An adaptive multiphase parallel power management system

Abstract

The application discloses a kind of self-adapting multiphase parallel power management system, it is related to electric power system intelligent regulation and control technical field, the system includes: feature extraction module, periodicity, trend and randomness features are extracted from historical power consumption data and real-time sensor data, time series decomposition algorithm is used to determine load surge probability, and load surge probability value is obtained;Load prediction module, if load surge probability value is higher than preset threshold, then real-time load data is processed by multilayer perception machine structure, load surge peak value and duration in hour level or minute level time window are predicted, and load surge prediction result is obtained;The self-adapting multiphase parallel power management system uses multi-algorithm cooperation and real-time dynamic optimization, significantly improves power grid operation stability and power supply continuity, reduces overload and voltage fluctuation risk.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for power systems, and specifically to an adaptive multiphase parallel power supply management system. Background Technology

[0002] As the lifeline of modern society, the stability and continuity of the power system are crucial to ensuring social production and daily life. In complex and ever-changing operating environments, the power system must cope with various emergencies such as sudden increases and decreases in load and abnormal fluctuations to ensure uninterrupted power supply.

[0003] However, existing solutions often fall short in the face of these dynamic changes due to a lack of forward-looking prediction and flexible dynamic adjustment capabilities. This makes it difficult to effectively prepare for emergencies, leading to decreased system stability or partial power outages. The main limitation of current methods in dealing with load changes lies in their over-reliance on static scheduling strategies and single response mechanisms, ignoring the complexity of load changes and the need for adaptation to multiple scenarios. This makes it difficult for the system to achieve accurate resource allocation and rapid response when facing rapidly changing loads, especially in extreme scenarios such as the sudden startup of critical equipment or partial grid failures, where the system may become unstable due to delayed response. The core challenges lie in two key technical attributes: the ability to predict emergencies and the ability to dynamically adjust loads. Insufficient predictive capabilities prevent the system from anticipating potential load fluctuations. For example, during peak electricity consumption periods in industrial parks, failure to allocate reserve capacity in advance can lead to voltage fluctuations or power outages. The lack of dynamic load adjustment capabilities makes it difficult for the system to quickly adjust generation and distribution strategies when facing sudden load changes, potentially triggering a chain reaction. For example, in a city's power grid, when electric vehicle charging stations start up simultaneously, causing a sudden surge in load, the system may be unable to adjust power distribution in time, potentially leading to localized grid overload and affecting the normal power supply to other areas. These two technical attributes are interrelated: accurate prediction is a prerequisite for dynamic adjustment, while inadequate dynamic adjustment further amplifies the risks arising from the lack of prediction. Summary of the Invention

[0004] The purpose of this invention is to provide an adaptive multiphase parallel power management system that, in complex and variable load environments, improves the ability to predict sudden events and dynamically adjust loads, thereby achieving stable operation and continuous power supply of the power system.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an adaptive multiphase parallel power management system, the system comprising: a feature extraction module, which extracts periodic, trend, and random features from historical electricity consumption data and real-time sensor data, and uses a time series decomposition algorithm to determine the probability of load surge, thereby obtaining a load surge probability value; a load prediction module, which, if the load surge probability value is higher than a preset threshold, processes real-time load data through a multilayer perceptron structure to predict the peak value and duration of load surge within an hourly or minute-level time window, thereby obtaining a load surge prediction result; a standby capacity allocation module, which, based on the peak value and duration in the load surge prediction result, obtains available capacity data from a standby power database, and uses a linear programming algorithm to optimize standby capacity allocation, thereby obtaining a standby capacity allocation scheme; a power grid status monitoring module, which, based on the standby capacity allocation scheme, obtains power grid operating status data through a real-time monitoring system, detects whether the power grid voltage deviates from the normal range, and determines whether voltage fluctuation risk exists; and a voltage stabilization adjustment module, which, if a voltage fluctuation risk is detected, adjusts the generator output through a dynamic adjustment algorithm, and, combined with the predicted peak value of the load surge, stabilizes the voltage output, thereby obtaining stable voltage output data.

[0006] Preferably, the feature extraction module extracts periodic, trend, and random features from historical electricity consumption data and real-time sensor data, and uses a time series decomposition algorithm to determine the probability of a load surge. The process for obtaining the load surge probability value includes: acquiring a raw dataset from historical electricity consumption data and real-time sensor data; using preprocessing techniques to remove noise and missing values ​​to obtain a standardized dataset; based on the standardized dataset, using a classic time series decomposition algorithm to extract periodic, trend, and random features to obtain a decomposed feature set; and for the decomposed feature set, if the amplitude of the periodic feature exceeds a preset threshold, it is marked as a high-periodic fluctuation. A periodic fluctuation label set is generated. Based on the decomposed feature set and the periodic fluctuation label set, the slope change of the trend feature is calculated. If the absolute value of the slope change is greater than a preset threshold, it is marked as a trend change, resulting in a trend change label set. Random features are extracted from the decomposed feature set, and the random fluctuation intensity is calculated using a Gaussian distribution fitting method, resulting in a random fluctuation intensity value. Based on the periodic fluctuation label set, the trend change label set, and the random fluctuation intensity value, a logistic regression algorithm is used to calculate the probability of a load surge, resulting in a load surge probability value. If the probability value of the load surge is greater than a preset threshold, a surge risk assessment is triggered, resulting in a risk assessment result.

[0007] Preferably, the load prediction module, if the probability value of a load surge is higher than a preset threshold, processes real-time load data using a multilayer perceptron structure to predict the peak load surge and duration within an hourly or minute-level time window, obtaining the load surge prediction result by: acquiring raw load data from real-time sensors, segmenting the data using a sliding window technique to obtain a segmented load dataset; calculating load fluctuation characteristics within each time window based on the segmented load dataset to obtain a fluctuation feature set; if the peak change rate in the fluctuation feature set is higher than a preset threshold, training the fluctuation feature set using a multilayer perceptron model to obtain a load surge prediction model; processing real-time load data using the load surge prediction model to predict the load peak within an hourly time window to obtain an hourly peak prediction result; calculating the load change trend within a minute-level time window based on the hourly peak prediction result to obtain a minute-level trend feature; if the absolute value of the slope of the minute-level trend feature is greater than a preset threshold, classifying the minute-level trend feature using a random forest algorithm to determine the duration of the load surge and obtain a duration prediction result; and fusing the hourly peak prediction result and the duration prediction result to generate a load surge prediction report and obtain the final prediction output.

[0008] Preferably, the backup capacity allocation module, based on the peak value and duration in the load surge prediction results, obtains available capacity data from the backup power database and optimizes the backup capacity allocation using a linear programming algorithm to obtain a backup capacity allocation scheme, including: obtaining load peak value and duration data from the load surge prediction results; querying the backup power database through a database interface to obtain backup capacity information; constructing a linear programming model based on the backup capacity information, where the objective function is to maximize allocation efficiency, and the constraints include load peak value and duration, to obtain an optimized allocation model; processing the backup capacity information through the linear programming model to calculate the allocation ratio of each power unit, to obtain a preliminary capacity allocation scheme; if the allocation ratio of any power unit in the preliminary capacity allocation scheme exceeds a preset threshold, adjusting the allocation ratio using a greedy algorithm to prioritize meeting the load peak value requirement, to obtain an adjusted allocation scheme; obtaining real-time power status data based on the adjusted allocation scheme, determining whether each power unit meets the capacity requirements of the allocation scheme, to obtain a power availability judgment result; updating the capacity status in the backup power database based on the power availability judgment result, to generate a final capacity allocation scheme; and scheduling the power management system to execute allocation operations based on the final capacity allocation scheme, to determine the allocation execution status.

[0009] Preferably, the power grid status monitoring module, for the standby capacity allocation scheme, acquires power grid operating status data through a real-time monitoring system, detects whether the power grid voltage deviates from the normal range, and determines whether voltage fluctuation risk exists, including: collecting power grid operating status data through the real-time monitoring system and generating a voltage data stream; if the voltage data exceeds a preset threshold range, a voltage anomaly detection algorithm is used to determine the voltage fluctuation risk and obtain a risk assessment result; based on the risk assessment result, an anomaly alarm mechanism is used to generate an alarm signal and determine the alarm priority; based on the alarm priority, power grid load level data is acquired to determine the power grid stability status; if the power grid stability status is lower than a preset stability threshold, a support vector machine algorithm is used to analyze the voltage data stream and the power grid load level to obtain stability optimization parameters; based on the stability optimization parameters, the data acquisition frequency is adjusted to generate an optimized monitoring data stream; and based on the optimized monitoring data stream, the real-time monitoring system configuration is updated to determine a dynamic adjustment scheme for the power grid operating status.

[0010] Preferably, the voltage stabilization module, if it detects a voltage fluctuation risk, adjusts the generator output using a dynamic adjustment algorithm, and stabilizes the voltage output by combining the predicted load surge peak. The stable voltage output data obtained includes: if a voltage fluctuation risk is detected, analyzing the load surge peak using a support vector machine algorithm, adjusting the generator output, and obtaining stable voltage output data; based on the stable voltage output data, acquiring real-time monitoring data streams of the power grid operating status, analyzing the voltage data change trend, and determining the stability of the power grid operating status; if the stability of the power grid operating status is lower than a preset threshold, processing the real-time monitoring data streams using a random forest algorithm, extracting voltage fluctuation features, and obtaining voltage fluctuation feature parameters; adjusting the data acquisition frequency based on the voltage fluctuation feature parameters, generating an optimized monitoring data stream, and determining a dynamic adjustment scheme for the power grid operating status; analyzing the relationship between the power grid load level and voltage fluctuation features using the optimized monitoring data stream, and obtaining load allocation optimization parameters; dynamically adjusting the generator output allocation ratio based on the load allocation optimization parameters, and generating an optimized output control scheme; updating the configuration parameters of the real-time monitoring system using the optimized output control scheme, and determining a stable control strategy for the power grid operating status.

[0011] Preferably, the system also includes a resource allocation optimization module. Based on the stable voltage output data, it obtains load distribution data of nodes and edge weights from the distribution network, and optimizes the resource allocation path using a graph cut algorithm to obtain the optimized resource allocation path. Specifically, this includes: obtaining node load and edge weight data through the real-time data stream of the distribution network, and generating an initial resource allocation path using a graph cut algorithm; extracting load distribution characteristics along the path based on the initial resource allocation path to determine the load distribution matrix; if any node load in the load distribution matrix exceeds a preset threshold, adjusting the resource allocation ratio using a linear programming algorithm to obtain an optimized resource allocation scheme; updating the edge weight data based on the optimized resource allocation scheme to generate a new network topology; recalculating the resource allocation path using a graph cut algorithm based on the new network topology to obtain an updated optimized path; adjusting the data acquisition frequency based on the updated optimized path to generate a high-frequency monitoring data stream and determine the load balance status of the distribution network; and analyzing the relationship between load distribution and stable voltage using the high-frequency monitoring data stream to determine voltage stability control parameters.

[0012] Preferably, it also includes an overload probability assessment module, which constructs a node and edge weight model of the distribution network for the optimized resource allocation path, simulates the current distribution under the load allocation path through Monte Carlo analysis, and determines the probability of grid overload by combining the overload threshold setting. Specifically, it includes: obtaining node load and edge weight data from the real-time data stream of the distribution network, constructing an initial network topology model, and obtaining the distribution characteristics of node and edge weights; performing multiple random simulations of the load allocation path in the initial network topology model using the Monte Carlo method to obtain the statistical characteristics of the current distribution; if the statistical characteristics of the current distribution show that the current value of a node or edge exceeds the preset overload threshold... If the current value exceeds the overload threshold, the load distribution ratio is adjusted using a linear programming algorithm to obtain an optimized load distribution scheme. Based on the optimized load distribution scheme, the edge weight data in the network topology model is updated to generate a new topology. Using the new topology, the Monte Carlo method is used to re-simulate the current distribution under the load distribution path to obtain updated current distribution characteristics. If the current value of any node or edge in the updated current distribution characteristics exceeds the overload threshold, the load distribution ratio is iteratively adjusted to generate the final load distribution scheme. Based on the final load distribution scheme, the overload probability of the power grid under the current load distribution path is calculated to determine the load balance status of the power grid.

[0013] Preferably, the system also includes a load priority control module. If the grid overload probability is higher than the power supply continuity index requirement, the load allocation priority is adjusted through a feedback control mechanism. Combining the predicted confidence interval and the fault scenario simulation results, the load allocation is optimized to obtain the final load balancing scheme. Specifically, this includes: acquiring grid load data through real-time data streams, calculating the grid overload probability, and determining whether it is higher than the power supply continuity threshold; if the grid overload probability is higher than the power supply continuity threshold, the feedback control mechanism is used to adjust the load allocation priority and generate an initial control parameter adjustment scheme; based on the initial control parameter adjustment scheme and combined with the load prediction model, the confidence interval of the load allocation is calculated to obtain the predicted load distribution characteristics; and through fault scenario simulation, the stability of the predicted load distribution characteristics under different fault conditions is analyzed to generate fault scenario constraints.

[0014] Preferably, the load priority control module, if the grid overload probability is higher than the power supply continuity index requirement, adjusts the load allocation priority through a feedback control mechanism, and optimizes the load allocation by combining the predicted confidence interval and the simulation results of the fault scenario to obtain the final load balancing scheme. This further includes: using a linear programming algorithm, combining the fault scenario constraints and the confidence interval, to optimize the load allocation priority and obtain an optimized load allocation scheme; updating the grid load data by optimizing the load allocation scheme, recalculating the grid overload probability, and determining whether the power supply continuity threshold is met; if the recalculated grid overload probability is still higher than the power supply continuity threshold, iteratively adjusting the control parameters to generate the final load balance state.

[0015] As can be seen from the above technical solution, the present invention has the following beneficial effects:

[0016] This adaptive multiphase parallel power management system addresses the complex scenario of power grids facing simultaneous load surges, voltage fluctuations, and overload risks. It integrates historical and real-time data, extracts features and predicts the probability of load surges through time series decomposition algorithms, and accurately predicts peak values ​​and durations using a multilayer perceptron. Furthermore, it optimizes reserve capacity allocation using linear programming to ensure power supply stability. Regarding voltage fluctuation risks, this invention adjusts generator output through dynamic regulation algorithms, optimizes resource allocation paths using graph cut algorithms, assesses overload probabilities using Monte Carlo analysis, and finally dynamically adjusts load allocation priorities through a feedback control mechanism to achieve load balance. The most significant innovation of this invention lies in the synergy of multiple algorithms and real-time dynamic optimization, which significantly improves power grid operational stability and power supply continuity, while reducing overload and voltage fluctuation risks. Attached Figure Description

[0017] Figure 1 This is a system connection diagram of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figure 1 As shown, this invention provides a technical solution: an adaptive multiphase parallel power management system, the system comprising: a feature extraction module, which extracts periodic, trend, and random features from historical electricity consumption data and real-time sensor data, and uses a time series decomposition algorithm to determine the probability of load surge, obtaining a load surge probability value; a load prediction module, which, if the load surge probability value is higher than a preset threshold, processes real-time load data through a multilayer perceptron structure to predict the peak value and duration of load surge within an hourly or minute-level time window, obtaining a load surge prediction result; a reserve capacity allocation module, which, based on the peak value and duration in the load surge prediction result, obtains available capacity data from a reserve power database, and uses a linear programming algorithm to optimize reserve capacity allocation, obtaining a reserve capacity allocation scheme; and a power grid status monitoring module, which, based on the reserve capacity allocation scheme, obtains power grid operating status data through a real-time monitoring system, detects whether the power grid voltage deviates from the normal range, and determines whether voltage fluctuation risk exists. The voltage stabilization module, upon detecting a risk of voltage fluctuation, adjusts the generator output using a dynamic adjustment algorithm. Combined with predicted load surge peaks, it stabilizes the voltage output, obtaining stable voltage output data. The resource allocation optimization module, based on the stable voltage output data, obtains load allocation data with node and edge weights from the distribution network and optimizes the resource allocation path using a graph cut algorithm, resulting in an optimized resource allocation path. The overload probability assessment module, for the optimized resource allocation path, constructs a node and edge weight model of the distribution network, simulates the current distribution under the load allocation path using Monte Carlo analysis, and determines the grid overload probability based on the overload threshold setting. The load priority control module, if the grid overload probability exceeds the power supply continuity index requirement, adjusts the load allocation priority through a feedback control mechanism. Combining the predicted confidence interval and fault scenario simulation results, it optimizes the load allocation to obtain the final load balancing scheme.

[0020] The system operates on the principle of a real-time monitoring and predictive response mechanism for the power grid's operating status. It collects key parameters such as load, voltage, and current in real time through embedded sensors and edge computing nodes. The feature extraction module first decomposes these data into periodic, trend, and random disturbance features, and then establishes a probabilistic model based on historical statistics of sudden load spikes to determine the likelihood of such events. Once the probability of a spike exceeds a threshold, the system triggers a neural network model to predict the peak value and trend of the short-term load curve, providing input for capacity allocation. The standby capacity allocation module then calls upon information on various energy storage devices or backup power sources in the database, optimizing scheduling through linear programming to improve response speed and power supply efficiency. Simultaneously, the power grid status monitoring module continuously analyzes voltage deviations. If an anomaly is detected, a voltage regulation algorithm is activated to dynamically adjust generator output, ensuring voltage stability. The resource allocation optimization module, based on the voltage regulation results and the load network topology, uses graph theory optimization to redistribute power supply paths, reducing network impedance and transmission losses. Subsequently, the system simulates the load current distribution under the new resource allocation scheme using the Monte Carlo method and assesses the power supply reliability risk based on the overload threshold. If the risk is too high, the load priority control module adjusts the power supply priority according to the importance level of different service loads to ensure continuous power supply to critical loads, ultimately forming a dynamically adjustable end-to-end load balancing scheme. The entire system implements a closed-loop control logic from load identification, prediction, response, stabilization to evaluation and optimization in a hierarchical control manner, improving the adaptive adjustment capability and disturbance rejection capability of the power distribution system.

[0021] The feature extraction module extracts periodic, trend, and random features from historical electricity consumption data and real-time sensor data. It then uses a time series decomposition algorithm to determine the probability of a load surge, obtaining the load surge probability value through the following steps: First, it obtains the raw dataset from the historical electricity consumption data and real-time sensor data. Second, it uses preprocessing techniques to remove noise and missing values ​​to obtain a standardized dataset. Third, based on the standardized dataset, it uses a classic time series decomposition algorithm to extract periodic, trend, and random features, obtaining a decomposed feature set. Finally, for the decomposed feature set, if the amplitude of the periodic feature exceeds a preset threshold, it is marked as a high-periodic fluctuation, resulting in a periodicity feature set. A fluctuation label set is generated. Based on the decomposed feature set and the periodic fluctuation label set, the slope change of the trend feature is calculated. If the absolute value of the slope change is greater than a preset threshold, it is marked as a trend change, and a trend change label set is obtained. Random features are extracted from the decomposed feature set, and the random fluctuation intensity is calculated using a Gaussian distribution fitting method to obtain the random fluctuation intensity value. Based on the periodic fluctuation label set, the trend change label set, and the random fluctuation intensity value, a logistic regression algorithm is used to calculate the probability of a load surge, and a load surge probability value is obtained. If the probability value of the load surge is greater than a preset threshold, a surge risk assessment is triggered, and a risk assessment result is obtained.

[0022] In this embodiment, during the system initialization phase, the feature extraction module first calls the historical database of the power system to retrieve historical electricity consumption data for the past 30 consecutive days. This data is stored with a sampling period of 5 minutes, forming 288 data points per day, for a total of 8640 sample points. Simultaneously, the module acquires real-time sensor data through the edge acquisition unit, including the active power, reactive power, voltage, current, and frequency signals of the current load. The sampling frequency is consistent with the historical data, maintaining a sampling frequency of once every 5 minutes to ensure data time alignment. After obtaining the original dataset, the system performs data preprocessing. First, it uses a median filtering method with a window length of 5 on all data, taking the median value of every 5 consecutive data points as the smoothed value of the center data point to remove outliers and spike noise. Then, for data missing regions caused by communication interruptions, a linear interpolation method is used to fill in the missing points. That is, the linear trend of the two nearest valid data points before and after the missing point is calculated and inserted as an estimate of the missing point. After noise removal and missing point filling are completed, the system performs maximum and minimum value normalization on all data, transforming each data point to the interval of 0 to 1 according to the ratio between its minimum and maximum values ​​in the sequence, thus obtaining a standardized dataset.

[0023] Subsequently, the system performs time series decomposition processing on the standardized dataset, employing a seasonal trend decomposition algorithm based on local weighted regression. The specific steps are as follows: A 24-hour periodic window is used, with each window containing 288 sampling points. The system fits the periodic component to the data within this window; then, a trend component is extracted using a weighted moving average, which continuously calculates the average load over the past 6 hours and smooths it to generate a trend curve; the remaining portion, the residual after removing the periodicity and trend, is considered a random component. After obtaining the periodic component, the system analyzes its amplitude window by window, calculating the difference between the maximum and minimum values ​​within each 24-hour period. If this difference exceeds 0.15, the period is marked as a high-periodic fluctuation. The 0.15 threshold is calculated using the 90th percentile of the periodic amplitude in the power system load data over the past 365 days and verified as representative through operational scenario testing. For the trend component, the system uses the slope change method to identify abrupt changes. This involves calculating the difference between the current and previous values ​​in the trend sequence every 30 minutes and dividing by the time interval to obtain the rate of trend change per unit time. A rate exceeding 0.08 is considered a trend abrupt change. This threshold is obtained by rounding down to 1.5 times the standard deviation of the trend change sequence obtained through continuous simulation. For the random component, the system uses maximum likelihood estimation to perform Gaussian fitting on the residual sequence to obtain the mean and standard deviation. The standard deviation represents the intensity of random fluctuations. A standard deviation exceeding 0.12 is considered a high-intensity random disturbance. This threshold is determined by the 95th percentile of the standard deviation of historical residual samples. These three labels constitute the periodic fluctuation label set, the trend abrupt change label set, and the random fluctuation intensity value set, respectively.

[0024] Next, the system uses the three datasets mentioned above as input variables and feeds them into a pre-trained logistic regression model. The training data for this model is selected from records of whether load surge events occurred within the past 90 days. Input features include whether periodic fluctuations are of high amplitude, whether the trend slope changes abruptly, and whether the residual standard deviation exceeds a threshold, encoded as 1 or 0 as discrete variables, and a regression function is established accordingly. The model training process uses the least squares method, with the final fixed regression weights being 0.3 for periodic fluctuation features, 0.4 for trend change features, and 0.3 for random fluctuation features. After substituting the current input features into the model, the calculated result is the load surge probability value at the current moment, a real number between 0 and 1. If this probability value is higher than the preset surge risk threshold of 0.65, it is considered that there is a current load surge risk, and the system immediately sends a surge warning signal to the upper-level risk response module, initiating subsequent capacity scheduling and voltage stabilization processes. The aforementioned threshold of 0.65 was determined comprehensively after balancing the model's accuracy, recall, and false alarm rate in a power system operation simulation platform. It has the best effect in avoiding false alarms while improving the ability to identify real sudden events.

[0025] The load prediction module, if the probability of a load surge exceeds a preset threshold, processes real-time load data using a multilayer perceptron structure to predict the peak load surge and its duration within an hourly or minute-level time window. The load surge prediction results include: acquiring raw load data from real-time sensors and segmenting the data using a sliding window technique to obtain segmented load datasets; calculating load fluctuation characteristics within each time window based on the segmented load datasets to obtain a fluctuation feature set; if the peak change rate in the fluctuation feature set exceeds a preset threshold, training the fluctuation feature set using a multilayer perceptron model to obtain a load surge prediction model; processing real-time load data using the load surge prediction model to predict the load peak within an hourly time window to obtain an hourly peak prediction result; calculating the load change trend within a minute-level time window based on the hourly peak prediction result to obtain minute-level trend characteristics; if the absolute value of the slope of the minute-level trend characteristics exceeds a preset threshold, classifying the minute-level trend characteristics using a random forest algorithm to determine the duration of the load surge and obtain a duration prediction result; and fusing the hourly peak prediction result and the duration prediction result to generate a load surge prediction report, resulting in the final prediction output.

[0026] When the system detects that the probability of a sudden load increase exceeds a preset threshold, the load prediction module immediately initiates the prediction process. The system first acquires raw load data for the current hour from real-time power sensors deployed within the distribution network. This data is collected once per minute, forming a load time series containing 60 data points. To facilitate subsequent feature extraction, the system uses a sliding window technique to process this time series, with a sliding window length of 10 minutes and a sliding step size of 1 minute. Specifically, starting from the first data point, 10 consecutive data points are selected to form the first window; then, starting from the second data point, another 10 consecutive data points are selected to form the second window; and so on, until the entire time series is divided into multiple overlapping subsequences of length 10. These subsequences constitute the segmented load dataset.

[0027] Subsequently, the system extracts load fluctuation features for each subsequence. Specifically, within each 10-minute subsequence, the system first identifies the maximum and minimum load values ​​and calculates the difference between them as the fluctuation amplitude. Then, it subtracts the maximum value of the previous window from the maximum value of the current window, and divides the difference by the maximum value of the previous window to obtain the peak change rate. If this peak change rate is greater than a set threshold of 0.18, the system determines that the subsequence has significant fluctuations and triggers the prediction process of the multilayer perceptron structure. The threshold is set as follows: based on the load data of the past 365 days, the peak change rate within all sliding windows is statistically analyzed, the 90th percentile of its distribution is taken, and combined with the simulation prediction accuracy evaluation, it is finally determined to be 0.18.

[0028] After satisfying the fluctuation amplitude and peak change rate conditions, the system uses five key features extracted within the window as input: fluctuation amplitude, maximum value location, average load change rate, ratio of maximum to minimum value, and difference between the current maximum value and the maximum value of the previous window. This 5-dimensional feature vector is then input into a pre-trained multilayer perceptron model. This model has a 3-layer structure, comprising one input layer, two hidden layers, and one output layer. The input layer has 5 nodes, each hidden layer has 64 nodes, and the output layer has 1 node. The model training process uses all subsequences within the past 180 days that satisfy a peak change rate exceeding 0.18 as training samples, with the training label being the corresponding maximum load value occurring within the next hour. Mean squared error is used as the loss function during training, and the optimization algorithm is gradient descent with momentum. The initial learning rate is set to 0.01, and the training batch size is 100 epochs, with each epoch iterating through all training samples once. The final output is the hourly peak prediction result, which is the predicted maximum load value within a 60-minute time window.

[0029] After completing hourly load peak prediction, the system continues with minute-level load trend analysis based on the currently obtained real-time load sequence. The system uses a 5-minute time window, a sampling frequency of 1 minute, and a 1-minute sliding window step, calculating the linear fitting slope within each segment of the window. The resulting slope represents the minute-level trend characteristic. To determine if there is a significant and sustained surge trend, the system compares the absolute value of this trend characteristic with a set threshold of 0.12, which is obtained by multiplying the standard deviation of historical trend characteristics by 1.5 and rounding up. If the absolute value of the trend characteristic exceeds 0.12, the system initiates a duration prediction process.

[0030] The duration prediction process uses a trained random forest model for classification. Input features include: minute-level trend slope, 5-minute fluctuation amplitude, and the difference between the current peak and the trend start point. The model output is the duration level, categorized into three classes: less than 10 minutes, 10 to 30 minutes, and greater than 30 minutes. The model is trained on minute-level time series samples of all load surge events occurring within the past 180 days. An information gain criterion is used to construct decision trees, with a maximum tree depth of 10, resulting in a forest of 100 trees. The final prediction is the possible duration of the current load surge event in the future.

[0031] Finally, the system merges the hourly load peak prediction results with the minute-level duration prediction results to generate a complete load surge prediction report, including the expected maximum load value and corresponding duration category for the next hour, which is used as input for subsequent standby capacity scheduling and resource allocation modules. The above prediction process automatically starts whenever the probability value of a load surge exceeds 0.65, where 0.65 is a probability threshold set based on system operating experience and historical false alarm rates, ensuring that the prediction system has sufficient sensitivity to abnormal events without generating too many false alarms.

[0032] The standby capacity allocation module, based on the peak load and duration data from the load surge prediction results, retrieves available capacity data from the standby power database and optimizes standby capacity allocation using a linear programming algorithm. The resulting standby capacity allocation scheme includes: obtaining peak load and duration data from the load surge prediction results and querying the standby power database through a database interface to obtain standby capacity information; constructing a linear programming model based on the standby capacity information, where the objective function is to maximize allocation efficiency, and constraints include peak load and duration, resulting in an optimized allocation model; processing the standby capacity information using the linear programming model to calculate the allocation ratio of each power unit, resulting in a preliminary capacity allocation scheme; if the allocation ratio of any power unit in the preliminary capacity allocation scheme exceeds a preset threshold, adjusting the allocation ratio using a greedy algorithm to prioritize meeting peak load demands, resulting in an adjusted allocation scheme; acquiring real-time power status data based on the adjusted allocation scheme to determine whether each power unit meets the capacity requirements of the allocation scheme, resulting in a power availability judgment result; updating the capacity status in the standby power database based on the power availability judgment result, generating a final capacity allocation scheme; and scheduling the power management system to execute allocation operations based on the final capacity allocation scheme, determining the allocation execution status.

[0033] First, the system receives the load surge prediction result from the preceding load prediction module. This result contains two key parameters: the load peak value and the duration. The load peak value represents the maximum power demand that may occur within the predicted time period, in kilowatts (kW), and the duration represents the duration of this maximum power demand, in minutes. Upon receiving this data, the system immediately queries the backup power database interface for real-time information on all backup power units currently in a dispatchable state. This information includes data such as the maximum remaining output power, maximum continuous output duration, startup response delay time, and current operating status for each power unit. The maximum remaining output power represents the maximum power that the power unit can allocate without affecting its stable operation. It is generally calculated by the real-time load monitoring system using the rated power minus the currently allocated power, in kilowatts. The maximum continuous output duration is determined by the equipment's historical operating records and the power supply thermal stability model. The response delay time is the average value of the last startup response record, in seconds.

[0034] Subsequently, the system constructs a linear programming model with load peak and duration as the main constraints. The objective function is defined as maximizing the efficiency of reserve capacity allocation, specifically the contribution of a unit of reserve power output to the total predicted load. During modeling, the system sets the load peak as the overall allocation target, and the sum of the allocated power of all selected reserve power units must be equal to or slightly higher than this peak. Simultaneously, the allocated power of each power unit must not exceed its maximum remaining output power; the continuous operating time of each power unit must be greater than or equal to the predicted duration; and the response delay of the power unit must be less than the maximum response threshold specified by the system, typically set at 5 seconds. This value is a predetermined value based on the response speed capability of the power grid control system, aiming to ensure the timeliness and accuracy of dispatch response. Furthermore, the system sets a minimum operating efficiency threshold of 80% for each power unit. This value is derived from stability experiments of this type of power equipment under load variation scenarios; power units with efficiencies below this threshold will not participate in allocation modeling.

[0035] After the model is established, the system uses the simplex method or an equivalent solver to obtain a preliminary capacity allocation scheme, i.e., the allocation ratio corresponding to each power supply unit. For example, for a scenario with a predicted peak load of 150 kW, if a power supply unit is allocated a ratio of 0.2 and its rated power is 100 kW, then it needs to undertake a power supply task of 20 kW. After allocation, the system verifies the allocation ratio of all power supply units one by one. If the allocated power of any power supply unit exceeds 80% of its maximum remaining power, the system considers the unit to be in an overload risk state. This 80% threshold comes from the recommended safe load range provided by the equipment manufacturer and is usually fixed in actual operation based on equipment performance test data. For power supply units exceeding this threshold, the system will trigger a greedy adjustment mechanism. The adjustment logic is as follows: first, sort by power supply response speed and operating efficiency, and prioritize the selection of power supply units with short response time and high efficiency to fill the allocation gap of the removed power supply units. If the peak load demand still cannot be met in the current resource pool, the system will trigger an early warning mechanism to prompt the management personnel to supplement backup power.

[0036] After adjustments are completed, the system enters the allocation availability verification phase. During this phase, the system acquires real-time operating status information for each power unit, including whether it is online, under maintenance, has a fault flag, and is being used by other load tasks. All status data is provided by the power monitoring system. The system generates a power availability judgment result based on whether the power supply meets the execution conditions. Only power units that meet all conditions are retained in the final allocation list. Finally, based on the verified final standby capacity allocation scheme, the system sends allocation commands to the control modules of each power unit, scheduling them to start power supply at the specified time. The system records and confirms the successful start-up status of each unit, ensuring the integrity of the execution and the implementation of closed-loop control. After this operation is completed, the system synchronously updates the remaining capacity and status information of each power unit in the standby power database, forming a complete record of one capacity allocation operation.

[0037] The power grid status monitoring module, based on the reserve capacity allocation scheme, acquires power grid operating status data through a real-time monitoring system to detect whether the power grid voltage deviates from the normal range and determine whether voltage fluctuation risks exist. This includes: collecting power grid operating status data through the real-time monitoring system and generating a voltage data stream; if the voltage data exceeds a preset threshold range, a voltage anomaly detection algorithm is used to determine the voltage fluctuation risk and obtain a risk assessment result; based on the risk assessment result, an anomaly alarm mechanism is used to generate alarm signals and determine alarm priorities; based on the alarm priorities, power grid load level data is acquired to determine the power grid stability status; if the power grid stability status is lower than a preset stability threshold, a support vector machine algorithm is used to analyze the voltage data stream and the power grid load level to obtain stability optimization parameters; based on the stability optimization parameters, the data acquisition frequency is adjusted to generate an optimized monitoring data stream; and based on the optimized monitoring data stream, the real-time monitoring system configuration is updated to determine a dynamic adjustment scheme for the power grid operating status.

[0038] In this implementation, the power grid status monitoring module first collects power grid operating status data through a real-time monitoring system deployed within the power system. The data collection targets are the voltage values ​​of all multi-phase parallel power supply nodes. The data collection cycle is once per second, and each sampling records the instantaneous voltage value of each node. The sampling equipment is directly connected to the system controller via a communication interface. All collected voltage values ​​constitute a continuous voltage data stream for subsequent analysis and judgment. The system presets a normal voltage range threshold of 5% above and below the rated voltage. For example, in a 220V system, this range is set to 209V to 231V. This threshold is specified by the State Grid standard and stored in the system configuration file, and is not dynamically adjusted with system operation. All monitoring nodes use a unified standard.

[0039] The system compares the voltage value at each sampling point with a preset range. If any voltage value is found to be below 209 volts or above 231 volts, it is considered abnormal voltage data. When the system detects an abnormal voltage at a node for two consecutive sampling cycles (i.e., two consecutive abnormal values ​​are collected within 2 seconds), the system invokes the voltage anomaly detection algorithm. This algorithm does not involve model training; it only performs logical judgment, that is, it statistically records whether the voltage value continuously exceeds the limit and generates a voltage fluctuation event record, which includes the node number, anomaly timestamp, actual voltage value, and deviation ratio. The deviation ratio is calculated by dividing the difference between the actual value and the rated value by the rated value. For example, if the voltage value is 240 volts, the deviation ratio is 9.09%.

[0040] After a voltage fluctuation event is recorded, the system automatically activates the anomaly alarm mechanism, generating an alarm signal containing anomaly details. The alarm signal level is divided into three levels: high, medium, and low, determined by two parameters: deviation ratio and anomaly duration. If the deviation ratio exceeds 10% and the anomaly duration is greater than 5 seconds, it is classified as high-level; if the deviation ratio is between 7% and 10% and the anomaly duration is greater than 3 seconds, it is classified as medium-level; all others are classified as low-level. Based on the alarm level, the system further calls the load level analysis module to read the current load power data of the node. This data is collected in real-time by a current sampler, and the power value is calculated. This value is then compared with the node's rated power to obtain the load utilization rate. If the load utilization rate is greater than 90%, the node is considered to be operating under high voltage. This 90% threshold is preset according to the power system safety design principles and is fixedly stored in the system settings.

[0041] When the system simultaneously determines that the alarm level of a node is high or medium and the load utilization rate exceeds 90%, indicating poor grid stability, the system invokes a support vector machine (SVM) algorithm to jointly analyze the current node's voltage data stream and load level data. This SVM model is trained offline, with historical voltage deviation ratios and load utilization rates as inputs, and a suggested data sampling frequency adjustment ratio as output. The model calculates by mapping the current input features to a high-dimensional space, classifying them, identifying whether the sampling frequency needs adjustment, and outputting specific adjustment suggestions.

[0042] If the model determines that the current stability is insufficient, the data sampling frequency needs to be increased from once per second to once every 0.67 seconds (1.5 times per second). The system controller then sends an update command to the sampling device via an interface to modify the data acquisition configuration. After the sampling frequency update, the new data stream begins to continuously collect voltage data according to the optimized cycle, and this frequency adjustment behavior is recorded in the monitoring system configuration file. The optimized data stream can respond to voltage fluctuations more promptly, improving the real-time performance and accuracy of grid status perception. Ultimately, through continuous monitoring and analysis, the system develops a dynamic adjustment scheme for abnormal grid voltage responses, achieving specific control over improving the stability of the power system.

[0043] The voltage stabilization module, upon detecting a risk of voltage fluctuation, adjusts the generator output using a dynamic adjustment algorithm. Combined with predicted load surge peaks, it stabilizes the voltage output, obtaining stable voltage output data including: if a voltage fluctuation risk is detected, a support vector machine algorithm is used to analyze the load surge peak and adjust the generator output to obtain stable voltage output data; based on the stable voltage output data, real-time monitoring data streams of the power grid operating status are acquired, and the voltage data change trend is analyzed to determine the stability of the power grid operating status; if the stability of the power grid operating status is lower than a preset threshold, a random forest algorithm is used to process the real-time monitoring data stream, extracting voltage fluctuation characteristics to obtain voltage fluctuation characteristic parameters; based on the voltage fluctuation characteristic parameters, the data acquisition frequency is adjusted to generate an optimized monitoring data stream, determining a dynamic adjustment scheme for the power grid operating status; through the optimized monitoring data stream, the relationship between the power grid load level and voltage fluctuation characteristics is analyzed to obtain load allocation optimization parameters; based on the load allocation optimization parameters, the generator output allocation ratio is dynamically adjusted to generate an optimized output control scheme; through the optimized output control scheme, the configuration parameters of the real-time monitoring system are updated to determine the stable control strategy for the power grid operating status.

[0044] When the power grid status monitoring module compares and analyzes the collected real-time voltage data streams and determines that there is a risk of voltage fluctuation in the power grid (i.e., the voltage deviates from the normal operating voltage range by more than 3% within two or more consecutive sampling periods), the system immediately activates the voltage stabilization module. First, the system uses the peak load surge output from the load prediction module as input data. This peak value, measured in kilowatts, represents the maximum load value within the prediction time window. This data is a predicted value obtained through a multilayer perceptron model after segmenting historical load data into sliding windows, extracting fluctuation features, and then analyzing the data. The system reads this peak value and compares it with the actual power output of each generator unit to calculate the power difference. For example, if the predicted peak load is 1600 kilowatts and the current total power output is 1300 kilowatts, then an additional 300 kilowatts of power generation capacity needs to be increased. Subsequently, the system activates a support vector machine model for discrimination, classifying the generator units based on historical samples of similar load surge events and outputting the specific percentage increase required for each generator, such as 10% for generator A and 5% for generator B. This ratio is derived from the model's calculations of the input features, which include the current node's load forecast, the generator set's current status, and its response capability level. Each output ratio value is multiplied by the corresponding generator's current rated power to determine the actual power increase required, updating the control commands for each generator set and completing the output adjustment.

[0045] After output adjustment, the system will monitor the latest data stream in real time and input it into the analysis module. It will extract the voltage change curve over a ten-second period, including the maximum, minimum, and average voltage values ​​and their slopes. Combined with load level and frequency changes, the system will determine the current grid stability. The stability threshold is defined as follows: the voltage change amplitude must not exceed 3% within ten consecutive seconds, the change frequency must not exceed once per second, and the average voltage value must not deviate from the rated value by more than 1%. If any condition is not met, the grid's operational stability is deemed insufficient. The system will immediately trigger a random forest model to classify the current data stream and extract voltage fluctuation characteristic parameters. The input features of the random forest model include the voltage standard deviation, mean voltage slope, frequency fluctuation range, load growth rate, and the voltage anomaly label from the previous period within a ten-second window. The model's output voltage fluctuation characteristic parameters include fluctuation type, severity level, and suggested adjustment method label.

[0046] Based on the aforementioned voltage fluctuation characteristic parameters, the system determines whether the current data sampling frequency of the monitoring system meets the stability control requirements. For example, if the fluctuation type is a sudden increase and the severity level is three, the system needs to increase the current sampling frequency from once per second to five times per second to ensure high-frequency capture of rapidly changing voltages and avoid information loss. The optimized sampling frequency is written into the monitoring system configuration file and takes effect immediately, generating an optimized data stream for subsequent analysis modules. The system then further calculates the correlation coefficient between the current load level of each node and its voltage fluctuation amplitude based on the optimized data stream. When the absolute value of the correlation coefficient exceeds 0.6, the system determines that there is a strong correlation between the load and the voltage fluctuation.

[0047] The system then converts this correlation coefficient into load allocation optimization parameters, specifically including the node number for which the load should be reduced, the reduction value (in kilowatts), and the adjustment time range. For example, if a node currently has a load of 1200 kilowatts, and the system suggests reducing it by 200 kilowatts within the next three minutes, the load optimization parameters would be the node number, 200 kilowatts, and three minutes. Based on these parameters, the system dynamically adjusts the generator output ratio, using a weighted average method to evenly distribute the load reduction to generators with adjustment capabilities, generating a new output control scheme, recording it in the scheduling database, and then sending it to the field equipment for execution by the control module.

[0048] Finally, based on this output control scheme, the system updates all configuration parameters of the real-time monitoring system, including monitoring frequency, voltage stability threshold, generator response delay time, etc., and generates a complete grid operation stability control strategy. This ensures that the voltage stability control process has traceability and automatic correction capabilities, achieving timely suppression of voltage fluctuations and improving system operation safety. The output of this strategy will continue to serve as input for a new round of analysis in the next monitoring cycle, realizing dynamic closed-loop control.

[0049] It also includes a resource allocation optimization module, which, based on stable voltage output data, obtains load distribution data of nodes and edge weights from the distribution network, and uses a graph cut algorithm to optimize the resource allocation path to obtain an optimized resource allocation path. Specifically, this includes: obtaining node load and edge weight data through the real-time data stream of the distribution network, and generating an initial resource allocation path using a graph cut algorithm; extracting load distribution characteristics on the path based on the initial resource allocation path to determine the load distribution matrix; if any node load in the load distribution matrix exceeds a preset threshold, adjusting the resource allocation ratio using a linear programming algorithm to obtain an optimized resource allocation scheme; updating the edge weight data based on the optimized resource allocation scheme to generate a new network topology; recalculating the resource allocation path using a graph cut algorithm based on the new network topology to obtain an updated optimized path; adjusting the data acquisition frequency based on the updated optimized path to generate a high-frequency monitoring data stream to determine the load balance status of the distribution network; and analyzing the relationship between load distribution and stable voltage through the high-frequency monitoring data stream to determine voltage stability control parameters.

[0050] During the execution of the resource allocation optimization module, the latest stable voltage output data is first obtained from the stable voltage regulation module. This data is collected twice per second by the generator set's output voltage detection device, including the voltage value, current value, and timestamp of each node, serving as input parameters for the distribution network status. Subsequently, the system obtains the load data of each power supply node and the edge weight data of the connecting lines between nodes through the distribution network monitoring module. The node load data is the instantaneous power value in kilowatts, measured by the power meter of the distribution node; the edge weight data consists of two parts: first, the equivalent impedance value of the cable line, obtained from a table based on the cable length and diameter; second, the load ratio of the line, which refers to the ratio of the current flowing through the cable to its rated current, calculated as the percentage obtained by dividing the current current by the rated current. The final edge weight is the impedance value multiplied by the load ratio.

[0051] After collecting node and edge weight data, the system constructs a power distribution network graph model, where nodes represent power supply nodes and edges represent power connection paths between nodes, with the previously calculated weight value labeled on each edge. Next, the system applies a graph cut algorithm to process the graph model, aiming to minimize the total path weight. Specifically, it finds a set of paths connecting the main power supply to all load nodes such that the sum of the weights of all edges along the path is minimized, thus deriving the initial resource allocation path. In this step, the graph cut algorithm performs maximum flow minimum cut calculations on the entire graph to determine the power supply path structure and outputs a list containing all nodes and their assigned paths.

[0052] After obtaining the initial resource allocation path, the system extracts the load values ​​of each node along the path and generates a load distribution matrix. This matrix is ​​stored as a two-dimensional array, with each row representing a power supply node and each column representing the load value at different times. Subsequently, the system performs a node-by-node analysis of the load distribution matrix and compares it with a preset load threshold. This threshold is 90% of the node's rated capacity to ensure a safe operating boundary. For example, if a node's rated capacity is 150 kilowatts, its threshold is set to 135 kilowatts. This parameter is stored in the system configuration file and set during the initialization phase.

[0053] If the load on a node exceeds its corresponding threshold, the system initiates a linear programming algorithm to reallocate resources. In this model, the objective function is to minimize the load adjustment of each node, and the constraints include: first, the total allocated load does not exceed the maximum output power of the power supply; second, the adjusted load value of each node does not exceed its threshold; and third, the original load of a node cannot be reduced by more than 20%. This linear programming problem is solved using the simplex method to determine the resource allocation ratio adjustment for each node, thus forming a new resource allocation scheme.

[0054] Based on the optimized resource allocation scheme, the system recalculates the weights of the connecting edges. The new weight of each edge is the product of the updated load current and the original line impedance. The current is calculated by dividing the new load power by the voltage. For example, if the new load of a node is 120 kilowatts and the voltage is 400 volts, the current will be 300 amps. If the line impedance is 0.1 ohms, the weight will be 30. After all edge weights are updated, the system generates the latest network topology graph based on the new weights and executes the graph cut algorithm again to redetermine the resource allocation path on this structure, obtaining the updated optimized path.

[0055] Based on the new path, the system increases the data acquisition frequency to five times per second to obtain a high-frequency monitoring data stream. This data stream covers voltage, current, and load changes at all nodes at different time points. The system processes this data stream in real time, statistically analyzing the voltage fluctuation amplitude and frequency of each node per unit time, and matching it with stable voltage output data to calculate the node voltage stability coefficient. This coefficient is defined as the ratio of the standard deviation to the mean of the voltage at a node per unit time; a smaller value indicates more stable voltage. If the stability coefficient of a node exceeds 10%, it indicates poor stability. The system analyzes the correlation between load distribution and voltage stability based on this data, calculates the regression coefficient between the load change rate and voltage change rate of each node, and extracts voltage stability control parameters. These parameters are used to control the generator output adjustment strategy and future data acquisition frequency adjustment logic. Finally, all allocation and adjustment processes are stored in the system log and fed back to the main control unit for updating the allocation strategy for the next cycle.

[0056] It also includes an overload probability assessment module, which constructs a node and edge weight model of the distribution network for the optimized resource allocation path. It simulates the current distribution under the load allocation path using Monte Carlo analysis, and, combined with an overload threshold setting, determines the probability of grid overload. Specifically, this includes: obtaining node load and edge weight data from the real-time data stream of the distribution network, constructing an initial network topology model, and obtaining the distribution characteristics of node and edge weights; using the Monte Carlo method to perform multiple random simulations of the load allocation path in the initial network topology model to obtain the statistical characteristics of the current distribution; if the statistical characteristics of the current distribution show that the current value of a node or edge exceeds a preset overload threshold... The load allocation ratio is adjusted using a linear programming algorithm to obtain an optimized load allocation scheme. Based on the optimized load allocation scheme, the edge weight data in the network topology model is updated to generate a new topology. Using the new topology, the Monte Carlo method is used to re-simulate the current distribution under the load allocation path to obtain updated current distribution characteristics. If the current value of any node or edge in the updated current distribution characteristics exceeds the overload threshold, the load allocation ratio is iteratively adjusted to generate the final load allocation scheme. Based on the final load allocation scheme, the overload probability of the power grid under the current load allocation path is calculated to determine the load balance status of the power grid.

[0057] In this embodiment, the overload probability assessment module, based on the optimized resource allocation path and under real-time power grid operation, first collects node load data and edge weight data from the distribution network using built-in sensors and communication devices. Node load data refers to the actual active power load value at a distribution node per unit time, measured in kilowatts, obtained in real-time by voltage and current sensors, and weighted averaged according to each monitoring cycle (e.g., every 10 seconds) using a data aggregation algorithm. Edge weight data represents the transmission capacity limitation of the line between two nodes. The value is determined by the line's impedance and the maximum allowable current. The impedance is calculated based on the line type, length, and material parameters, while the maximum current is determined based on the rated current carrying capacity of the line equipment, typically preset in the equipment manual or power grid design parameters.

[0058] After data acquisition, the system constructs an initial network topology model based on the node connections of the current network. In the model, each node represents a load point or power supply point, and each edge represents a transmission line, with corresponding load and weight parameters assigned to nodes and edges. Subsequently, the system uses the Monte Carlo simulation method to simulate the current distribution under the load distribution path in the constructed topology model. Specifically, during the simulation, the load of each node is perturbed according to a Gaussian distribution, and the standard deviation is set to the average load change rate of that node over the past 30 minutes based on the original acquired values. The total number of simulation samples is set to 1000.

[0059] In each simulation, the system progressively adds node loads according to the topology path and calculates the current value of each edge based on the edge weights. The current value is derived by inversely calculating the relationship between power and voltage, with the voltage in the calculation formula fixed at the rated voltage, assumed to be 380 volts. During the statistical analysis phase, the system records the current distribution of each node and edge in all simulations, including the average, maximum, minimum, and standard deviation. If the analysis finds that the maximum current of some nodes or edges exceeds the overload threshold, the system will trigger a load adjustment procedure. Here, the overload threshold refers to 120% of the rated maximum current carrying capacity of a single line or device. This value is set to ensure the safe operation of the equipment and is based on the State Grid equipment operation standards and historical operating data.

[0060] Upon triggering, the system adjusts the load distribution ratio using a linear programming algorithm. Specifically, the objective function is set to minimize the system's maximum edge current, with constraints including a constant total load, current on each edge not exceeding its rated capacity, and load adjustment not exceeding 20% ​​of its historical average load change. The algorithm uses the simplex method to obtain a new load distribution ratio that satisfies the above constraints. Based on the optimization results, the system reassigns the weights of each edge to reflect the updated load and current ratios.

[0061] Subsequently, the system repeats the Monte Carlo simulation based on the newly generated topology, with the process being completely identical to the initial simulation, to verify whether the load adjustment effectively eliminates the overload phenomenon. If any node or edge current still exceeds the threshold, the system enters the iterative process, calling linear programming again for adjustment and updating the weights, until the maximum current values ​​of all nodes and edges in the simulation results are within the allowable range, or 5 iterations have been performed. At this point, the system automatically records the optimal feasible solution and stops the simulation.

[0062] Finally, the system counts the number of times the current exceeded the overload threshold in all simulation results and divides this number by the total number of simulations to obtain the overload probability for each path. For example, if a path experiences 25 current overloads in 1000 simulations, the overload probability for that path is 2.5%. Based on the overload probabilities of all paths, the system assesses the overall load balance of the power grid. If the overload probability of all paths is below 5%, the power grid is considered to be operating stably; otherwise, an alarm signal is issued indicating the need for further optimization of the load distribution strategy or increase of reserve capacity. These steps ensure accurate assessment results, reasonable parameter settings, and standardized simulation procedures, thereby meeting the requirements for high-reliability power grid operation monitoring.

[0063] It also includes a load priority control module. If the grid overload probability is higher than the power supply continuity index requirement, the load allocation priority is adjusted through a feedback control mechanism. Combining the predicted confidence interval and the simulation results of fault scenarios, the load allocation is optimized to obtain the final load balancing scheme. Specifically, this includes: acquiring grid load data through real-time data streams, calculating the grid overload probability, and determining whether it exceeds the power supply continuity threshold; if the grid overload probability is higher than the power supply continuity threshold, the feedback control mechanism is used to adjust the load allocation priority and generate an initial control parameter adjustment scheme; based on the initial control parameter adjustment scheme and combined with the load prediction model, the confidence interval of the load allocation is calculated to obtain the predicted load distribution characteristics; through fault scenario simulation, the stability of the predicted load distribution characteristics under different fault conditions is analyzed to generate fault scenario constraints; using a linear programming algorithm, combined with the fault scenario constraints and confidence interval, the load allocation priority is optimized to obtain an optimized load allocation scheme; by optimizing the load allocation scheme, the grid load data is updated, the grid overload probability is recalculated, and it is determined whether the power supply continuity threshold is met; if the recalculated grid overload probability is still higher than the power supply continuity threshold, the control parameters are iteratively adjusted to generate the final load balance state.

[0064] In this embodiment, the load priority control module first acquires the load data of each node in the current power grid through real-time data streams, including the instantaneous active power and reactive power values ​​of each node, in kilowatts and kilovars, respectively. The system then determines whether the acquired data and the overall power grid overload probability value output by the overload probability assessment module exceed the power supply continuity threshold. This power supply continuity threshold is set at 2%, and its value is determined based on the continuous power supply requirements for national level-one loads. If the system detects that the current power grid overload probability is greater than 2%, it immediately activates the feedback control mechanism.

[0065] The feedback control mechanism first classifies all load nodes in the power grid into three levels: Level 1 consists of critical loads with the highest requirements for power supply continuity, such as hospitals and dispatch centers; Level 2 includes loads in ordinary residential areas and schools; and Level 3 comprises highly controllable loads such as industrial plants and commercial loads. The level of each load node is determined in the system initialization configuration file based on load type and historical power supply priority, and cannot be dynamically changed after initialization. Based on the current overload probability state, the system sets initial control parameters for load nodes of different levels, including the maximum allowable load percentage and power reduction ratio for each level. For example, Level 1 loads are set to have a maximum allowable load of 100% with no reduction; Level 2 is set to 90%; and Level 3 is set to 70%.

[0066] Next, the system invokes the load prediction model to predict the load change trend of each node within the next 30 minutes, and sets a confidence interval based on the average absolute percentage error generated from the model's historical running data. This error value is derived from the comparison between the predicted and actual loads collected by the system over the past 7 days, with an average error of 10%. Therefore, the upper and lower limits of the confidence interval are the predicted value plus or minus 10%. Based on this, the system constructs upper and lower limit fluctuation ranges for each predicted load sample, generating a complete predicted load distribution feature dataset.

[0067] Subsequently, the system sequentially introduces multiple common fault events through the fault scenario simulation module, including line breakage, sudden voltage drop at nodes, and switch failure to operate. A total of five typical fault types are set, with each fault simulated 50 times, resulting in 250 fault scenarios to ensure coverage of major system operational anomalies. Response data for each node in each fault scenario is collected, recording indicators such as maximum current change, voltage drop, and duration of grid instability. These are then combined with operational standards to form a set of fault scenario constraints, such as critical node voltage not falling below 90% of its rated value and secondary load power outage time not exceeding 180 seconds. All values ​​are set according to State Grid technical specifications.

[0068] The system inputs initial control parameters, predicted load confidence intervals, and the aforementioned fault constraints into a linear programming model. It constructs an objective function to minimize the power outage probability of the first-level load and, assuming that the remaining nodes do not violate fault constraints, solves for the optimized load allocation priority scheme. The scheme specifies the final target power supply value for each load node, the power supply sequence, and the maximum allowable load fluctuation range. This scheme is pushed to each control node in real time for execution.

[0069] After implementing the optimization scheme, the system re-obtains the current power grid status from the monitoring data and recalculates the total power grid overload probability. If the calculation result is still greater than 2%, the system automatically enters an iterative process, adjusting control parameters. For example, the maximum allowable load of level 3 loads is further reduced from 70% to 60%, and linear programming calculations are performed again to generate a new load priority control scheme. This iterative process continues until the recalculated power grid overload probability is lower than 2%. The final determined control parameters are the final load balancing scheme, used to ensure the safe and stable operation of the power grid.

[0070] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An adaptive multiphase parallel power supply management system, characterized in that, The system includes: The feature extraction module extracts periodic, trend, and random features from historical electricity consumption data and real-time sensor data, and uses a time series decomposition algorithm to determine the probability of load surge and obtain the load surge probability value. If the probability value of a load surge is higher than a preset threshold, the load prediction module processes real-time load data through a multilayer perceptron structure to predict the peak load surge and duration within an hourly or minute-level time window, thus obtaining the load surge prediction result. The standby capacity allocation module obtains available capacity data from the standby power database based on the peak value and duration of the load surge prediction results, and optimizes the standby capacity allocation using a linear programming algorithm to obtain a standby capacity allocation scheme. The power grid status monitoring module, based on the standby capacity allocation scheme, acquires power grid operation status data through a real-time monitoring system, detects whether the power grid voltage deviates from the normal range, and determines whether there is a risk of voltage fluctuation. If the voltage stabilization module detects a risk of voltage fluctuation, it adjusts the generator output through a dynamic adjustment algorithm, combines the predicted peak load surge, stabilizes the voltage output, and obtains stable voltage output data. It also includes a resource allocation optimization module, which, based on stable voltage output data, obtains load distribution data of node and edge weights from the distribution network, and uses a graph cut algorithm to optimize the resource allocation path, resulting in an optimized resource allocation path, specifically including: By acquiring node load and edge weight data through the real-time data stream of the power distribution network, the initial resource allocation path is generated using the graph cut algorithm. Based on the initial resource allocation path, extract the load distribution characteristics on the path and determine the load distribution matrix; If any node in the load distribution matrix has a load exceeding a preset threshold, the resource allocation ratio is adjusted using a linear programming algorithm to obtain an optimized resource allocation scheme. Based on the optimized resource allocation scheme, update the edge weight data and generate a new network topology. Using the new network topology, the resource allocation path is recalculated using the graph cut algorithm to obtain the updated optimized path. Based on the updated optimized path, the data acquisition frequency is adjusted to generate a high-frequency monitoring data stream to determine the load balance status of the power distribution network. By monitoring the high-frequency data stream, the relationship between load distribution and stable voltage is analyzed to determine the voltage stability control parameters.

2. The adaptive multiphase parallel power management system according to claim 1, characterized in that, The feature extraction module extracts periodic, trend, and random features from historical electricity consumption data and real-time sensor data, and uses a time series decomposition algorithm to determine the probability of a load surge, obtaining the load surge probability value including: The raw dataset is obtained from historical electricity consumption data and real-time sensor data. Preprocessing techniques are used to remove noise and missing values ​​to obtain a standardized dataset. Based on the standardized dataset, the classic time series decomposition algorithm is used to extract periodic features, trend features, and random features to obtain the decomposed feature set; For the decomposed feature set, if the amplitude of the periodic feature exceeds a preset threshold, it is marked as a high periodic fluctuation, thus obtaining a periodic fluctuation label set. Based on the decomposed feature set and the periodic fluctuation label set, the slope change of the trend feature is calculated. If the absolute value of the slope change is greater than the preset threshold, it is marked as a trend change and a trend change label set is obtained. Randomness features are extracted from the decomposed feature set, and the random fluctuation intensity is calculated using the Gaussian distribution fitting method to obtain the random fluctuation intensity value; Based on the periodic fluctuation label set, the trend change label set, and the random fluctuation intensity value, the logistic regression algorithm is used to calculate the probability of load surge and obtain the load surge probability value. If the probability value of a sudden load increase is greater than a preset threshold, a sudden load risk assessment is triggered, and the risk assessment result is obtained.

3. The adaptive multiphase parallel power management system according to claim 1, characterized in that, If the load surge probability value is higher than a preset threshold, the load prediction module processes real-time load data through a multilayer perceptron structure to predict the peak load surge and duration within an hourly or minute-level time window, obtaining the load surge prediction result including: Raw load data is acquired from real-time sensors, and the data is segmented using a sliding window technique to obtain a segmented load dataset. Based on the segmented load dataset, calculate the load fluctuation characteristics within each time window to obtain the fluctuation characteristic set; If the peak change rate in the fluctuation feature set is higher than a preset threshold, the fluctuation feature set is trained using a multilayer perceptron model to obtain a load surge prediction model. A load surge prediction model is used to process real-time load data and predict the load peak within an hourly time window to obtain hourly peak prediction results. Based on the hourly peak prediction results, the load change trend within the minute-level time window is calculated to obtain the minute-level trend characteristics. If the absolute value of the slope of the minute-level trend feature is greater than the preset threshold, the minute-level trend feature is classified by the random forest algorithm to determine the duration of the load surge and obtain the duration prediction result. Based on the hourly peak prediction results and the duration prediction results, the combined prediction results are used to generate a load surge prediction report, resulting in the final prediction output.

4. The adaptive multiphase parallel power management system according to claim 1, characterized in that, The backup capacity allocation module, based on the peak value and duration of the load surge prediction results, obtains available capacity data from the backup power database, and optimizes the backup capacity allocation using a linear programming algorithm to obtain a backup capacity allocation scheme including: The load peak and duration data are obtained from the load surge prediction results, and the backup power database is queried through the database interface to obtain backup capacity information; Based on the reserve capacity information, a linear programming model is constructed, where the objective function is to maximize allocation efficiency, and the constraints include load peak and duration, thus obtaining the optimized allocation model. The standby capacity information is processed by a linear programming model to calculate the allocation ratio of each power unit and obtain a preliminary capacity allocation scheme. If the allocation ratio of any power unit in the initial capacity allocation scheme exceeds the preset threshold, the allocation ratio is adjusted by a greedy algorithm to prioritize meeting the peak load demand, thus obtaining the adjusted allocation scheme. Based on the adjusted allocation scheme, real-time power status data is obtained to determine whether each power unit meets the capacity requirements of the allocation scheme, and the power availability judgment result is obtained. Based on the power availability assessment results, update the capacity status in the backup power database and generate the final capacity allocation plan. Based on the final capacity allocation scheme, the scheduling power management system executes the allocation operation and determines the allocation execution status.

5. The adaptive multiphase parallel power supply management system according to claim 1, characterized in that: The power grid status monitoring module, based on the reserve capacity allocation scheme, acquires power grid operating status data through a real-time monitoring system, detects whether the power grid voltage deviates from the normal range, and determines whether voltage fluctuation risks exist, including: The system collects power grid operation status data through a real-time monitoring system and generates a voltage data stream. If the voltage data exceeds the preset threshold range, the voltage fluctuation risk is determined by the voltage anomaly detection algorithm, and the risk assessment result is obtained. Based on the risk assessment results, an abnormal alarm mechanism is used to generate alarm signals and determine alarm priorities. By prioritizing alarms, we can obtain grid load level data and determine the grid stability status. If the grid stability is below the preset stability threshold, the voltage data stream and grid load level are analyzed using the support vector machine algorithm to obtain stability optimization parameters. Based on the stability optimization parameters, the data acquisition frequency is adjusted to generate an optimized monitoring data stream; By optimizing the monitoring data stream, the configuration of the real-time monitoring system is updated, and a dynamic adjustment plan for the power grid operation status is determined.

6. The adaptive multiphase parallel power supply management system according to claim 1, characterized in that: If the voltage stabilization module detects a risk of voltage fluctuation, it adjusts the generator output using a dynamic adjustment algorithm, and combines this with the predicted peak load surge to stabilize the voltage output. The stable voltage output data includes: If a voltage fluctuation risk is detected, the peak load surge is analyzed using a support vector machine algorithm, and the generator output is adjusted to obtain stable voltage output data. Based on the stable voltage output data, obtain the real-time monitoring data stream of the power grid operation status, analyze the changing trend of the voltage data, and judge the stability of the power grid operation status; If the stability of the power grid operation is lower than a preset threshold, the real-time monitoring data stream is processed by the random forest algorithm to extract voltage fluctuation characteristics and obtain voltage fluctuation characteristic parameters. Based on voltage fluctuation characteristic parameters, the data acquisition frequency is adjusted to generate an optimized monitoring data stream and determine a dynamic adjustment scheme for the power grid operation status. By analyzing the relationship between the power grid load level and voltage fluctuation characteristics through the optimized monitoring data stream, load allocation optimization parameters are obtained. Based on the load distribution optimization parameters, the output distribution ratio of the generator set is dynamically adjusted to generate an optimized output control scheme. By optimizing the power output control scheme, updating the configuration parameters of the real-time monitoring system, and determining the stable control strategy for the power grid operation status.

7. The adaptive multiphase parallel power supply management system according to claim 1, characterized in that, It also includes an overload probability assessment module, which constructs a node and edge weight model of the distribution network for the optimized resource allocation path, simulates the current distribution under the load allocation path through Monte Carlo analysis, and determines the probability of grid overload by combining the overload threshold setting. Specifically, it includes: Node load and edge weight data are obtained from the real-time data stream of the power distribution network to construct an initial network topology model and obtain the distribution characteristics of nodes and edge weights. The Monte Carlo method was used to perform multiple random simulations of the load distribution path in the initial network topology model to obtain the statistical characteristics of the current distribution. If the current value of a node or edge exceeds the preset overload threshold in the statistical characteristics of the current distribution, the load distribution ratio is adjusted by linear programming algorithm to obtain an optimized load distribution scheme. Based on the optimized load distribution scheme, update the edge weight data in the network topology model to generate a new topology. Using the new topology, the Monte Carlo method is used to re-simulate the current distribution under the load distribution path, and the updated current distribution characteristics are obtained. If the current value of any node or edge in the updated current distribution characteristics exceeds the overload threshold, the load distribution ratio is adjusted iteratively to generate the final load distribution scheme. Based on the final load allocation scheme, calculate the overload probability of the power grid under the current load allocation path and determine the load balance status of the power grid.

8. The adaptive multiphase parallel power supply management system according to claim 1, characterized in that, It also includes a load priority control module. If the probability of grid overload exceeds the power supply continuity requirement, the load allocation priority is adjusted through a feedback control mechanism. Combining the predicted confidence interval and the simulation results of fault scenarios, the load allocation is optimized to obtain the final load balancing scheme, which specifically includes: By acquiring grid load data through real-time data streams, calculating the grid overload probability, and determining whether it exceeds the power supply continuity threshold; If the probability of grid overload is higher than the power supply continuity threshold, a feedback control mechanism is adopted to adjust the load allocation priority and generate an initial control parameter adjustment scheme. Based on the initial control parameter adjustment scheme and combined with the load prediction model, the confidence interval of load distribution is calculated to obtain the predicted load distribution characteristics. By simulating fault scenarios, the stability of load distribution characteristics under different fault conditions is analyzed and predicted, and fault scenario constraints are generated.

9. An adaptive multiphase parallel power supply management system according to claim 8, characterized in that: The load priority control module, if the grid overload probability is higher than the power supply continuity index requirement, adjusts the load allocation priority through a feedback control mechanism. Combining the predicted confidence interval and fault scenario simulation results, it optimizes load allocation to obtain the final load balancing scheme. This also includes: By employing a linear programming algorithm, combined with fault scenario constraints and confidence intervals, the load allocation priority is optimized to obtain an optimized load allocation scheme. By optimizing the load distribution scheme, updating the grid load data, recalculating the grid overload probability, and determining whether the power supply continuity threshold is met; If the recalculated grid overload probability is still higher than the power supply continuity threshold, the control parameters are iteratively adjusted to generate the final load balance state.

Images