Control method of centralized control platform and electrical system

Abstract

The application relates to the technical field of power network power supply control, and particularly discloses a control method of a centralized control center station and an electrical system. In the control process of the centralized control center station, by mining source side, network side, load side and storage side full-dimension cross-category data association, constructing multi-time scale prediction, control logic and closed-loop optimization mechanism, the problem of information island existing in the data processing in the control process of the centralized control station in the prior art is solved, meanwhile, the problems of lack of time scale adaptability of prediction and control, lack of control foresight and precision are solved, stable operation of the power grid, uninterrupted power supply of sensitive loads and efficient utilization of energy storage are realized, and the prediction precision, control adaptability and equipment operation reliability and utilization efficiency are improved.

Description

Technical Field

[0001] This application relates to the field of power grid power supply control technology, and specifically discloses a control method and electrical system for a centralized control platform. Background Technology

[0002] With the large-scale development of sensitive loads such as data centers and the widespread grid connection of distributed energy sources (photovoltaics and wind power), the power system is characterized by "intensified source-load fluctuations, complex multi-device coordination, and high requirements for fault response." The centralized control console connected to the grid is the central control unit connecting distributed energy sources, power generation equipment, and the main power grid. Its core function is to achieve safe grid connection and orderly disconnection of energy sources from the grid through data acquisition, status monitoring, decision control, and coordinated dispatching, while ensuring stable grid operation and efficient energy utilization.

[0003] The existing centralized control platform control methods mainly suffer from the following technical defects:

[0004] For example, patent application CN117200288A discloses a data center energy storage UPS control method based on dynamic load. This method activates the characteristics of the lithium battery system and improves the charging and discharging performance of the lithium battery energy storage system by utilizing surplus capacity to participate in peak shaving and valley filling of the power grid, overcoming the shortcomings of traditional UPS systems that require long-term floating charging. However, in the control process, it only adopts a single time-scale prediction plus unified control mode, which is difficult to meet the planning requirements of long-term energy storage peak-valley scheduling and cannot cope with the rapid response requirements of short-term sensitive load pulse fluctuations. This leads to inefficient long-term energy storage utilization and short-term control lag, resulting in problems such as insufficient peak-segment energy storage and short-term load surges impacting the power grid.

[0005] Meanwhile, existing centralized control consoles are in a state of information silos when processing the collected data: existing methods only perform independent classification processing on the diverse and heterogeneous data from the source side, network side, load side, and storage side, without mining the implicit correlations between cross-category data, resulting in the prediction model having a single input feature.

[0006] In summary, existing control methods applied to centralized control platforms for grid connection have not achieved deep integration of cross-category data correlation and multi-timescale prediction-control, making it difficult to meet the control requirements of integrated source-grid-load-storage systems in smart grids.

[0007] This invention provides a control method and electrical system for a centralized control platform to solve the above-mentioned problems. Summary of the Invention

[0008] The purpose of this invention is to provide a control method and electrical system for a centralized control platform. By mining cross-category data correlation across all dimensions of the source side, grid side, load side, and storage side, constructing multi-timescale prediction, control logic, and closed-loop optimization mechanism, it can achieve stable grid operation, uninterrupted power supply to sensitive loads, and efficient utilization of energy storage.

[0009] To achieve the above objectives, the basic solution of the present invention provides a control method for a centralized control platform, comprising the following steps:

[0010] Step S1: Collect multi-source heterogeneous data including source side, network side, load side, storage side, environment, and scenario-related data, and obtain a standardized dataset after preprocessing;

[0011] Step S2: Split the standardized dataset into subsets according to each category, calculate the linear association strength of key data pairs across categories, establish a paired linear quantization model, and extract the cross-category association feature set;

[0012] Step S3: Integrate the corresponding cross-category association features into various data subsets to obtain enhanced subsets, and extract time-series features in parallel according to the three time scales of long, medium and short time scales to obtain a three-level time scale feature set;

[0013] Step S4: Input the three-level time scale feature set into the pre-established prediction model adapted to different time scales, and perform parallel prediction to output the full-scene prediction results and trigger thresholds;

[0014] Step S5: Identify four core scenarios based on scenario association data and match them with preset control strategies. Based on the full scenario prediction results and trigger thresholds, perform hierarchical control according to long-term peak-valley energy storage scheduling, medium-term power grid stability control and fault prevention, and short-term pulse load replenishment.

[0015] Step S6: Monitor the control results in real time using multi-dimensional indicators, continuously optimize the prediction model, correlation features, and fault prevention and control strategies based on the monitoring results, and feed them back to steps S2 to S5 respectively.

[0016] Furthermore, in step S2, the linear correlation strength is calculated based on the Pearson correlation coefficient. Based on the total number of data, the deviations of the normalized values ​​of the two types of data from their respective means are compared, and the correlation coefficient is obtained by the ratio of covariance to standard deviation.

[0017] Furthermore, the cross-category key data pairs include key data pairs between source-side data and load-side data, between source-side data and grid-side data, between load-side data and storage-side data, and between grid-side data and storage-side data.

[0018] Furthermore, in step S3, various data subsets and paired cross-category association features are horizontally concatenated to obtain enhanced subsets.

[0019] Furthermore, the three-level time-scale feature set includes a long-term feature set, a medium-term feature set, and a short-term feature set;

[0020] The long-term scale feature set includes source-side power output trend features, load-side power accumulation features, and grid-side peak-valley electricity price cycle features.

[0021] The mid-timescale feature set includes grid-side voltage stability features, source-side power output fluctuation features, and cross-category correlation sensitivity features.

[0022] The short-time scale feature set includes load-side pulse load features and storage-side supercapacitor features.

[0023] Furthermore, in step S4, the full-scenario prediction results and trigger thresholds include long-term energy storage charging and discharging thresholds, medium-term voltage deviation and fault probability, short-term power mutation prediction values, and energy replenishment thresholds.

[0024] Furthermore, in step S4, for the three time scales of long time, medium time, and short time, LSTM neural network models, GRU neural network models, and lightweight CNN models combined with attention mechanisms are established respectively.

[0025] Furthermore, in step S5, a reverse correction mechanism is established to achieve cross-scale regulation coordination, transitioning from short-term to medium-term and then to long-term.

[0026] The basic solution of the present invention also provides an electrical system for a centralized control platform, used to implement the control method of the centralized control platform according to the above description, including:

[0027] The data acquisition and preprocessing module is used to collect multi-source heterogeneous data, including source side, network side, load side, storage side, environment, and scenario-related data, and obtain a standardized dataset after preprocessing.

[0028] The cross-category association feature set extraction module is used to split the standardized dataset into data subsets according to each category, calculate the linear association strength of key cross-category data pairs, establish a paired linear quantization model, and extract the cross-category association feature set.

[0029] The three-level time scale feature set extraction module is used to integrate corresponding cross-category correlation features into various data subsets to obtain enhanced subsets, and extract time-series features in parallel according to long, medium and short time scales to obtain three-level time scale feature sets;

[0030] The full-scene prediction module is used to input the three-level time scale feature set into a pre-built prediction model adapted to different time scales, perform parallel prediction, and output the full-scene prediction result and trigger threshold.

[0031] The hierarchical control module is used to identify four core scenarios based on scenario-related data and match preset control strategies. Based on the full-scenario prediction results and trigger thresholds, hierarchical control is performed for long-term peak-valley energy storage scheduling, medium-term power grid stability control and fault prevention, and short-term pulse load replenishment.

[0032] The optimization module is used to monitor the control results in real time with multi-dimensional indicators. Based on the monitoring results, it continuously optimizes the prediction model, correlation features, and fault prevention and control strategies, and feeds them back to the cross-category correlation feature set extraction module, the three-level time scale feature set extraction module, the full-scenario prediction module, and the hierarchical control module, respectively.

[0033] The principle and effect of this solution are as follows:

[0034] Compared with existing technologies, this invention solves the problem of information silos in the centralized control console during data processing by mining cross-category data correlations across the source, grid, load, and storage sides, constructing multi-timescale prediction, control logic, and closed-loop optimization mechanisms. It also addresses the lack of time-scale adaptability in prediction and control, as well as insufficient foresight and accuracy in control. This achieves stable grid operation, uninterrupted power supply to sensitive loads, and efficient utilization of energy storage, improving prediction accuracy, control adaptability, equipment reliability, and utilization efficiency. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 A flowchart of a control method for a centralized control platform proposed in an embodiment of this application is shown. Detailed Implementation

[0037] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0038] A control method for a centralized control platform, implementing, for example... Figure 1 As shown, it includes the following steps:

[0039] Step S1: Collect multi-source heterogeneous data including source side, network side, load side, storage side, environment, and scenario-related data, and obtain a standardized dataset after preprocessing.

[0040] Multi-source heterogeneous data is collected based on pre-deployed sensors and monitoring modules.

[0041] Specifically, in this embodiment, a high-frequency acquisition sensor with a sampling frequency ≥20Hz and a fixed sampling interval ≤50ms is deployed to collect dynamic data from the source-side data and grid-side data. A status monitoring module is also deployed to collect device status data such as the SOC and supercapacitor terminal voltage of the energy storage BESS. The fixed sampling interval of the status monitoring module is consistent with that of the high-frequency acquisition sensor. A scene signal receiving module with a reception delay ≤100ms is also deployed to receive scene-related data such as grid dispatch instructions and computing power leasing orders in real time.

[0042] The collected source-side data includes photovoltaic irradiance (W / m²). 2 The data includes wind speed (m / s) and real-time output data (kW) of distributed energy sources. Distributed energy sources include photovoltaic energy and wind power.

[0043] The collected grid-side data includes grid voltage (kV), current (A), frequency (Hz), grid fault signals, real-time peak-valley electricity price signals (yuan / kWh), grid dispatch instructions, grid maintenance notices, and grid fault levels (levels 1-4). Among them, grid fault signals include undervoltage and overcurrent, and grid dispatch instructions include full grid connection and power rationing.

[0044] The collected load-side data includes data center sensitive load power (kW, including 1Hz pulse load), load priority basic tags, computing power task priority (P0-P3 level), and service interruption loss coefficient (yuan / minute). Among them, the load priority basic tags include core servers, network equipment, office load, and air conditioning auxiliary units.

[0045] The collected energy storage side data includes the SOC (%), single cell voltage (V), charge / discharge rate (C), supercapacitor SOC (%), and terminal voltage (V) of the BESS.

[0046] The collected environmental data includes data center room temperature (°C) and photovoltaic panel surface temperature (°C). The collected scenario-related data includes computing power rental order signals (including order time period and computing power demand kW).

[0047] After collecting the above multi-source heterogeneous data, it is preprocessed. The preprocessing process includes numerical unification, identification and removal of outliers, normalization and time alignment to obtain a standardized dataset.

[0048] Specifically, the numerical standardization process, including standardizing the numerical ranges of various types of data in multi-source heterogeneous data, is carried out using the following formula:

[0049] ;

[0050] In the formula, These are intermediate data after numerical standardization. This is the historical statistical average of this type of data over the past 72 hours of normal operation. For data that needs to be standardized, This is the minimum value among the physically reasonable values ​​for this type of data. This represents the maximum value within the physically reasonable range of values ​​for this type of data.

[0051] For the unified intermediate data, calculate the difference between it and the average of the normal operating data of the same category in the past hour. Use the absolute value of this difference to identify abnormal data under the 3σ criterion. Intermediate data with a difference that meets the 3σ criterion is considered normal data, while intermediate data with a difference that does not meet the 3σ criterion is considered abnormal data. The case where the difference does not meet the 3σ criterion is that the absolute value of the difference exceeds the range of the 3σ criterion.

[0052] Next, normal data of different dimensions, such as voltage, power, and temperature, are normalized to the [0,1] interval using the following formula:

[0053] ;

[0054] In the formula, For normalized data, This is normal data. This is the historical minimum value for this type of data. This represents the historical maximum value for this type of data.

[0055] Furthermore, using 50ms as a unified time base, time alignment is performed on asynchronously sampled data, i.e., data with different sampling intervals, according to the following process:

[0056] For data with a sampling interval of less than 50ms, the mean aggregation method is used to compress the data to a 50ms interval. For data with a sampling interval of more than 50ms, the linear interpolation method is used to complete the data to a 50ms interval.

[0057] Step S2: Split the standardized dataset into subsets according to each category, calculate the linear association strength of key data pairs across categories, establish a paired linear quantization model, and extract the cross-category association feature set. The established linear quantization models include source-load side, source-grid side, and load-storage side linear quantization models.

[0058] After obtaining the standardized dataset, the standardized dataset is divided into four data subsets according to the four categories of source side, network side, load side, and storage side.

[0059] The source-side data subset includes normalized data on photovoltaic irradiance, wind speed, and photovoltaic or wind power output. The grid-side data subset includes normalized data on grid voltage, current, frequency, fault signals, electricity prices, dispatch instructions, and fault levels. The load-side data subset includes normalized data on load power, computing task priority, service interruption loss coefficient, and load base tag. The storage-side data subset includes normalized data on the SOC, single-cell voltage, and charge / discharge rate of BESS (Body Storage System Escalator), and the SOC and terminal voltage of supercapacitors.

[0060] In this embodiment, the linear correlation strength of key data pairs between different data subsets is calculated and quantified using the Pearson correlation coefficient. This coefficient ranges from -1 to 1, with a larger absolute value indicating a stronger linear correlation. Specifically, the calculation is based on the total number of time-aligned data points. The deviations of the normalized values ​​of the two data types from their respective means are compared, and the correlation coefficient is obtained by the ratio of the covariance to the standard deviation. The calculation formula is as follows:

[0061] ;

[0062] In the formula, Let Pearson correlation coefficient be the correlation coefficient between data A and data B. Let A be the normalized value of data A at time point k. Let B be the normalized value at time point k. Let A be the normalized mean of data A. Let M be the normalized mean of data B, and M be the total number of data points.

[0063] Furthermore, after calculating the linear correlation strength of each key data pair, a paired linear quantization model is established according to the following process:

[0064] To correlate source-side and load-side data, a linear quantification model of photovoltaic irradiance and air conditioning load is established, which addresses the Pearson correlation coefficient between photovoltaic irradiance and air conditioning load. In this case, for every 100W / m increase in irradiance 2 The physical law governing an 8-12kW increase in air conditioning load is calculated using the following formula:

[0065] ;

[0066] In the formula, For air conditioning load, The correlation coefficient, , Photovoltaic irradiance, This is a constant term.

[0067] For the correlation between source-side data and grid-side data, the sensitivity coefficient between wind power output and grid voltage is calculated using the following formula. :

[0068] ;

[0069] In the formula, To provide power for wind power, This represents the change in wind power output (kW). This is the grid voltage. The sensitivity coefficient ranges from -0.006 to -0.01 kV / kW, corresponding to a decrease in output power and a decrease in voltage.

[0070] Regarding the correlation between load-side and storage-side data, the linear relationship between core load power and energy storage SOC consumption rate is fitted using the following formula, which corresponds to a 0.5% / min increase in SOC consumption rate for every 20kW increase in load. The calculation formula is as follows:

[0071] ;

[0072] In the formula, The energy storage SOC consumption rate, This is the consumption rate coefficient. , For core load power, Based on the consumption rate.

[0073] The correlation between grid-side and energy storage-side data includes the correlation between grid voltage and BESS charge / discharge rate, grid frequency and BESS SOC, and peak / valley electricity price and energy storage charge / discharge capacity. Specifically, a linear quantification model for the correlation between grid voltage and BESS charge / discharge rate is constructed using the following formula, addressing cases where the absolute value of the Pearson correlation coefficient between rated voltage and charge / discharge rate is ≥0.7:

[0074] ;

[0075] ;

[0076] In the formula, The relationship between grid voltage and BESS charging rate. The relationship between grid voltage and BESS discharge rate is shown. Rated voltage, This is the grid voltage. , .

[0077] A linear quantitative model relating grid frequency to the State of Charge (SOC) of a BESS (Based Energy Storage System) is constructed using the following formula, to address the case where the Pearson correlation coefficient between frequency and SOC is ≤ -0.65:

[0078] ;

[0079] In the formula, For the power grid frequency, For the rated frequency, , , .

[0080] A linear quantitative model relating peak-valley electricity prices to energy storage charging and discharging is constructed using the following formula, addressing cases where the Pearson correlation coefficient between electricity price and discharge is ≤-0.8 and ≥0.8:

[0081] ;

[0082] ;

[0083] In the formula, For energy storage charging capacity, For energy storage discharge capacity, For peak-valley electricity pricing, This represents the maximum peak-valley electricity price. , .

[0084] After obtaining the above linear correlations, they are converted into quantitative features to form a cross-category correlation feature set. The cross-category correlation feature set includes Pearson correlation coefficients, sensitivity values, and linear fitting curve parameters, where the linear fitting curve parameters include the coefficients in the linear model.

[0085] Step S3: Integrate the corresponding cross-category association features into various data subsets to obtain enhanced subsets, and extract time-series features in parallel according to the three time scales of long, medium and short time to obtain a three-level time scale feature set.

[0086] Based on the cross-category association feature set obtained in step S2, it is split into categories and horizontally concatenated with the corresponding data subsets to obtain enhanced subsets of joint cross-category association features. Specifically, the horizontal concatenation process is as follows: the source-side data subset is integrated with source-load-side and source-side-network association features to obtain a source-side enhanced subset; the network-side data subset is integrated with source-network-side and network-storage-side association features to obtain a network-side enhanced subset; the load-side data subset is integrated with source-load-side and load-side-storage-side association features to obtain a load-side enhanced subset; and the storage-side data subset is integrated with load-side-storage-side and network-storage-side association features to obtain a storage-side enhanced subset.

[0087] In this embodiment, a differentiated sliding window and feature extraction logic are set up to meet the prediction requirements of long, medium and short time scales. For the enhanced subsets of source side, network side, load side and storage side, the appropriate time-series features are extracted in parallel at the long, medium and short time scales respectively to obtain the long time scale feature set, the medium time scale feature set and the short time scale feature set.

[0088] The long-term feature extraction spans 1-3 hours, with a fixed time step of 5 minutes and a sliding window length of 12, covering 1 hour of data. This yields a long-term feature set including source-side power output trend features, load-side cumulative power features, and grid-side peak-valley electricity price cycle features. Specifically, the source-side power output trend features include photovoltaic power output trends and wind power output trends; the load-side cumulative power features include core load accumulation and non-core load accumulation; and the grid-side peak-valley electricity price cycle features include peak-period electricity price cycles and valley-period electricity price cycles.

[0089] The long-term scale feature extraction targets the source-side enhanced subset, the load-side enhanced subset, and the network-side enhanced subset. The extraction process is as follows:

[0090] Source-side power output trend characteristics: Calculated using the moving average method, the formula is as follows:

[0091] ;

[0092] In the formula, The source-side power output trend characteristics at the k-th long-term time step include one of the photovoltaic power output trend and the wind power output trend. For the i-th original data, the normalized value of the photovoltaic / wind power output is... The length of the long-term sliding window. This is the long-term time step number.

[0093] Load-side power accumulation characteristics: obtained by accumulating the load power at each time step, calculated as follows:

[0094] ;

[0095] In the formula, The cumulative load-side power characteristic at the k-th long-time step. The original data sampling interval, , For long-term fixed time steps, The number of raw data points contained in each time step.

[0096] The cyclical characteristics of grid-side peak-valley electricity prices are obtained by extracting the periodic factor from the electricity price time series, and the calculation formula is as follows:

[0097] ;

[0098] In the formula, For the first The characteristics of electricity price cycles over a long period of time. This represents the normalized mean of electricity prices over a long timescale. The normalized standard deviation of electricity prices over a long timescale. For the first The normalized value of peak and valley stores over a long period of time.

[0099] Mesoscale feature extraction spans 5-15 minutes, with a fixed time step of 1 minute and a sliding window length of 10, covering 10 minutes of data. This yields a mesoscale feature set including grid-side voltage stability features, source-side power output fluctuation features, and cross-category correlation sensitivity features. Specifically, grid-side voltage stability features include voltage standard deviation and voltage coefficient of variation; source-side power output fluctuation features include photovoltaic power output fluctuation and wind power output fluctuation; and cross-category correlation sensitivity features include source-grid sensitivity time series, grid-load sensitivity time series, source-storage sensitivity time series, and load-storage sensitivity time series.

[0100] Specifically, the extraction targets for mesoscale feature extraction are the source-side enhanced subset, the network-side enhanced subset, and the load-side enhanced subset. The extraction process is as follows:

[0101] Grid-side voltage stability characteristics include the calculation of the voltage standard deviation over standard fluctuation intensity and the calculation of the voltage anomaly coefficient, which characterizes the stability index. The calculation formulas are as follows:

[0102] ;

[0103] In the formula, Let be the standard deviation of the voltage at the k-th time step. The length of the sliding window when in the middle. The normalized value of the grid voltage for the i-th original data point. This represents the normalized mean voltage within the k-th sliding window.

[0104] ;

[0105] In the formula, This is the voltage variation coefficient; the smaller the value, the more stable the voltage.

[0106] Source-side power output fluctuation characteristics: obtained by calculating the mean of the absolute values ​​of the first-order differences of the power output sequence, as shown in the following formula:

[0107] ;

[0108] In the formula, The source-side output power fluctuation characteristics at the k-th intermediate time step are shown. This is the valid data for the first-order difference.

[0109] Cross-category association sensitivity features: obtained by extracting the time-series changes of sensitivity coefficients, calculated as follows:

[0110] ;

[0111] In the formula, For the cross-class sensitivity feature at the k-th mid-time step, Let be the source-to-network sensitivity coefficient at the k-th time step. The mean of the sensitivity coefficient within the medium timescale. denoted as the standard deviation of the sensitivity coefficient within the medium timescale.

[0112] Short-timescale feature extraction spans from 100ms to 1 second, with a fixed time step of 100ms and a sliding window length of 5, covering 500ms of data. This yields a short-timescale feature set including load-side pulsed load characteristics and storage-side supercapacitor characteristics. The load-side pulsed load characteristics include load peak value and pulse frequency, while the storage-side supercapacitor characteristics include SOC change rate and terminal voltage change rate.

[0113] Specifically, the short-term scale feature extraction targets the load-side enhanced subset and the reservoir-side enhanced subset, and the extraction process is as follows:

[0114] Load-side pulse load characteristics: This includes the calculation of load peak characteristics and pulse frequency characteristics. Specifically, the formula for calculating load peak characteristics is as follows:

[0115] ;

[0116] In the formula, The load peak characteristics are for the k-th short-time step. This represents the normalized load power value of the i-th original data. is the length of the sliding window, and k is the short-time step number.

[0117] ;

[0118] In the formula, Let be the pulse load frequency at the k-th short-time step. This is an indicator function (it takes the value 1 if the condition is met, and 0 otherwise). The threshold for determining pulse load. .

[0119] Supercapacitor SOC change rate characteristics: obtained by calculating the instantaneous change rate of the SOC sequence, as shown in the following formula:

[0120] ;

[0121] In the formula, Let SOC be the rate of change of the supercapacitor at the k-th short-time step. This is the normalized SOC value of the supercapacitor at the k-th short-time step. This is the normalized SOC value of the supercapacitor at the (k-1)th short-time step. It is a short fixed time step.

[0122] Step S4: Input the three-level time scale feature set into the pre-established prediction model adapted to different time scales, and perform parallel prediction to output the full-scene prediction results and trigger thresholds.

[0123] In this embodiment, LSTM neural network models, GRU neural network models, and lightweight CNN models with attention mechanisms are established for three time scales: long-term, medium-term, and short-term. Through parallel prediction, the output includes full-scenario prediction results and trigger thresholds, including long-term energy storage charging and discharging thresholds, medium-term voltage deviation and fault probability, short-term power surge prediction values, and energy replenishment thresholds.

[0124] Specifically, long-term forecasting targets long-term planning needs of 1-3 hours. It uses a pre-established LSTM neural network model to predict the trend of source-side output and load-side demand, and outputs energy storage charging and discharging thresholds.

[0125] In the long-term prediction process, the concatenation result of the long-term feature set and the source-load correlation features is used as the input feature of the LSTM neural network model. The long-term feature model is achieved by using the forget gate, input gate, cell state and output gate built into the LSTM neural network model. Specifically, the operation process of the gating unit is as follows:

[0126] Forgotten Gate: ;

[0127] Input Gate: ;

[0128] ;

[0129] Cell status update: ;

[0130] Output gate: ;

[0131] ;

[0132] In the formula, Let be the input features at time t. For the previous moment: Hidden layer output at any given time, For the previous moment: Cellular state at any given moment , , , These are the weight matrices, , , , These are the bias terms, It is the sigmoid activation function. This is element-wise multiplication.

[0133] Finally, the total load demand forecast for the next 1-3 hours is obtained through mapping using the fully connected layer. Forecast value of total output of distributed energy resources ,in .

[0134] Then, combining peak and off-peak electricity price signals, the energy storage charging and discharging thresholds are calculated using the following formula, including the charging threshold. Discharge threshold The calculation is as follows:

[0135] ;

[0136] ;

[0137] In the formula, This is a peak-valley electricity price signal. This represents the minimum value of the peak-valley electricity price. These represent the minimum and maximum values ​​of peak-valley electricity prices.

[0138] Based on the calculated energy storage charging and discharging thresholds, under the condition of dynamically adapting to changes in electricity prices, the peak hours (00:00-06:00) in a normal scenario are as follows. Peak hours (10:00-18:00) .

[0139] Specifically, the medium-time scale prediction targets the 5-15 minute power grid stability forecast. It uses a pre-established GRU neural network model to predict voltage deviation and fault probability, and outputs the UPS mode switching threshold.

[0140] In the mesoscale prediction process, the concatenation result of the mesoscale feature set and the source-network correlation features is used as the input feature of the GRU neural network model. The input feature is processed by the reset gate, update gate, and candidate hidden layer to hidden layer update built into the GRU neural network model. Specifically, the operation process of the gating unit is as follows:

[0141] Reset Door: ;

[0142] Update Gate: ;

[0143] Candidate hidden layers: ;

[0144] Hidden layer update: ;

[0145] In the formula, For the input features at time step t, , , These are the weight matrices, , , These are the bias terms, and the meanings of the other symbols are the same as those for LSTM.

[0146] Finally, the predicted value of the grid voltage deviation for the next 5-15 minutes is obtained. Grid failure probability ,in, .

[0147] Furthermore, a UPS mode switching threshold is set based on a voltage deviation threshold and a fault probability threshold. In this embodiment, the UPS mode switching threshold is specifically set as follows: voltage deviation threshold... Fault probability threshold That is, when or When this occurs, a UPS mode switching command is triggered (which commands are included).

[0148] Specifically, short-time scale prediction addresses the high-frequency response requirements of 100ms-1 second by combining a lightweight CNN model with an attention mechanism to achieve rapid prediction of load power mutations.

[0149] In the short-term prediction process, the concatenation result of the short-term feature set and the load-storage correlation feature is used as the input feature. By combining the lightweight structure of 2 convolutional layers + 1 pooling layer built into the lightweight CNN model with attention mechanism, the following calculations are performed:

[0150] Convolutional layer 1: ;

[0151] Pooling layer: ;

[0152] Convolutional layer 2: ;

[0153] In the formula, , These are the convolution kernels (3×1 and 2×1 respectively). , These are the step sizes (each is 1). Pooling core (size 2×1). Let be the pooling step size (1). , These are the bias terms, This is the activation function.

[0154] During processing, a channel attention mechanism is introduced to strengthen the weights of key features, calculated as follows:

[0155] ;

[0156] ;

[0157] In the formula, For global average pooling, For max pooling, It is a fully connected layer. Here is the channel attention weight vector. This is the attention-weighted feature map.

[0158] Finally, the predicted load power surge values ​​are obtained through a flat layer and a single-output fully connected layer. , .

[0159] Furthermore, the delay control formula is calculated during the short-term forecasting process:

[0160] ;

[0161] In the formula, For convolution computation time, Calculate the time for the attention mechanism. The computation time for the fully connected layer is calculated to ensure that short-time response requirements are met.

[0162] In this embodiment, the following supercapacitor power replenishment trigger threshold is set. :

[0163] ;

[0164] In the formula, This represents the current load power. For load priorities (0-1) based on load priority base labels, core load The time threshold is lowered to ensure a fast response.

[0165] Step S5: Identify four core scenarios based on scenario-related data and match them with preset control strategies. Based on the full-scenario prediction results and trigger thresholds, perform tiered control according to long-term peak-valley energy storage scheduling, medium-term grid stability control and fault prevention, and short-term pulse load replenishment. Furthermore, in this step, a reverse correction mechanism is established to transition from short-term to medium-term and then to long-term control to achieve cross-scale control coordination.

[0166] Specifically, based on the scene-related data collected in step S1, four core scene types are automatically identified through quantitative judgment logic, and matched with a preset control strategy framework, providing a basis for scene adaptation for tiered control. In the process of identifying core scenes based on scene-related data, scene identification coefficients are defined. Based on data such as power grid dispatch instructions, computing power rental orders, maintenance notices, and extreme weather warnings, the scenario type is determined by weighted summation using the following formula:

[0167] ;

[0168] In the formula, This is an indicator function that returns 1 if the scene trigger condition is met, and 0 otherwise. These represent the weights for each scenario. In this embodiment, , , , .

[0169] In the formula, This refers to scenario one, where distributed energy resources are fully connected to the grid. When... ,the remaining At that time, it indicates that the scheduling instruction is to access the internet at full capacity. .

[0170] In the formula, This refers to scenario two, where data center computing power is rented out during off-peak hours. ,the remaining This indicates that there are computing power orders during off-peak periods. .

[0171] In the formula, This is a scenario involving power rationing due to grid maintenance, corresponding to scenario three. When... ,the remaining At that time, it indicates that a power grid maintenance notice has been received. .

[0172] In the formula, This is an extreme weather scenario, corresponding to scenario four. When... ,the remaining At that time, it indicates that a typhoon / rainstorm warning has been received. .

[0173] Furthermore, in this embodiment, control targets and core parameter ranges are pre-established for different scenarios.

[0174] For scenario one: the core control objective is to prioritize ensuring the electricity supply to the grid from distributed energy sources, with key parameters including a BESS charge / discharge rate ≤ 0.3C and no discharge during peak hours.

[0175] For scenario two: the core control objective is to balance the power supply to the computing load and the cost of energy storage. The key parameter range is the charging rate of the BESS during off-peak hours of 0.4-0.5C and the discharging rate of 0.3-0.4C.

[0176] For scenario 3: The core control objective is to ensure uninterrupted power supply to the core load, and the key parameter range is that the SOC of the energy storage BESS is ≥90% and the non-core load shedding threshold is ≥0.6.

[0177] For scenario four: the core control objective is to address the risk of a sudden drop in power output on the source side, with key parameters ranging from supercapacitor SOC ≥ 50% to BESS backup capacity ≥ 30%.

[0178] After scene recognition is completed, three levels of time-scale modulation are performed: long-term scale modulation, medium-term scale modulation, and short-term scale modulation, as follows:

[0179] For long-term planning needs of 1-3 hours, the focus is on optimizing peak-valley scheduling of energy storage BESS. Based on long-term forecast results and combined with matching scenario strategies, the charge and discharge rates and thresholds of the energy storage BESS are dynamically adjusted.

[0180] Specifically, the charge / discharge rate is adaptively adjusted based on the predicted source-load difference and scene type, and the calculation formula is as follows:

[0181] ;

[0182] ;

[0183] In the formula, This refers to the charging rate. In Scenario 3, the energy storage BESS maintains a SOC ≥ 90% and discharges at a rate of 0.2C only when there is a power supply gap in the core load.

[0184] Furthermore, to prevent the real-time SOC from deviating from the target range, a closed-loop calibration of the charging and discharging thresholds is performed based on the real-time state of charge (BESS) of the energy storage system. This ensures that the energy storage operation always meets the needs of the scenario, avoiding both insufficient energy storage during peak periods and wasted charging during off-peak periods. In this embodiment, the energy storage charging threshold is combined with... Discharge threshold Based on real-time SOC value dynamic calibration, the calculation formula is as follows:

[0185] ;

[0186] ;

[0187] In the formula, For real-time state of charge of BESS energy storage, Set the target SOC for the scenario (e.g., 90% for scenario 3). For the calibrated charging threshold, This is the calibrated discharge threshold.

[0188] For medium-time scale regulation of 5-15 minutes, based on the medium-time prediction results and the source-grid correlation characteristics, grid voltage stability regulation and active fault prevention are implemented to avoid grid fluctuations and fault propagation at the medium-time scale.

[0189] The power grid voltage stabilization control process is as follows: when the predicted voltage deviation... When the UPS switches to voltage regulation mode, it dynamically adjusts the output voltage to ensure that the voltage deviation after adjustment is ≤±1%. The adjustment formula is as follows:

[0190] ;

[0191] In the formula, The rated voltage of the power grid. This is the real-time voltage of the power grid. This is the voltage regulation coefficient.

[0192] The proactive fault prevention and control process is as follows: First, based on the source-network correlation characteristics, the fault risk level is defined. The calculation formula is as follows:

[0193] ;

[0194] In the formula, The slope of wind power output fluctuation. This is the predicted value of voltage deviation. The threshold for the fluctuation slope. .

[0195] Furthermore, tiered prevention and control will be implemented through the following process:

[0196] when The risk level is determined to be low. The UPS will remain in grid connection mode, and the source-grid parameters will be continuously monitored.

[0197] when The risk level was determined to be medium, and the UPS was instructed to enter voltage stabilization mode, with the BESS energy storage pre-charged to [value missing]. The expression is: .

[0198] when Determined to be high-risk, core loads were locked in advance, and BESS (Based Energy Storage System) backup capacity was utilized. This ensures rapid power replenishment in case of failure.

[0199] In this embodiment, the UPS is equipped with four operating modes: double-conversion online mode, voltage regulation mode, emergency power supply mode, and bypass mode. Double-conversion online mode provides daily power supply and BESS (Battery Safe Energy Storage) replenishment, laying the foundation for medium-time scale regulation. Voltage regulation mode enables medium-time scale grid stability control. Emergency power supply mode ensures uninterrupted power supply to sensitive loads under fault conditions. Bypass mode provides equipment protection.

[0200] Specifically, when a sudden drop in wind / solar power output is predicted, the compensation amount for the sudden drop in output is calculated using the following formula. And compensation will be made based on this:

[0201] ;

[0202] In the formula, This is the predicted source-side power output under extreme weather conditions, and the energy storage BESS uses... Replenish energy in advance according to the multiplier.

[0203] For short-term control of 100ms-1s, based on short-term prediction results, the supercapacitor is triggered to quickly replenish energy to cope with 1Hz-level pulse loads and ensure short-term power supply stability.

[0204] Specifically, supercapacitor power replenishment Based on the predicted power surge and the dynamic adjustment of the current load power, the calculation formula is as follows:

[0205] ;

[0206] In the formula, This is the energy redundancy coefficient. The current load power is used to ensure that the replenishment covers the peak and fluctuation margin of the pulse load.

[0207] Furthermore, a bidirectional DC / DC converter is used to achieve rapid power exchange between the supercapacitor and the load side. The expression for the control strategy is as follows:

[0208] ;

[0209] In the formula, The converter duty cycle, This is the terminal voltage of the supercapacitor. The target voltage for the load. This is the control factor.

[0210] When the pulse load eases to Alternatively, if the supercapacitor's SOC is less than 20%, stop replenishing energy to prevent over-discharge of the supercapacitor.

[0211] Furthermore, in this embodiment, a reverse correction mechanism is also established to transition from short-term to medium-term and then to long-term, which corrects the previous-term control instructions in real time based on the execution results of the next-level control, ensuring the coordination and consistency of the three-level scale control.

[0212] Specifically, when the supercapacitor's state of charge (SOC) is replenished on a short timescale, At that time, the BESS (Best Energy Storage System) provides energy replenishment to the supercapacitor, with a replenishment ratio of [missing value]. The calculation formula is as follows:

[0213] ;

[0214] The target is to replenish the supercapacitor to 50% SOC, and the replenishment rate is dynamically adjusted according to the current SOC to avoid over-replenishing the supercapacitor.

[0215] The actual output value of the source side was detected on the time scale. Compared with long-term forecast values deviation When correcting the long-term energy storage charging and discharging rates, the peak (or valley) energy storage charging and discharging rates are adjusted downwards (or upwards) as follows:

[0216] ;

[0217] ;

[0218] In the formula, For long-term correction time windows, , For energy storage charging and discharging efficiency, , Rated capacity for BESS energy storage.

[0219] Step S6: Monitor the control results in real time using multi-dimensional indicators, continuously optimize the prediction model, correlation features, and fault prevention and control strategies based on the monitoring results, and feed them back to steps S2 to S5 respectively.

[0220] Specifically, the multi-dimensional indicators include grid indicators, energy storage indicators, load indicators, and scenario indicators.

[0221] Among these, power grid indicators include voltage fluctuation and frequency stability, which directly reflect the quality of power supply. Voltage fluctuation The voltage difference between the maximum and minimum values ​​during the control period is expressed as a percentage of the rated voltage, calculated as follows:

[0222] ;

[0223] In the formula, The maximum voltage of the power grid during the control period. For minimum voltage, The rated voltage of the power grid. .

[0224] Frequency stability is measured by the standard deviation of frequency deviation. The calculation formula is as follows:

[0225] ;

[0226] In the formula, Let k be the power grid frequency at the kth sampling point. For the rated frequency, , The number of sampling points during the monitoring period. .

[0227] Among these, energy storage indicators include the SOC fluctuation range and charge / discharge efficiency of BESS and supercapacitors to avoid overcharging, over-discharging, and inefficient operation. The SOC fluctuation range of BESS and supercapacitors is also a key factor. The calculation formula is as follows:

[0228] ;

[0229] ;

[0230] In the formula, The maximum SOC of BESS energy storage For the minimum SOC of BESS energy storage, This represents the maximum SOC of the supercapacitor. This represents the minimum state of charge (SOC) of a supercapacitor. , .

[0231] Energy storage system charge and discharge efficiency The ratio of discharge to charge is used, and the calculation formula is as follows:

[0232] ;

[0233] In the formula, Total charging amount This is the total discharge amount. This refers to the self-loss of the energy storage system. .

[0234] Load indicators include the power outage time of sensitive loads and the core business guarantee rate, in order to ensure the reliability of power supply.

[0235] Load power outage time The cumulative duration of actual power outages during the control period is calculated as follows:

[0236] ;

[0237] In the formula, Let k be the duration of the power outage. Number of power outages .

[0238] Core business coverage rate Based on dynamic load priority calculation, the calculation formula is as follows:

[0239] ;

[0240] In the formula, The core load priority for the k-th time period is... , This refers to the duration of stable power supply to the core load during this period. This is the total duration of the period. .

[0241] The scenario indicators include the rate of electricity consumption meeting the target in the full-scale internet access scenario, the energy storage cost saving rate in the off-peak computing power scenario, the core load guarantee rate in the maintenance and power rationing scenario, and the success rate of responding to the sudden drop in output in the extreme weather scenario, so as to cover the full scenario requirements of the four scenarios.

[0242] Specifically, the rate of electricity consumption meeting the internet usage target. Calculate using the following formula:

[0243] ;

[0244] In the formula, This refers to the actual amount of electricity generated by distributed energy sources connected to the grid. This is the target value for internet power consumption predicted by S4.1. .

[0245] Energy storage cost savings Calculate using the following formula:

[0246] ;

[0247] In the formula, The operating cost of energy storage under conventional control mode, To optimize the operating costs of the control mode, .

[0248] The core load guarantee rate is calculated using the following formula:

[0249] ;

[0250] In the formula, To ensure stable power supply duration for core loads during maintenance periods, This represents the total maintenance time for the power grid. .

[0251] The success rate of responding to a sudden drop in output is calculated using the following formula:

[0252] ;

[0253] In the formula, To successfully compensate for the sudden drop in output, This represents the total power output that drops sharply under extreme weather conditions. .

[0254] Subsequently, based on the monitoring results under multi-dimensional indicators, the parameters of the prediction model and cross-category association features were optimized in a targeted manner to continuously improve the prediction accuracy and association mining accuracy.

[0255] The prediction model parameter optimization process uses gradient descent to iteratively update the weights and biases of the LSTM / GRU / CNN model. The core update formula is as follows:

[0256] ;

[0257] ;

[0258] In the formula, To predict the weights before the model update, The bias before the prediction model is updated. To predict the updated weights of the model, This is the bias after the prediction model is updated. For learning rate, , This is the gradient of the loss function with respect to the parameters. The loss function is calculated based on the monitored prediction error.

[0259] LSTM model: Updating the weight matrices of the forget gate, input gate, and output gate ( , , ) and bias term to optimize the accuracy of long-term trend prediction;

[0260] GRU model: Update reset gate, update gate weights ( , ), to improve the accuracy of mid-term fluctuation prediction;

[0261] Lightweight CNN Models: Updating Convolutional Kernels ( , ) and attention weights ( This reduces the prediction error of short-term power mutations.

[0262] Cross-category association feature optimization includes iterative optimization of association parameters, specifically adjustment of the source-grid association sensitivity coefficient and optimization of the load-storage association curve.

[0263] The process of adjusting the source-grid correlation sensitivity coefficient is as follows: If voltage fluctuations The sensitivity coefficient is corrected using the following formula: .

[0264] The optimization process of the load-storage correlation curve is as follows: If the energy storage SOC deviates from the target range Refit the load-storage correlation curve parameters using the following formula:

[0265] ;

[0266] In the formula, This represents the SOC deviation.

[0267] Subsequently, based on the monitoring results, the fault prevention and control strategy was optimized to achieve orderly reset after a fault and active blocking of fault propagation. The specific process includes the construction of fault chain blocking quantification logic and the optimization of fault reset strategy.

[0268] The fault chain blocking quantization logic construction includes fault determination coefficient calculation and fault chain blocking execution. During the fault determination coefficient calculation process, a fault determination coefficient is defined, which combines the fault signal strength and the associated influence range. The calculation formula is as follows:

[0269] ;

[0270] In the formula, The fault signal strength (0-1, 1 for undervoltage / overcurrent). The number of affected device nodes. This represents the total number of nodes.

[0271] Fault chain blocking execution refers to: when When a fault chain interruption command is triggered, the connection between the faulty branch and the main grid is immediately severed (e.g., a faulty branch of a photovoltaic inverter). Simultaneously, the energy storage system is instructed to rapidly replenish energy, with the replenishment power... This ensures that the core load is not affected.

[0272] Fault reset strategy optimization occurs after fault mitigation, that is... At that time, the system performs reverse resets according to short-term, medium-term, and long-term scales, with the reset threshold dynamically calculated.

[0273] ;

[0274] ;

[0275] ;

[0276] In the formula, , , These are the reset delay times on short, medium, and long time scales, respectively. The lower the risk of residual fault, the faster the reset, thus avoiding impact on the power grid.

[0277] The optimized prediction model, correlation features, and fault prevention strategies will be synchronized to the core processes of steps S2 to S5, realizing a complete closed loop from monitoring to evaluation, optimization, and iteration, continuously improving grid stability, load reliability, and energy storage utilization efficiency.

[0278] Furthermore, the present invention also provides an electrical system for a centralized control platform, used to implement the above-mentioned control method for a centralized control platform, comprising:

[0279] The data acquisition and preprocessing module is used to collect multi-source heterogeneous data, including source side, network side, load side, storage side, environment, and scenario-related data, and obtain a standardized dataset after preprocessing.

[0280] The cross-category association feature set extraction module is used to split the standardized dataset into data subsets according to each category, calculate the linear association strength of key cross-category data pairs, establish a paired linear quantization model, and extract the cross-category association feature set.

[0281] The three-level time scale feature set extraction module is used to integrate corresponding cross-category correlation features into various data subsets to obtain enhanced subsets, and extract time-series features in parallel according to long, medium and short time scales to obtain three-level time scale feature sets;

[0282] The full-scene prediction module is used to input the three-level time scale feature set into a pre-built prediction model adapted to different time scales, perform parallel prediction, and output the full-scene prediction result and trigger threshold.

[0283] The hierarchical control module is used to identify four core scenarios based on scenario-related data and match preset control strategies. Based on the full-scenario prediction results and trigger thresholds, hierarchical control is performed for long-term peak-valley energy storage scheduling, medium-term power grid stability control and fault prevention, and short-term pulse load replenishment.

[0284] The optimization module is used to monitor the control results in real time with multi-dimensional indicators. Based on the monitoring results, it continuously optimizes the prediction model, correlation features, and fault prevention and control strategies, and feeds them back to the cross-category correlation feature set extraction module, the three-level time scale feature set extraction module, the full-scenario prediction module, and the hierarchical control module, respectively.

[0285] The above data acquisition and preprocessing modules, cross-category correlation feature set extraction modules, three-level time scale feature set extraction modules, full-scene prediction modules, hierarchical control modules, and optimization modules are all mounted on the software layer of the electrical system, working in conjunction with the centralized control platform to achieve data acquisition, analysis, and control decisions.

[0286] This invention addresses the problem of information silos in existing centralized control consoles during data processing by mining cross-category data from the source, grid, load, and storage sides, constructing multi-timescale prediction and control logic, and establishing a closed-loop optimization mechanism. It also solves the problems of insufficient time-scale adaptability in prediction and control, and inadequate foresight and accuracy in control. This results in stable grid operation, uninterrupted power supply to sensitive loads, and efficient utilization of energy storage, improving prediction accuracy, control adaptability, equipment reliability, and utilization efficiency.

[0287] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any indirect modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A control method of a centralized control center, characterized by, The method comprises the following steps: Step S1: Collecting multi-source heterogeneous data including source side, network side, load side, storage side, environment, and scene correlation data, and obtaining a standardized data set after preprocessing; Step S2: Splitting the standardized data set according to each side category to obtain a data subset, calculating the linear correlation strength of the cross-category key data pair, establishing a paired linear quantification model, and extracting a cross-category correlation feature set; Step S3: Integrating the corresponding cross-category correlation features into each type of data subset to obtain an enhanced subset, and extracting time sequence features according to long, medium, and short time scales in parallel to obtain a three-level time scale feature set; Step S4: Inputting the three-level time scale feature set into a pre-established prediction model adapted to different time scales to perform parallel prediction to output a full-scene prediction result and a trigger threshold; Step S5: Identifying four types of core scenes based on scene correlation data and matching preset control strategies, and performing hierarchical control based on the full-scene prediction result and the trigger threshold according to long-time peak-valley energy storage scheduling, medium-time power grid stability control and fault prevention, and short-time pulse load energy compensation; Step S6: Real-time multi-dimensional index monitoring of the control result, continuous optimization of the prediction model, correlation features, and fault prevention strategies based on the monitoring result, and feedback to steps S2 to S5.

2. The control method of a centralized control center according to claim 1, wherein In step S2, the linear correlation strength is calculated based on the Pearson correlation coefficient. Based on the total number of data, the deviation of the normalized value of the two types of data from the respective mean value is compared, and the correlation coefficient is obtained by the ratio of the covariance to the standard deviation.

3. The control method of a centralized control center according to claim 1, wherein The cross-category key data pair includes key data pairs between source side data and load side data, between source side data and network side data, between load side data and storage side data, and between network side data and storage side data.

4. The control method of a centralized control center according to claim 3, wherein In step S3, each type of data subset and the paired cross-category correlation features are obtained by horizontal splicing to obtain an enhanced subset.

5. The control method of a centralized control center according to claim 4, wherein The three-level time scale feature set includes a long-time scale feature set, a medium-time scale feature set, and a short-time scale feature set; The long-time scale feature set includes source side output trend features, load side power accumulation features, and network side peak-valley price cycle features; The medium-time scale feature set includes network side voltage stability features, source side output fluctuation features, and cross-category correlation sensitivity features; The short-time scale feature set includes load side pulse load features and storage side super capacitor features.

6. The control method of a centralized control center according to claim 1, wherein In step S4, the full-scene prediction result and the trigger threshold include long-time energy storage charging and discharging thresholds, medium-time voltage deviation and fault probability, and short-time power mutation prediction value and energy compensation threshold.

7. The control method of a centralized control center according to claim 6, wherein In step S4, LSTM neural network models, GRU neural network models, and lightweight CNN models combined with attention mechanisms are established for long-time, medium-time, and short-time time scales.

8. The control method of a centralized control center according to claim 1, wherein In step S5, a reverse correction mechanism is also established for realizing cross-scale control collaboration from short-time to medium-time, and then to long-time.

9. An electrical system of a centralized control center, characterized by A control method for a centralized control platform according to claim 1, comprising: a data acquisition and preprocessing module for acquiring multi-source heterogeneous data including source side, network side, load side, storage side, environment, and scene correlation data, and obtaining a standardized data set after preprocessing; The cross-category association feature set extraction module is configured to split the standardized data set according to each side category to obtain a data subset, calculate the linear association strength of a cross-category key data pair, establish a paired linear quantification model, and extract a cross-category association feature set; The three-level time scale feature set extraction module is configured to fuse the corresponding cross-category association features into each category data subset to obtain an enhanced subset, and extract time sequence features according to long, medium and short three-level time scales to obtain a three-level time scale feature set; The full-scene prediction module is configured to input the three-level time scale feature set into a prediction model adapted to different time scales, and perform parallel prediction to output a full-scene prediction result and a trigger threshold; The hierarchical regulation module is configured to identify four types of core scenes based on scene association data and match a preset regulation strategy, perform hierarchical regulation according to the full-scene prediction result and the trigger threshold, and execute long-time peak-valley energy storage scheduling, medium-time power grid stability regulation and fault prevention, and short-time pulse load energy compensation; The optimization module is configured to monitor the regulation result in real time in multiple dimensions, continuously optimize the prediction model, the association feature, the fault prevention strategy based on the monitoring result, and feed back to the cross-category association feature set extraction module, the three-level time scale feature set extraction module, the full-scene prediction module, and the hierarchical regulation module, respectively.

Images