient idleness due to severe overcapacity. A neutral switching prediction indicator is generated when the comprehensive score of the trend continuation parameters is between 0.4 and 0.7, with the switching decision weight shifting to the price-sensitive threshold range. The confidence field of the switching prediction flag directly records the comprehensive score value of the trend continuation parameter. The expected effective time field is estimated based on the duration of the acceleration segment dominated by the load fluctuation trend characteristics. Switching prediction flags with a confidence level below 0.5 are marked with an uncertainty label to prevent false triggering. After the switching prediction flag is generated, it also needs to be compared with the output result of the previous sliding window for directional consistency. In the case of directional reversal, a directional reversal label is automatically superimposed and the confidence level is corrected downward.

[0032] The switching trigger condition parameters are determined by a weighted fusion calculation based on the price-sensitive threshold range and the switching prediction indicator. The three-level threshold boundaries of the price-sensitive threshold range and the direction and confidence level of the switching prediction indicator constitute the two inputs for the weighted fusion calculation. The formula for calculating the weighted fusion score is S = w_p × P_level + w_t × T_conf, where S is the fusion score, P_level is the three-level sensitivity band encoding value of the price-sensitive threshold range (1.0 for high sensitivity, 0.6 for medium sensitivity, and 0.2 for low sensitivity), T_conf is the confidence level value of the switching prediction indicator, and w_p and w_t are the price weight and trend weight, respectively, with values ​​of 0.6 and 0.4. When the positive switching prediction indicator falls into the high sensitivity band, the fusion score usually exceeds 0.8, and the trigger type of the switching trigger condition parameter is determined as immediate switching; when the positive indicator falls into the medium sensitivity band or the confidence level is low, the trigger type is determined as delayed switching; when the negative switching prediction indicator or the running value is below the low sensitivity benchmark, the trigger type is determined as suppressed switching. During off-peak electricity pricing at night, the price sensitivity threshold range falls below the low-sensitivity benchmark. The confidence level of the handover prediction flag is insufficient to raise the fusion score above the trigger threshold. Therefore, the handover trigger condition parameters output a suppression type during this period, effectively preventing invalid handover actions during low-value nighttime hours. The handover trigger condition parameters consist of three parts: trigger type, trigger strength score, and trigger time estimate. The trigger time estimate is calculated based on the expected effective time of the handover prediction flag. When the price sensitivity threshold range boundary approaches, the trigger strength score of the handover trigger condition parameters automatically increases by 5% to enhance sensitivity near the boundary.

[0033] Step S130: Perform time-series nonlinear prediction on the set of operating status parameters to generate a handover probability distribution, extract low-probability handover segments from the handover probability distribution to generate handover suppression parameters, and fuse handover triggering condition parameters based on the handover suppression parameters to form a handover scheduling feature set.

[0034] In some embodiments, the step of generating a switching probability distribution by performing time-series nonlinear prediction on the set of operating state parameters includes: extracting historical time-series patterns from the set of operating state parameters to generate a time-series feature matrix; extracting renewable energy fluctuation components from the time-series feature matrix to generate a fluctuation correction factor; performing nonlinear prediction calculations based on the fluctuation correction factor to generate a prediction output sequence; and performing probability density estimation based on the prediction output sequence to generate a switching probability distribution.

[0035] Historical time-series patterns were extracted from the operating status parameter set to generate a time-series feature matrix. The operating status parameter set was divided into multiple historical samples with a 24-hour window length and a 1-hour step size. Within each sample, four statistical features were extracted for load power, electricity price component, and equipment efficiency time series: mean, standard deviation, first-order difference mean, and difference standard deviation. The mean reflects the overall intensity of the level within the time period, the standard deviation characterizes the degree of dispersion, the difference mean describes the direction of trend change, and the difference standard deviation captures the degree of fluctuation. These four features together constitute a multi-dimensional time-series description of the sample. During the continuous rainy weather of the plum rain season, photovoltaic output remained low throughout the day and fluctuated frequently at the minute level. The difference standard deviation of the photovoltaic output dimension of the operating status parameter set was significantly higher under this weather pattern than under sunny weather, and the corresponding column values ​​in the time-series feature matrix were consistently high, indicating that the prediction model needs to focus on the impact of output randomness on switching decisions under this pattern. The statistical features of each dimension of the operating status parameter set are concatenated into row vectors. The row vectors of all historical samples are stacked in chronological order to form a time-series feature matrix. The number of rows in the matrix equals the total number of historical samples, and the number of columns equals the sum of the number of features in each dimension. The matrix is ​​normalized to zero mean and unit variance to eliminate dimensional differences and ensure a balanced contribution weight of features of different magnitudes to the prediction model. If the number of rows in the time-series feature matrix is ​​less than 30, the number of historical samples is too small. The operating status parameter set needs to be extended and supplemented before re-extraction. The prediction confidence of the time-series feature matrix generated under insufficient sample conditions is downgraded to the warning level.

[0036] For example, the step of extracting renewable energy fluctuation components from the time-series feature matrix to generate a fluctuation correction factor includes: performing frequency domain transformation based on the time-series feature matrix to generate a spectrum distribution map; extracting dominant frequency components from the spectrum distribution map to generate a main frequency set; performing cross-cycle stability verification on the main frequency set to generate a stable main frequency set; and performing energy proportion analysis based on the stable main frequency set to generate a fluctuation correction factor.

[0037] A spectral distribution map is generated by performing a frequency domain transformation based on the time-series feature matrix. The column vectors corresponding to photovoltaic (PV) and wind power outputs in the time-series feature matrix are extracted as renewable energy time series. A discrete Fourier transform is applied to the renewable energy time series to map the power fluctuation sequence in the time domain to the frequency domain. The squared amplitude of each frequency component constitutes the power spectral density, which is arranged from low to high frequency to form the spectral distribution map. The horizontal axis of the spectral distribution map represents frequency, and the vertical axis represents the corresponding amplitude intensity. High amplitude positions in the spectral distribution map reveal the periodic dominant component of renewable energy output fluctuations, while low amplitude positions correspond to random noise background. The daily variation period is reflected as a low-frequency main peak approximately 1 / 24 hour, while short-term shading caused by rapid cloud movement is reflected as minute-level high-frequency components. The relative amplitude differences between the two types of components in the spectral distribution map are significant under different weather conditions. Under clear weather conditions, the low-frequency main peak in the spectral distribution map is prominent and sharp, while high-frequency noise is flat. Under continuous rainy weather conditions, the amplitude of the low-frequency main peak in the spectral distribution map shrinks significantly, and high-frequency dispersion increases, with the overall spectral shape changing from a single-peak dominant to a broadband dispersion pattern. The longer the historical coverage period of the time-series feature matrix, the higher the frequency resolution of the low-frequency components in the spectrum distribution map. When the resolution is insufficient, adjacent peaks in the low-frequency region of the spectrum distribution map are prone to overlap, affecting the accurate identification of the dominant period.

[0038] Dominant frequency components are extracted from the spectral distribution map to generate a main frequency set. Frequency components with amplitudes exceeding the global mean plus two standard deviations in the spectral distribution map are identified as dominant frequency components. This threshold ensures that only significant frequency points that substantially contribute to power output fluctuations are retained, filtering out weak periodic components with amplitudes close to the noise floor. When the amplitude difference between adjacent frequency points in the spectral distribution map is less than 10% of the local mean, the peak frequency of the corresponding continuous frequency segment is represented as the entire group to avoid spectral leakage that would cause a single physical cycle to be split into multiple adjacent peaks and counted repeatedly in the spectral distribution map. The frequency value, amplitude, and corresponding physical cycle duration of each dominant frequency component are extracted and aggregated into the main frequency set. The number of elements in the main frequency set reflects the number of identifiable independent periodic components in the renewable energy power output fluctuations within the current operating cycle. The main frequency set of photovoltaic microgrids in industrial parks typically includes two types of dominant components: the sunrise power cycle and the rapid morning cloud movement. The former corresponds to a stable peak in the low-frequency band, and the latter corresponds to a wide-envelope peak in the mid-to-high frequency band. The coexistence of these two types of components in the main frequency set reflects that photovoltaic power output is simultaneously driven by both solar radiation rhythms and local meteorological conditions. A small number of elements in the dominant frequency set suggests a simple power output fluctuation pattern and relatively low randomness, while a large number of elements in the dominant frequency set suggests the coexistence of multiple fluctuation mechanisms and strong power output randomness. When no frequency component in the spectrum distribution diagram exceeds the judgment threshold, the dominant frequency set is assigned an empty set, indicating that the current renewable energy time series is close to white noise, with no discernible regular fluctuation cycle.

[0039] A stable main frequency set is generated by performing cross-cycle stability verification on the main frequency set. Each frequency component in the main frequency set is identified as the dominant component in the current analysis window. However, the spectral characteristics of a single window may exhibit non-representative peaks due to occasional weather or short-term disturbances. Cross-cycle stability verification compares the spectral distribution of each component in the main frequency set with the spectral distribution maps of multiple historical analysis windows window by window. The ratio of the number of times each frequency component consistently appears as the dominant component in historical windows to the total number of historical windows is used as the cross-cycle co-occurrence rate of that component. Frequency components in the main frequency set with a cross-cycle co-occurrence rate exceeding 0.6 are marked as stable components and retained in the stable main frequency set. Components with a co-occurrence rate below 0.3 are marked as occasional components and removed from the stable main frequency set. Components in the middle range are retained after linear weighting according to their co-occurrence rate; the weights reflect the stability of the component within the historical windows. When continuous rainy weather lasts for more than 3 days, the amplitude of the photovoltaic diurnal variation cycle shrinks significantly in the spectral distribution diagram during this period. The cross-cycle co-occurrence rate of the corresponding component shows a phased decrease in the short term, and the weight of this component in the stable dominant frequency set decreases accordingly to dynamically reflect the actual contribution of this cycle under the current weather background. When the dominant frequency set is empty, the stable dominant frequency set is also empty. The number of elements in the stable dominant frequency set does not exceed the number of elements in the dominant frequency set. The difference between the two reflects the number of unstable fluctuation components introduced by occasional disturbances in the current window. A larger difference indicates that the current window is significantly affected by occasional weather interference.

[0040] A fluctuation correction factor is generated based on the energy proportion analysis of the stable main frequency set. The energy of each component in the stable main frequency set is determined by the proportion of the square of the amplitude of the corresponding frequency point in the total energy of the entire frequency band in the spectrum distribution diagram. The fluctuation correction factor is synthesized by weighting the energy proportion of each component with its cross-period co-occurrence rate. The calculation formula is F_mod=Σ(E_i / E_total×r_i), where F_mod is the fluctuation correction factor, E_i is the energy of the i-th component in the stable main frequency set, E_total is the total time-series energy of renewable energy, r_i is the cross-period co-occurrence rate of the i-th component, the sum of E_i / E_total does not exceed 1 and the value of r_i is in the range of 0 to 1, and the value of F_mod is in the range of 0 to 1. A higher fluctuation correction factor indicates that the current power output fluctuations are dominated by historical stable cycle components. The regularity of the stable cycle amplitude allows the prediction interval to be expanded accordingly to fully cover the known fluctuation range, resulting in a larger adjustment range for the prediction model parameters. Conversely, a lower fluctuation correction factor indicates insufficient contribution from stable cycle components and a higher proportion of random disturbances. The prediction interval expansion is correspondingly increased to cover a wider range of uncertainties introduced by random disturbances. In coastal wind power microgrids during the monsoon stable period, wind speed exhibits strong diurnal variation patterns and high cross-cycle recurrence rates. The stable dominant frequency energy is highly concentrated in a few low-frequency components, and the fluctuation correction factor is close to 0.9. During typhoons, the wind field structure is disordered, and the proportion of stable cycle components drops sharply, causing the fluctuation correction factor to fall below 0.3. The significant difference in fluctuation correction factors between these two scenarios directly drives the differentiated expansion of the subsequent prediction interval. When the stable dominant frequency set is empty, the fluctuation correction factor is assigned a value of 0, indicating that there is no identifiable stable fluctuation pattern in the current renewable energy output. The prediction interval is expanded by 30% according to the expansion rule below 0.3 to accommodate high randomness.

[0041] After correcting the time-series feature matrix based on a fluctuation correction factor, nonlinear prediction calculations are performed to generate a predicted output sequence. The prediction interval width of the renewable energy output-related columns in the time-series feature matrix is ​​proportionally expanded according to the fluctuation correction factor value. When the fluctuation correction factor is higher than 0.6, the prediction interval width is expanded by 20% to fully cover the amplitude range of known stable periods; when it is lower than 0.3, the prediction interval width is expanded by 30% to accommodate the high uncertainty caused by random disturbances. For intermediate ranges, the expansion ratio is determined by linear interpolation. Newly added corrected columns are individually normalized to zero mean and unit variance, while previously normalized columns remain unchanged. The nonlinear prediction model employs a multilayer perceptron structure, with ReLU selected as the activation function in the hidden layer to capture the nonlinear coupling relationship between load and electricity price. Each row of the time-series feature matrix is ​​sequentially fed into the model to obtain the predicted value of the switching trigger intensity at the corresponding time. The predicted values ​​for all times are arranged in chronological order to form the predicted output sequence; higher values ​​indicate a stronger expected driving force for triggering switching actions at that time. When the fluctuation correction factor is low, uncertainty labels are added to the relevant times in the predicted output sequence, indicating that the prediction reliability is limited by the randomness of output fluctuations. Outliers in the predicted output sequence are detected by the moving median at the previous 5 time points. Predicted values ​​that deviate from the median by 3 times the absolute deviation are replaced with the median to eliminate isolated prediction errors.

[0042] A switching probability distribution is generated by estimating the probability density based on the predicted output sequence. The predicted output sequence is fitted to the probability distribution curve using a kernel density estimation method. A Gaussian kernel is selected as the kernel function, and the bandwidth is adaptively determined according to the Silverman criterion. The bandwidth expands as the variance of the predicted output sequence increases to smooth the estimation results during periods of high volatility. The probability density curve after kernel density estimation is integrated and normalized over the predicted value domain. The normalized result forms the switching probability distribution, with the horizontal axis representing the predicted switching trigger intensity and the vertical axis representing the corresponding probability density. During sunny midday weather with sufficient photovoltaic output and stable electricity demand, the predicted output sequence values ​​are concentrated in the low to medium range, and the switching probability distribution exhibits a narrow peak shape, indicating low uncertainty in switching behavior. Under extreme high-temperature weather, the factory's air conditioning operates at full load, and photovoltaic output frequently drops due to efficiency decline caused by high temperatures. The predicted output sequence values ​​have large dispersion, and the switching probability distribution exhibits a low-flat, wide-tail shape, indicating the highest uncertainty in switching behavior during these periods. The switching probability distribution records the cumulative probability values ​​of each trigger intensity level by time period. A steep slope in the cumulative probability curve indicates a high concentration of the switching probability distribution, while a gentle slope indicates a high dispersion. The longer the historical coverage period of the operating status parameter set, the more sufficient the statistical basis of the switching probability distribution and the more robust the tail estimation. When the switching probability distribution exhibits a thick tail characteristic at the high trigger intensity end, it indicates that the probability of extremely high driving intensity events has not decreased sharply. Tail thickness information is additionally recorded in the attribute fields of the switching probability distribution as skewness and kurtosis statistics, providing a quantitative reference for tail risk in subsequent quantile classification.

[0043] In some embodiments, the step of extracting low-probability handover segments from the handover probability distribution to generate handover suppression parameters includes: dividing the handover probability distribution into probability grading intervals based on quantiles; locating the lowest probability level segment from the probability grading interval to generate a low-probability segment set; sorting the low-probability segment set by handover cost to generate handover cost grading parameters; and integrating the handover cost grading parameters with the low-probability segment set to generate handover suppression parameters.

[0044] The switching probability distribution is divided into quantiles to generate probability grading intervals. The cumulative probability curve of the switching probability distribution is divided into four levels using the 10th, 30th, and 70th quantiles as boundaries. The interval below the 10th quantile is defined as the extremely low probability zone, the interval between the 10th and 30th quantiles as the low probability zone, the interval between the 30th and 70th quantiles as the medium probability zone, and the interval above the 70th quantile as the high probability zone. This four-level division balances the fine-grained stratification of suppression intensity with the feasibility of practical operation. The switching probability distribution is mapped to the quantile division results by time period. Each time period is assigned to the corresponding level based on the position of its predicted switching trigger intensity value in the switching probability distribution. Time periods assigned to the extremely low probability zone mean that switching actions at that intensity level were rarely actually executed in historical data, corresponding to the strongest constraint interval for switching suppression. After the factory shuts down completely late at night, the park enters a minimum maintenance state. Security lighting and a small number of monitoring devices are kept at extremely low loads, and electricity prices remain at their lowest off-peak levels. The probability density of the switching probability distribution at this time of day is almost zero, classifying this period as an extremely low probability zone. The probability grading intervals here are clearly marked as low. The overall shape of the switching probability distribution changes with the seasons and work-rest patterns. The significant alternation between peak and off-peak periods in summer causes the boundaries of the probability grading intervals to adjust dynamically. The threshold values ​​for the probability grading intervals are updated synchronously with the switching probability distribution to ensure that the classification of each level always aligns with the latest operational patterns.

[0045] The lowest probability level segment is located within the probability grading interval to generate a low-probability segment set. All time periods belonging to the extremely low probability zone within the probability grading interval are extracted and arranged chronologically. Adjacent consecutive time periods belonging to the same level are merged into one segment. The merged segment is labeled with its start and end times, ensuring that short-interval-separated low-probability time periods are identified as complete and continuous suppression segments rather than multiple discrete fragments. When the extremely low probability zone time periods are concentrated within the probability grading interval, the merged segment sets form long-term low-probability segment sets, typically corresponding to late-night shutdowns or full-day shutdowns during holidays. During the Spring Festival holiday when the park is completely shut down, the low-probability segment set may cover entire periods of several consecutive days, indicating that there are almost no effective handover-driven scenarios during this period, and handover suppression forms a continuous and strong barrier during such periods. Time periods within the probability grading interval that only reach the low-probability zone but not the extremely low-probability zone are also included in the low-probability segment set, but with a secondary label to distinguish them from the extremely low-probability segments. These secondary-labeled segments receive a relatively lower suppression upper limit when assigning suppression weights later. Each element in the low-probability segment set records three pieces of information: the segment's time range, average probability density, and its level label. Segments with lower average probability density receive higher suppression priority in subsequent cost ranking. When the low-probability segment set is empty, it means that there are no obvious low-probability switching periods in the current running cycle, the switching suppression requirement is weak, and the scheduling logic is mainly driven by trigger signals.

[0046] The switching cost is sorted for low-probability segment sets to generate switching cost classification parameters. The physical cost of switching operations for each segment in the low-probability segment set is comprehensively quantified by three indicators: the number of mechanical actions of the circuit breaker, the contactor switching loss, and the relay protection verification delay. The switching cost value of each segment is obtained by weighted summation of the three indicators. The calculation formula is Cost = w_b × N_b + w_c × L_c + w_r × T_r, where Cost is the switching cost value, N_b is the normalized value of the number of mechanical actions of the circuit breaker, L_c is the normalized value of the contactor switching loss, T_r is the normalized value of the relay protection verification delay, and w_b, w_c, and w_r are the corresponding weights calibrated according to the sensitivity of equipment life loss. The mechanical life of the circuit breaker is usually measured by the number of operations, with the loss weight having the highest weight. The sum of the three weights is 1, and the value of Cost ranges from 0 to 1. During the low-cost, late-night hours when energy storage batteries are fully charged, there is no further room for charging. Forcibly triggering a switchover during these periods not only increases the number of mechanical actions of the circuit breaker but may also cause additional actions in the protection circuit due to overcharge protection. The switching cost during these periods is compounded by the additional losses from equipment protection actions, resulting in a significantly higher cost than during normal switching periods. This high-cost characteristic exhibits a stable regularity in historical operation records. In the low-probability segment set, segments are arranged in descending order of switching cost value. The top 30% of segments are defined as high-cost segments, the middle 40% as medium-cost segments, and the bottom 30% as low-cost segments. These three levels constitute the switching cost grading parameters. The grading boundaries of the switching cost grading parameters are dynamically determined based on the cost distribution within the low-probability segment set. During operating cycles with generally high cost values, the three grading boundaries shift upwards synchronously to ensure that the grading results always reflect the relative distribution characteristics of the switching cost within the current cycle. The switching cost grading parameters are output in the form of a segment identifier-cost level mapping table.

[0047] The handover cost grading parameters are integrated with the low-probability segment set to generate handover suppression parameters. The average probability density of each segment in the low-probability segment set and the cost level of the handover cost grading parameters jointly determine the suppression weight of that segment. The comprehensive evaluation of these two dimensions ensures that the suppression strength reflects both the historical scarcity of handover behavior and the physical cost of operation. In the handover cost grading parameters, high-cost-level segments correspond to the upper limit of the suppression weight, and low-cost-level segments correspond to the lower limit of the suppression weight. The suppression weight calculation rules are as follows: the suppression weight of a segment with an extremely low probability and a high cost level is assigned 1.0, the suppression weight of a segment with an extremely low probability and a medium cost level is assigned 0.85, and the suppression weight of a segment with a low probability and a high cost level is assigned 0.75. Other combinations are linearly converted according to the product coefficient of the probability level and the cost level. During the period of complete shutdown of the park at night, both the extremely low probability and high cost conditions are met. The suppression weight of the handover suppression parameters reaches its full value during this period. Any trigger signal with a suppression strength lower than the full value cannot break through the suppression barrier during this period, effectively protecting the equipment from frequent operation losses during periods with no substantial benefits. The handover suppression parameters arrange the suppression weights for each time period in chronological order to form a suppression weight sequence. The sequence length is the same as the number of time periods in the handover probability distribution. When the low-probability segment set is empty, all suppression weights in the handover suppression parameters are assigned a value of 0, indicating that no active handover suppression is required in the current operating cycle, and all time periods are dominated by the trigger signal for scheduling decisions. The handover suppression parameters also include a label indicating the physical suppression reason for each segment. The reason label distinguishes between probability-driven suppression and cost-driven suppression, which helps operations and maintenance personnel quickly locate the source logic of the suppression signal when reviewing scheduling records, providing a traceable decision-making basis for the continuous optimization of subsequent handover strategies.

[0048] A switching scheduling feature set is formed by fusing switching suppression parameters with switching trigger condition parameters. The suppression weight sequence of the switching suppression parameters and the trigger intensity score sequence of the switching trigger condition parameters are aligned by the same time period index. The two types of signals form an antagonistic relationship in the same time period, and the final scheduling tendency depends on the net difference between the two. During the continuous high temperature weather in summer, the surface temperature of photovoltaic panels rises and the conversion efficiency declines. At the same time, the central air conditioning in the factory area is at full load and the cooling systems of each workshop are turned on simultaneously, and the load climbs to the highest point of the day. The trigger intensity of the switching trigger condition parameters reaches its peak. During this period, the suppression weight of the switching suppression parameters is at a low level, and the net difference is significantly biased towards the execution direction. The switching scheduling feature set outputs a strong execution switching signal during this period, driving the backup diesel generator to switch into grid to supplement the capacity gap. After the production lines are shut down one after another late at night, all high-power equipment enters standby mode. The power consumption of the park is only for security and basic maintenance loads. The trigger intensity of the switching trigger condition parameters drops to the lowest point of the day, and the suppression weight of the switching suppression parameters rises to a high point simultaneously. The net difference is biased towards the suppression direction. The switching scheduling feature set outputs a strong suppression signal during this period to avoid unnecessary mechanical losses caused by performing equipment switching during the late night period when there is no substantial benefit. The switching scheduling feature set arranges the fused scheduling signals into a multi-dimensional feature vector by time period. Each element of the vector carries three attributes: trigger strength net score, suppression weight, and scheduling tendency. The scheduling tendency is distinguished by three categories of labels: execution, delay, and suppression. The net score of the execution label period crosses the preset threshold, while the net score of the delay label period is observed to be close to the threshold.

[0049] Step S140: Identify the non-critical equipment delay switching margin of the switching scheduling feature set to generate the scheduling buffer capacity; identify inter-cluster migration events of the switching scheduling feature set to generate anomaly triggering identifiers; and extract state transition features based on the fusion of the scheduling buffer capacity and the anomaly triggering identifiers to construct a switching decision matrix.

[0050] In some embodiments, the step of identifying the non-critical equipment delayable switching margin of the switching scheduling feature set to generate scheduling buffer capacity includes: classifying the switching scheduling feature set into equipment priorities to generate an equipment priority distribution; identifying a set of non-critical equipment from the equipment priority distribution to generate delayable equipment identifiers; estimating the switching margin based on the delayable equipment identifiers to generate a delayable switching margin group; and performing operational constraint verification on the delayable switching margin group to generate scheduling buffer capacity.

[0051] Equipment priority distribution is generated by prioritizing equipment in the switching scheduling feature set. The functional attributes, process dependence, and downtime tolerance of each equipment in the switching scheduling feature set constitute the input dimensions for priority evaluation. After weighting each of the three parameters, a weighted sum is obtained to obtain the priority score of each equipment. The calculation formula is: Score = w_f × F_func + w_q × P_proc + w_s × T_stop, where Score is the equipment priority score, F_func is the normalized value of the functional attribute, P_proc is the normalized value of the process dependence, T_stop is the normalized inverse value of the downtime tolerance, and w_f, w_q, and w_s are the corresponding weights, with the sum of the three being 1. The Score value ranges from 0 to 1. If the injection molding machine cooling system is interrupted during switching, it will cause work-in-process to be scrapped, and its priority score is close to the maximum value; a short-term shutdown of the auxiliary ventilation equipment does not affect any production process, and its priority score is low. Equipment priority distribution is represented by the complete set of priority scores for each device. Devices with scores above 0.8 are classified as critical devices, those with scores between 0.5 and 0.8 as secondary critical devices, and those with scores below 0.5 as non-critical devices. Production microgrids typically form peak clusters of core production equipment in the high-score range and peak clusters of auxiliary facilities in the low-score range. Equipment priority distribution is updated regularly based on changes in device access or periodic assessment cycles. Regular updates are independent of changes in the scheduling tendency markers of the switching scheduling feature set, maintaining an independent update mechanism. Temporary downgrading of secondary critical devices is a dynamic adjustment based on the current scheduling tendency markers on the basis of the equipment priority distribution, without changing the score records of the equipment priority distribution itself.

[0052] The non-critical equipment set is identified from the equipment priority distribution to generate deferred equipment identifiers. All equipment in the non-critical equipment layer of the equipment priority distribution is extracted into a non-critical equipment set. Equipment in the secondary critical equipment layer with no active tasks in the current time period is temporarily downgraded and included in the non-critical equipment set. The two sets of equipment are merged to form the complete non-critical equipment set for the current time period. In a commercial complex microgrid, after closing at night, elevators, escalators, and public lighting are all in standby mode. Although elevators are secondary critical equipment during normal business hours, they are temporarily downgraded to non-critical equipment when there is no passenger demand late at night. The switching operation can be postponed without affecting any business operations. The temporary downgrade of secondary critical equipment in the equipment priority distribution is determined based on the scheduling tendency flag of the current time period in the switching scheduling feature set. During the marked period, secondary critical equipment maintains its original level without downgrading, while the suppressed or delayed marked period allows for temporary downgrading to expand the adjustable space. Delayable device identifiers are a list of device numbers in the non-critical device set. Each device number in the list includes a degrading reason for the current time period, distinguishing between regular non-critical and temporary degrading sources. The length of the delayable device identifier list reflects the scale of devices that can participate in elastic scheduling for the current time period; the longer the list, the greater the degree of flexibility in adjusting the switching scheduling feature set for the current time period. Temporarily degraded devices also have a recovery priority label in the delayable device identifier list. Once the scheduling tendency of the switching scheduling feature set changes from suppression or delay to execution, temporary degraded devices with high recovery priority must be the first to leave the non-critical device set and recover to the second-critical level to participate in normal scheduling.

[0053] A delay switching margin group is generated based on the delayable device identifiers. Each device in the delayable device identifier list has its maximum delay time estimated based on process tolerance and historical delay records. Process tolerance is determined by combining the allowable power outage duration specified in the equipment manual with actual operational monitoring data. Historical delay records are used as the upper confidence limit, representing the longest time without any alarms triggered after performing delay operations under such environmental conditions. The smaller of the two values ​​is taken to ensure a safety margin. The process tolerance of auxiliary ventilation equipment is typically constrained by both the workshop temperature rise rate and worker safety procedures. In high-temperature summer environments, the maximum delay time for auxiliary ventilation equipment is reduced to 5 minutes, while in low-temperature winter environments, it can be delayed to 15 minutes. The maximum delay duration of each device in the list of delayable devices is multiplied by its rated switching power to obtain the delay switching power-time product for each device. The physical meaning of the power-time product is the maximum elastic switching energy that the device can bear during the longest delay period. The delay switching power-time products of all devices are summed to form a delay switching margin group. Each element of the delay switching margin group corresponds to the margin contribution of a single device. The device with the largest margin contribution is usually an auxiliary device with high rated power and loose process tolerance. This type of device is the main source of scheduling buffer capacity. If the list of delayable devices is an empty set of the time-space delay switching margin group, it means that all devices in the current time period are critical or strongly time-constrained devices with no elastic scheduling space.

[0054] The scheduling buffer capacity is generated by verifying the operational constraints of the delay switching margin group. Before being put into use, the margin duration of each device in the delay switching margin group needs to be verified against two types of constraints. The first type is the local overload constraint, which checks whether the real-time load rate of the feeder where the non-critical device is located exceeds 90% of the rated value during the delay switching period. If it does, the margin duration of the corresponding device is shortened proportionally to the feeder margin, ensuring that the feeder does not trigger protection actions due to continuous full-load operation during the delay switching period, thus affecting the normal power supply of other devices. The second type is the energy storage state of charge constraint. If the energy storage needs independent power supply during the delay switching period of non-critical devices, the predicted decrease in state of charge within the delay duration must not cause it to fall below the 20% protection lower limit. If there is a risk of falling below this level, the margin duration is shortened until the state of charge just reaches the protection lower limit, preventing over-discharge of the energy storage from forcibly interrupting subsequent charging cycles. In the delay handover margin group, the margin duration of devices that pass both types of constraint checks is retained, while the margin duration of devices that fail is adjusted to the upper limit allowed by the constraints. After passing the checks, the effective margin duration of each device is weighted and summed to obtain the scheduling buffer capacity. The weights are determined based on the proportion of each device's rated power to the total power of non-critical devices, using the formula D_buf=Σ(P_i / P_total×T_i), where D_buf is the scheduling buffer capacity, P_i is the rated power of the i-th device, P_total is the total rated power of non-critical devices, and T_i is the effective margin duration after checks. A larger scheduling buffer capacity means that more handover actions can be postponed in the current time period, providing flexibility for the path selection of the subsequent handover decision matrix. When all devices in the delay handover margin group are constrained to zero, the scheduling buffer capacity is assigned a value of 0, indicating that there is no flexible scheduling space in the current time period.

[0055] In some embodiments, the step of identifying inter-cluster migration events and generating anomaly trigger identifiers from the switching scheduling feature set includes: vectorizing load features from the switching scheduling feature set to generate a load feature vector group; initializing and iteratively allocating cluster centers from the load feature vector group to generate load pattern clusters; performing a weighted comprehensive score of cluster boundary distance and outlier degree on the load pattern clusters to generate an outlier degree score distribution; and identifying anomaly events and generating anomaly trigger identifiers based on the outlier degree score distribution.

[0056] Load feature vector groups are generated from the handover scheduling feature set by vectorizing load features. The three attributes extracted from each time period in the handover scheduling feature set—net trigger strength score, suppression weight, and scheduling tendency encoding value—are concatenated to form the feature vector for that time period. Scheduling tendency is numerically encoded using a value of 1.0 for execution, 0.5 for delay, and 0.0 for suppression. The concatenated vector has a dimension of 3, covering three orthogonal information dimensions: economic pressure, suppression signal, and tendency judgment. This dimensional design ensures that different operating modes form a distinguishable distribution in the feature space. The feature vectors for each time period are arranged chronologically to form the load feature vector group. The load feature vector group is then normalized to zero mean and unit variance to eliminate dimensional differences between dimensions. The normalization retains the original mean and standard deviation parameters for each dimension for subsequent inverse normalization. During the off-peak hours of the night shift, the microgrid in the industrial park exhibits low net trigger strength scores, high suppression weights, and a suppression-oriented scheduling tendency. The corresponding feature vectors in the load feature vector group for this period stably cluster in the low-value region of the feature space. During the morning shift startup phase, the net trigger strength score rises sharply, the suppression weights decrease, and the scheduling tendency shifts to execution. The feature vectors in the load feature vector group jump into the high-value region. These two typical scenarios form naturally occurring clusters with clear intervals in the feature space, providing a clear data foundation for subsequent clustering analysis. The more abundant the sample size in the load feature vector group, the clearer the distribution profile of each typical operating mode in the feature space. However, the sparse samples in the short-term load feature vector group during the historical coverage period of the switching scheduling feature set result in coarse cluster boundary definitions and limited anomaly identification accuracy. It is recommended that the coverage period be no less than 30 days to support stable pattern learning.

[0057] Load pattern clusters are generated by initializing and iteratively assigning cluster centers from the load feature vector group. The load feature vector group is clustered using the K-Means algorithm. The number of clusters K is automatically selected between 2 and 6 based on the elbow rule. The elbow rule determines the optimal number of clusters by calculating the inflection point of the sum of variances within clusters for different K values. At the inflection point, the variance reduction caused by increasing the K value significantly narrows, achieving an optimal balance between classification accuracy and computational complexity. Cluster center initialization uses the K-Means++ strategy. The first center is randomly selected from the load feature vector group, and subsequent centers are selected sequentially with the squared probability of their distance from the previously selected centers. This strategy ensures a uniform distribution of initial centers in the feature space, effectively avoiding unstable cluster partitioning results caused by random initialization falling into local optima. The feature vectors of each time period in the load feature vector group are iteratively assigned to the nearest cluster center. The iteration converges when the cluster assignments for all time periods no longer change. After convergence, each time period and its assigned cluster center together constitute the load pattern cluster. A typical microgrid operating day usually forms 3 to 4 load pattern clusters, corresponding to the late-night off-peak mode, daytime peak mode, transitional ramp-up mode, and sudden anomaly mode. The center vector of each load pattern cluster represents the typical characteristic combination of the corresponding operating mode. After the cluster center vector is inversely normalized and restored to the physical dimension space, it can be used by operation and maintenance personnel to intuitively understand the actual trigger intensity level and suppression degree of each operating mode. When a new sample is added to the load feature vector group, the load pattern cluster is periodically retrained to update the cluster center. When the operating mode structure changes due to seasonal changes, the number of load pattern clusters may be adjusted accordingly. The retraining trigger condition is that the number of new samples exceeds 20% of the total number of existing samples.

[0058] An outlier score distribution is generated by weighting cluster boundary distance and outlier degree for load pattern clusters. The Euclidean distance between the feature vector of each time period and its corresponding cluster center is defined as the intra-cluster distance. The minimum Euclidean distance between the feature vector of the same time period and non-cluster centers is defined as the inter-cluster distance. A larger ratio of intra-cluster distance to inter-cluster distance indicates that the feature vector of that time period is closer to the cluster boundary and has a lower certainty of belonging to the current cluster. Cluster affiliation is measured using a silhouette coefficient variant, calculated as O_score=(d_out-d_in) / max(d_in,d_out), where O_score is the cluster affiliation score, d_in is the intra-cluster distance, and d_out is the inter-cluster distance. The value of O_score ranges from -1 to 1. A higher value indicates that the operating state of that time period is more stably belonging to the current cluster, while a lower value indicates that the operating state of that time period is closer to the cluster boundary or even crosses into the domain of other pattern clusters. The outlier score distribution is formed by arranging the O_scores of all load mode clusters in chronological order. A sudden drop in the outlier score distribution corresponds to the transition point where the operating state migrates from one stable mode to another. The larger the drop, the more drastic the migration and the higher the degree of change in operating state. During the closing and clearing phase of a commercial complex's microgrid, lighting, air conditioning, and POS systems are sequentially powered off. The load mode cluster feature vector continuously drifts from the peak operating mode cluster to the late-night standby mode cluster within approximately 30 minutes. The outlier score distribution remains consistently low during this transition phase, reflecting that the operating state is in the inter-cluster migration channel. Extremely short drops in the outlier score distribution typically correspond to sudden migration events such as equipment failure or unexpected circuit breaker tripping, clearly distinguishing it from normal gradual mode switching in terms of score distribution pattern. This serves as an important signal for subsequent anomaly event identification.

[0059] Anomaly trigger markers are generated based on outlier score distribution. Points in the outlier score distribution with scores below the global mean minus two standard deviations are identified as candidate anomalies. This threshold ensures that only operational state migration events significantly deviating from normal fluctuation ranges are captured. Three or more consecutive candidate anomalies constitute an anomaly event segment. Two adjacent anomaly event segments with an interval of less than two time points are merged into one segment. The merged segment is marked with its outer boundary to represent the complete anomaly period. The merging operation uniformly handles consecutive anomaly scenarios where outlier scores briefly rise and then fall again. Planned inter-cluster migrations caused by normal mode switching in the outlier score distribution may also trigger candidate anomaly points. To distinguish between planned and unplanned migrations, inter-cluster migrations during the execution-marked periods in the switching scheduling feature set are marked as planned migrations and are not included in the anomaly trigger markers. Only migration events occurring outside the execution-marked periods are identified as anomaly triggers. This differentiation mechanism effectively reduces the interference of normal scheduling operations on anomaly identification results and improves the accuracy of anomaly trigger markers. An inverter cooling fan malfunction in a certain energy storage microgrid triggered automatic switching due to equipment temperature rise protection. During this period, the switching scheduling characteristic set showed a suppression tendency, and the outlier score distribution dropped sharply. These two conditions combined confirmed an unplanned anomaly. The anomaly trigger identifier recorded the occurrence time, the source and target cluster numbers, and the outlier score. When no anomaly event was identified in the outlier score distribution, the anomaly trigger identifier was set to an empty set, indicating that all inter-cluster migrations within the current operating cycle were planned operations, and the microgrid's operating state was stable.

[0060] A switching decision matrix is ​​constructed by fusing state transition features extracted from scheduling buffer capacity and abnormal triggering indicators. State transition features are extracted from the time-period correspondence between scheduling buffer capacity and abnormal triggering indicators. Periods with ample scheduling buffer capacity but missing abnormal triggering indicators are defined as stable and adjustable states; periods with abnormal triggering indicators and insufficient scheduling buffer capacity are defined as tense and mandatory states; periods with abnormal triggering indicators but still sufficient scheduling buffer capacity are defined as early warning and flexible states; and periods with missing abnormal triggering indicators and insufficient scheduling buffer capacity are defined as rigid constraint states. Under rigid constraint states, all equipment is at a critical operating level with no flexible scheduling space, and switching actions can only be executed when the trigger intensity of the switching triggering condition parameters exceeds the high-sensitivity upper bound. The four states are arranged chronologically to form a state sequence, and the transition relationship between state markers in adjacent time periods is defined as a state transition event. The switching decision matrix is ​​organized in a directed graph structure, with graph nodes corresponding to the four operating states. The edge weights are obtained by dividing the historical occurrence count of each type of transition event by the total number of transition events. Under extreme summer heat, the edge weights for the transition from stable and adjustable states to tense and mandatory states are significantly higher, reflecting the historical pattern of rapid deterioration of operating states under this weather condition. The high-frequency transfer paths in the switching decision matrix correspond to the most common switching pressure evolution trajectory of microgrids, while low-frequency but high-cost transfer paths correspond to low-probability extreme events. Transfer paths in the switching decision matrix starting from the early warning elastic state have bidirectional reachability; they may either return to a stable and adjustable state or deteriorate into a tense and forced state. The historical transfer probabilities in each direction at the bifurcation node and the remaining proportion of the current scheduling buffer capacity together constitute the quantitative basis for path selection.

[0061] Step S150: Perform convergence iterative correction on the handover trigger condition parameters and handover scheduling feature set to determine the handover decision weights, and output the device handover scheduling command by integrating the handover decision matrix through the handover decision weights.

[0062] Specifically, convergent iterative correction is performed on the handover triggering condition parameters and the handover scheduling feature set to determine the handover decision weights. The trigger intensity score sequence of the handover triggering condition parameters and the scheduling tendency label sequence of the handover scheduling feature set exhibit consistent or contradictory correspondences across different time periods. Consistency is shown in the scheduling tendency being execution when the trigger intensity is high and suppression when the trigger intensity is low. Contradiction is shown in the case of high trigger intensity of the handover triggering condition parameters but a suppression tendency in the handover scheduling feature set, or low trigger intensity of the handover triggering condition parameters but an execution tendency in the handover scheduling feature set. The existence of contradictory periods indicates that the trade-off between the two types of signals has not yet reached a stable equilibrium, requiring iterative correction to eliminate the contradiction. Iterative correction starts from the initial handover decision weights, with the initial weights set at 0.5 for trigger intensity and 0.5 for scheduling tendency. In each iteration, the trigger intensity score of the handover triggering condition parameters and the scheduling tendency encoding value of the handover scheduling feature set are weighted and summed based on the current handover decision weights to generate a comprehensive scheduling score. Both are normalized to the 0-1 interval, and the comprehensive scheduling score subsequently takes a value within the same interval. The overall scheduling score is compared with the handover execution results in the actual historical operation records to generate a comparison error. In each iteration, the weight adjustment direction is determined based on the correlation direction between the comparison error of each time period and the deviation of the two types of components: when the error contribution of the time period when the trigger strength score is higher than the scheduling tendency code value is positive, the trigger strength weight is reduced and the scheduling tendency weight is increased; when the error contribution of the time period when the scheduling tendency code value is higher than the trigger strength score is positive, the scheduling tendency weight is reduced and the trigger strength weight is increased. The adjustment step size is fixed at 0.05 in each round until the change amplitude of the handover decision weight is less than 0.01 in two consecutive iterations, at which point convergence is determined. After the handover decision weight converges, it is represented by two values: the trigger strength weight and the scheduling tendency weight, the sum of which is 1. The handover decision weight is periodically recalibrated as the operating mode changes.

[0063] The switching decision matrix outputs device switching scheduling commands by integrating switching decision weights. Each operating state node and state transition edge in the switching decision matrix constitutes a structured knowledge framework for scheduling decisions. Switching decision weights serve as guiding parameters for matrix traversal; when the weights favor trigger strength, matrix traversal prioritizes high-trigger-strength transition paths; when the weights favor scheduling tendency, matrix traversal prioritizes flexible scheduling paths. The current operating state is located by comparing the scheduling tendency marker of the current time period in the switching scheduling feature set with the state definition in the switching decision matrix. The historical occurrence frequency and trigger strength of each transition path are weighted and evaluated based on the switching decision weights. The path with the highest comprehensive score corresponds to the optimal switching action sequence. In the afternoon peak period of the photovoltaic-storage-diesel microgrid, the high-frequency transition path from a stable and adjustable state to a tense and forced state in the switching decision matrix scores the highest. The switching decision weights simultaneously indicate a high trigger tendency; both point to a switching action sequence that immediately connects the diesel generator to the grid and reduces the energy storage discharge depth. Equipment switching scheduling instructions are generated one by one for each device involved in the optimal transfer path. Each instruction includes the target device number, switching action type, execution time, and priority label. The execution time is determined jointly based on the historical average duration of path transfers in the switching decision matrix and the current scheduling buffer capacity. Equipment switching scheduling instructions corresponding to high-cost, low-frequency paths in the switching decision matrix are appended with a manual confirmation label to prevent irreversible operational errors in fully automated scheduling under extreme scenarios. Equipment switching scheduling instructions are output in a structured list format, arranged in ascending order of execution time, with the highest priority switching instruction at the top of the list.

[0064] To implement the above-described method embodiments, a microgrid critical equipment switching method based on big data analysis is proposed to achieve the corresponding functional and technical effects. See also... Figure 2 , Figure 2 This paper illustrates a structural block diagram of a microgrid critical equipment switching device 200 based on big data analysis, according to an embodiment of this application. The device includes: The data acquisition module 201 is used to collect load demand data, energy price data and equipment efficiency data of the microgrid, and to perform time-series synchronous preprocessing on the load demand data, energy price data and equipment efficiency data to form a set of operating status parameters. Threshold decision module 202 is used to perform load mean variance analysis on the set of operating status parameters to generate a price-sensitive threshold range, extract load fluctuation trend features from the set of operating status parameters to generate a switching prediction identifier, and perform weighted fusion calculation based on the price-sensitive threshold range and the switching prediction identifier to determine the switching trigger condition parameters. The probability prediction module 203 is used to perform time-series nonlinear prediction on the set of operating status parameters to generate a handover probability distribution, extract low-probability handover segments from the handover probability distribution to generate handover suppression parameters, and fuse the handover triggering condition parameters based on the handover suppression parameters to form a handover scheduling feature set. Matrix construction module 204 is used to identify the non-critical equipment delay switching margin of the switching scheduling feature set to generate a scheduling buffer capacity, perform inter-cluster migration event identification on the switching scheduling feature set to generate an abnormal triggering identifier, and extract state transition features based on the fusion of the scheduling buffer capacity and the abnormal triggering identifier to construct a switching decision matrix. The instruction output module 205 is used to perform convergence iterative correction on the handover trigger condition parameters and the handover scheduling feature set to determine the handover decision weights, and to integrate the handover decision matrix through the handover decision weights to output the device handover scheduling instruction.

[0065] The aforementioned microgrid critical equipment switching device 200 based on big data analysis can implement a microgrid critical equipment switching method based on big data analysis in the above-described 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 embodiment can be referred to the content of the above method embodiments, and will not be repeated in this embodiment.

[0066] The above embodiments are not an exhaustive list based on the present invention, and there may be many other embodiments not listed. Any substitutions and improvements made without departing from the concept of the present invention are within the protection scope of the present invention.

Claims

1. A method for switching critical equipment in a microgrid based on big data analysis, characterized in that, include: Collect load demand data, energy price data, and equipment efficiency data of the microgrid, and perform time-series synchronization preprocessing on the load demand data, energy price data, and equipment efficiency data to form a set of operating status parameters; A load mean-variance analysis is performed on the set of operating status parameters to generate a price-sensitive threshold range. Load fluctuation trend features are extracted from the set of operating status parameters to generate a switching prediction identifier. A weighted fusion calculation is performed based on the price-sensitive threshold range and the switching prediction identifier to determine the switching trigger condition parameters. A handover probability distribution is generated by performing time-series nonlinear prediction on the set of operating status parameters. Low-probability handover segments are extracted from the handover probability distribution to generate handover suppression parameters. The handover triggering condition parameters are then fused with the handover suppression parameters to form a handover scheduling feature set. The non-critical equipment in the switching scheduling feature set can be delayed to generate a scheduling buffer capacity. The inter-cluster migration event is identified in the switching scheduling feature set to generate an abnormal triggering identifier. The state transition features are extracted by fusing the scheduling buffer capacity and the abnormal triggering identifier to construct a switching decision matrix. The handover triggering condition parameters and the handover scheduling feature set are subjected to convergence iterative correction to determine the handover decision weights. The handover decision matrix is ​​then integrated with the handover decision weights to output the device handover scheduling command.

2. The method according to claim 1, characterized in that, The step of performing load mean-variance analysis on the set of operating status parameters to generate a price-sensitive threshold range includes: The set of operating status parameters is processed by load averaging over time periods to generate a load averaging sequence; A variance statistics method is used to generate a fluctuation intensity distribution from the load mean sequence; The price fluctuation coupling parameter is generated by weighted coupling calculation of the fluctuation intensity distribution and the energy price data in the operating state parameter set; Based on the price fluctuation coupling parameter and the equipment efficiency data in the operating status parameter set, efficiency correction is performed to generate a price-sensitive threshold range.

3. The method according to claim 1, characterized in that, The step of extracting load fluctuation trend features from the set of operating status parameters to generate a switching prediction identifier includes: Perform trend direction analysis on the set of operating status parameters to generate a load trend vector; The load fluctuation trend features are generated by identifying trend acceleration segments from the load trend vector. Based on the load fluctuation trend characteristics, a trend continuation assessment is performed to generate trend continuation parameters; A switching prediction identifier is generated by combining the trend continuation parameter and the load fluctuation trend characteristics.

4. The method according to claim 1, characterized in that, The step of generating a switching probability distribution by performing time-series nonlinear prediction on the set of operating state parameters includes: Historical time-series patterns are extracted from the set of operating status parameters to generate a time-series feature matrix; A volatility correction factor is generated by extracting renewable energy volatility components from the time-series feature matrix. After correcting the time series feature matrix based on the fluctuation correction factor, nonlinear prediction calculations are performed to generate a prediction output sequence. Based on the predicted output sequence, probability density estimation is performed to generate a switching probability distribution.

5. The method according to claim 1, characterized in that, The step of extracting low-probability handover segments from the handover probability distribution to generate handover suppression parameters includes: Based on the switching probability distribution, quantile division is performed to generate probability grading intervals; A low-probability segment set is generated by locating the lowest probability level segment from the probability grading interval; The low-probability segment set is sorted by handover cost to generate handover cost classification parameters; The handover cost classification parameters are integrated with the low-probability segment set to generate handover suppression parameters.

6. The method according to claim 1, characterized in that, The non-critical equipment that identifies the switching scheduling feature set can delay switching margin to generate scheduling buffer capacity, including: The switching scheduling feature set is used to generate a device priority distribution by classifying device priorities. Generate deferred device identifiers by identifying a set of non-critical devices from the device priority distribution; Based on the delayed device identifier, a delay switching margin group is generated by estimating the switching margin. The scheduling buffer capacity is generated by performing operational constraint verification on the delay switching margin group.

7. The method according to claim 1, characterized in that, The step of generating an anomaly trigger identifier by identifying inter-cluster migration events in the switching scheduling feature set includes: Load feature vector groups are generated by vectorizing the load features from the switching scheduling feature set; The load pattern clusters are generated by initializing and iteratively allocating cluster centers from the load feature vector group. An outlier score distribution is generated by weighting the cluster boundary distance and outlier degree for the load pattern clusters. Anomaly trigger identifiers are generated based on the outlier score distribution to identify abnormal events.

8. The method according to claim 3, characterized in that, The process of generating trend continuation parameters based on the load fluctuation trend characteristics includes: A trend change rate sequence is generated by performing differential operations based on the load fluctuation trend characteristics. Identify abrupt increase segments in the rate of change from the trend change rate sequence to generate a set of abrupt increase intervals; Perform cross-time period consistency verification on the set of sudden increase intervals to generate effective acceleration segments across time periods; Based on the trend continuation parameters of the effective acceleration segment generation across time periods.

9. The method according to claim 4, characterized in that, The step of extracting renewable energy volatility components from the time-series feature matrix to generate volatility correction factors includes: A frequency domain transformation is performed based on the aforementioned time-series feature matrix to generate a spectral distribution map; The dominant frequency components are extracted from the spectrum distribution map to generate a main frequency set; Perform cross-cycle stability verification on the main frequency set to generate a stable main frequency set; Based on the stable main frequency set, an energy proportion analysis is performed to generate a fluctuation correction factor.

10. A switching device for critical equipment in a microgrid based on big data analysis, characterized in that, include: The data acquisition module is used to collect load demand data, energy price data and equipment efficiency data of the microgrid, and to perform time-series synchronous preprocessing on the load demand data, energy price data and equipment efficiency data to form a set of operating status parameters. The threshold decision module is used to perform load mean-variance analysis on the set of operating status parameters to generate a price-sensitive threshold range, extract load fluctuation trend features from the set of operating status parameters to generate a switching prediction identifier, and perform weighted fusion calculation based on the price-sensitive threshold range and the switching prediction identifier to determine the switching trigger condition parameters. The probability prediction module is used to perform time-series nonlinear prediction on the set of operating status parameters to generate a handover probability distribution, extract low-probability handover segments from the handover probability distribution to generate handover suppression parameters, and fuse the handover triggering condition parameters based on the handover suppression parameters to form a handover scheduling feature set. The matrix construction module is used to identify the non-critical equipment delayable switching margin of the switching scheduling feature set to generate a scheduling buffer capacity, perform inter-cluster migration event identification on the switching scheduling feature set to generate an abnormal triggering identifier, and extract state transition features based on the fusion of the scheduling buffer capacity and the abnormal triggering identifier to construct a switching decision matrix. The instruction output module is used to perform convergent iterative correction on the handover trigger condition parameters and the handover scheduling feature set to determine the handover decision weights, and to integrate the handover decision matrix through the handover decision weights to output the device handover scheduling instruction.

Images