Microgrid energy optimization scheduling method based on deep reinforcement learning

Abstract

The application discloses a micro-grid energy optimization scheduling method based on deep reinforcement learning, relates to the technical field of power dispatching, and comprises the following steps: performing multi-dimensional feature extraction on micro-grid operation state data and executing nonlinear mapping deduction, taking a micro-grid comprehensive operation optimization target as a target, generating a preliminary power distribution instruction, and mapping the preliminary power distribution instruction into a dynamic physical safety constraint set to obtain an optimal scheduling action; executing a control instruction corresponding to the optimal scheduling action to generate a transient response waveform, and combining node voltage and branch power flow distribution to obtain operation feedback and scheduling gain; constructing a scheduling gain criterion based on the operation feedback and the scheduling gain, and executing parameter feedback iteration to update internal correlation parameters in the nonlinear mapping deduction according to the scheduling gain criterion to obtain an optimization strategy mapping logic; and the economic benefit and dynamic response quality of energy scheduling are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of power dispatching technology, and in particular to a microgrid energy optimization dispatching method based on deep reinforcement learning. Background Technology

[0002] With the increasing penetration of distributed energy resources and the widespread application of microgrid technology, traditional microgrid energy management systems (EMS) often employ scheduling schemes based on model predictive control (MPC) or heuristic optimization algorithms. These technologies construct mathematical models that include distributed power sources, energy storage systems, and flexible loads. Under the premise of satisfying static constraints such as power balance and voltage stability, they utilize mathematical programming methods to solve for the optimal output sequence of each execution unit. These technologies have demonstrated good economic efficiency and stability in scenarios such as industrial park distribution networks and isolated power supply systems, achieving basic power quality assurance and initial integration of distributed resources.

[0003] However, when dealing with microgrid environments characterized by high randomness and time-varying features, existing technologies often face the challenge of balancing computational real-time performance with the rigidity of physical constraints. Traditional energy dispatch schemes often rely on pre-defined distribution modeling of system uncertainties, but in the context of explosive growth in multidimensional state space, their online optimization process suffers from severe computational lag. At the same time, existing intelligent optimization logics struggle to simultaneously extract deep implicit features and decouple dynamic physical safety boundaries in real time when generating control decisions, resulting in dispatch commands that may cross physical safety red lines during transient processes or complex operating conditions, lacking an effective adaptive closed-loop verification mechanism. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a microgrid energy optimization scheduling method based on deep reinforcement learning to solve the problem of decreased decoupling accuracy caused by the lack of a real-time correction mechanism in dynamic environments.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a microgrid energy optimization scheduling method based on deep reinforcement learning, which includes: S1, acquire microgrid operating status data, and construct a dynamic physical security constraint set based on the uncertainty distribution of historical operating samples; S2 extracts multi-dimensional features from the microgrid operating status data and performs nonlinear mapping deduction. With the microgrid comprehensive operation optimization objective as the goal, it generates preliminary power allocation instructions and maps them into the dynamic physical security constraint set to obtain the optimal scheduling action. S3 executes the control command corresponding to the optimal scheduling action, generates a transient response waveform, and obtains operation feedback and scheduling gain by combining the node voltage and branch power flow distribution. S4. Based on the operation feedback and scheduling gain, a scheduling gain criterion is constructed, and the internal correlation parameters in the nonlinear mapping derivation are updated iteratively according to the parameter feedback to obtain the optimization strategy mapping logic. S5 utilizes the optimization strategy mapping logic to perform rolling adaptive control of energy flow in continuous scheduling cycles and synchronously refreshes the dynamic physical safety constraint set.

[0007] Preferably, the method for constructing the dynamic physical security constraint set includes: Collect multi-source heterogeneous data including distributed power generation, load demand, battery state of charge and tiered electricity pricing, and generate microgrid operation status data through normalization and time series feature enhancement processing. Based on historical operating samples, the probability distribution characteristics of distributed power generation output and load demand are analyzed to determine the uncertainty distribution parameters and construct the corresponding random disturbance set. The worst-case scenario search is performed on the random disturbance set, and the boundary optimization is performed in combination with the multi-dimensional physical operation constraints of the microgrid to obtain the dynamic physical security constraint set.

[0008] Preferably, the method for generating the initial power allocation command includes: Multi-level residual mapping processing is performed on the microgrid operation status data. By establishing cross-layer short-circuit connections between multiple linear processing layers, multi-dimensional correlation features in the microgrid operation status data are fused to obtain a multi-dimensional feature vector. By using a nonlinear mapping function to perform nonlinear dimensional transformation and feature inference on multidimensional feature vectors, the nonlinear fitting strength between scheduling features and decision logic is enhanced, and the action probability distribution is obtained. Multi-dimensional operational index data collected during the microgrid operation process are standardized, and weighted fusion is performed using preset scheduling preference weight coefficients to construct a comprehensive microgrid operation optimization target. Taking the comprehensive microgrid operation optimization target as the optimization direction, gradient evolution is performed on the action probability distribution to generate preliminary power allocation instructions.

[0009] Preferably, the method for obtaining the optimal scheduling action includes: Extract the nominal power values ​​of each energy unit in the preliminary power allocation command, construct the power vector to be checked, and transform the dynamic physical security constraint set into the geometric boundary constraints of the power space; Determine whether the power vector to be checked is inside the geometric boundary constraint; if the power vector to be checked is outside the geometric boundary constraint, then use the projection operator to perform spatial mapping to map the power vector to be checked to the boundary point on the geometric boundary constraint that is closest to the power vector to be checked, and obtain the corrected power vector; The corrected power vector is identified as the optimal scheduling action; if the power vector to be checked does not exceed the limit, it is identified as the optimal scheduling action.

[0010] Preferably, the method for generating the transient response waveform includes: Based on the power settings of each energy unit in the optimal scheduling action, a standardized communication command stream is generated; Standardized communication command streams are sent to the execution terminal to drive each energy unit to adjust its output state according to a preset response slope; Monitor instantaneous fluctuation data during the physical state adjustment process within the microgrid; summarize the instantaneous fluctuation data, perform time series feature alignment, and generate transient response waveforms for the dynamic process of power balance regulation in the microgrid.

[0011] Preferably, the method for obtaining operational feedback and scheduling gain includes: Calculate the sliding standard deviation of the transient response waveform within the sampling window; when the sliding standard deviation is continuously within the preset fluctuation threshold range, the microgrid is determined to have entered steady state, and the voltage of each physical node and the power flow distribution of each branch are extracted to generate a steady-state operation profile. Extract transient performance indicators from transient response waveforms and construct operational feedback that includes both dynamic and static features by combining steady-state operational sections; The initial economic and safety indicators of the microgrid before executing the optimal scheduling action are obtained. These indicators are compared with the real-time data of multi-dimensional indicators in the operation feedback to obtain the improvement of operation indicators. Multi-dimensional indicator weight aggregation is performed to quantitatively evaluate the degree of improvement of the microgrid operation performance by the optimal scheduling action and obtain the scheduling gain.

[0012] Preferably, the method for constructing the scheduling gain criterion includes: The runtime feedback is time-stamped and combined with the scheduling gain to construct a spatiotemporal feedback sequence; and the deviation of the scheduling gain from the expected value of the baseline gain is determined. The deviation magnitude is corrected using transient performance indicators in the spatiotemporal feedback sequence to implement a punitive reduction when the transient performance indicators exceed a preset safety threshold. By summarizing the deviation magnitude and spatiotemporal feedback sequence after safety correction, a scheduling gain criterion is established with the evaluation orientation of maximizing scheduling gain and minimizing physical constraint default rate.

[0013] Preferably, the method for obtaining the optimized strategy mapping logic includes: Retrieve the internal correlation parameters corresponding to each processing level in the nonlinear mapping derivation, and establish the correlation gradient sensitivity matrix of the internal correlation parameters relative to the scheduling gain criterion; By using the correlation gradient sensitivity matrix to perform weight reshaping and direction correction on the gradient flow generated by the scheduling gain criterion, the feedback iteration direction of the internal correlation parameters is determined; error backpropagation is performed according to the feedback iteration direction to realize the iterative update of the internal correlation parameters and obtain the optimization strategy mapping logic.

[0014] Preferably, the method for rolling adaptive control of energy flow in continuous scheduling cycles includes: At the start of each control step, acquire the latest multi-source heterogeneous data and update the microgrid operating status data for the current step; input the microgrid operating status data into the optimization strategy mapping logic to deduce and generate the preliminary power allocation command for the current step. The initial power allocation command is mapped to the synchronously refreshed dynamic physical security constraint set using the projection operator, the optimal scheduling action for the current step size is determined and issued for execution; After the current step size control is completed, the time window is slid forward by one control step size, and the state update, instruction deduction and constraint correction actions are executed cyclically to achieve rolling adaptive control of energy flow in continuous scheduling cycles.

[0015] Preferably, the method for synchronously refreshing the dynamic physical security constraint set includes: Within each control step, real-time measurements of distributed power and load demand are acquired and the historical operating samples are incrementally updated. The updated historical operating samples are then used to update the uncertainty distribution parameters and reconstruct the random disturbance set. The current topology connection status of the microgrid and the latest multidimensional physical operation constraints of the microgrid are monitored as optimization constraints. The worst-case scenario search and boundary optimization are performed on the random disturbance set to obtain a new envelope optimization boundary and confirm it as the dynamic physical security constraint set of the current step size, so as to complete the synchronization update.

[0016] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the microgrid energy optimization scheduling method based on deep reinforcement learning as described in the first aspect of the present invention.

[0017] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the microgrid energy optimization scheduling method based on deep reinforcement learning as described in the first aspect of the present invention.

[0018] The beneficial effects of this invention are as follows: By establishing a deep reinforcement learning architecture with physical boundary perception and dynamic constraint self-verification capabilities, the technical problem of traditional artificial intelligence scheduling instructions easily violating physical constraints under complex random disturbances is solved, and the deep coupling between the algorithm decision logic and the power system safety boundary is realized. By using a feedback mechanism that integrates transient and steady-state multi-dimensional features to guide the adaptive evolution of the strategy, the shortcomings of single-scale scheduling under complex operating conditions are effectively overcome. On the basis of ensuring the "red line" of physical safety operation of the microgrid throughout the entire cycle, the economic benefits and dynamic response quality of energy scheduling are greatly improved, and a dual leap in system operation reliability and intelligent scheduling capabilities is achieved. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the microgrid energy optimization scheduling method based on deep reinforcement learning in this invention.

[0021] Figure 2 This is a flowchart of obtaining the dynamic physical security constraint set in this invention.

[0022] Figure 3 This is a flowchart for determining the optimal scheduling action in this invention.

[0023] Figure 4 This is a flowchart of the rolling adaptive control in this invention. Detailed Implementation

[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0025] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0026] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0027] Reference Figure 1 , Figure 2 , Figure 3 and Figure 4 This is one embodiment of the present invention, which provides a microgrid energy optimization scheduling method based on deep reinforcement learning, including the following steps: Methods for constructing dynamic physical security constraint sets include: Multi-source heterogeneous data, including distributed power generation, load demand, battery state of charge, and tiered electricity pricing, are collected and processed through normalization and time-series feature enhancement to generate microgrid operation status data.

[0028] Specifically, the system acquires distributed power generation, load demand, battery state of charge, and tiered electricity pricing data in real time through power quality monitoring instruments and energy management units within the microgrid, aggregating these data to form multi-source heterogeneous data. A normalization operator is then used to perform linear mapping on parameters of different dimensions within the multi-source heterogeneous data, mapping them all to a preset per-unit value range (example value is 0.95) to eliminate dimensional differences. Subsequently, a sliding window operator is used to group the mapped multi-source heterogeneous data by time dimension. The differential characteristics of distributed power generation and load demand between adjacent control steps are calculated, and these differential characteristics are then concatenated and encapsulated with the mapped multi-source heterogeneous data in a time series format to complete time-series feature enhancement processing (example value is 0.02 for power change trends). Finally, microgrid operating status data containing both physical operating status and time evolution characteristics is generated.

[0029] Based on historical operating samples, the probability distribution characteristics of distributed power generation output and load demand are analyzed to determine the uncertainty distribution parameters and construct the corresponding random disturbance set.

[0030] Specifically, the historical observation sequence of distributed power and load demand stored in the historical operation sample is retrieved. By performing probability density fitting on the occurrence frequency of sample points in the historical observation sequence, the parameter estimate that maximizes the occurrence probability of the observation sequence is found, and the uncertainty distribution parameter reflecting the fluctuation range of distributed power and load demand is determined. The uncertainty distribution parameter is exemplarily the mean and variance. Based on the determined uncertainty distribution parameters, the uncertainty variable of each dimension is divided into equally probable intervals. For example, the probability distribution curve of photovoltaic output power is divided into ten intervals with equal cumulative probabilities along the horizontal axis. Each interval carries an exemplary 10% probability of occurrence. Hierarchical sampling is performed within each interval. The sampled variables of each dimension are combined in a multidimensional space in a random arrangement to generate multiple scenario combinations that include distributed power fluctuation components and load demand fluctuation components. The maximum deviation of a single power fluctuation amplitude is taken as 10% of the nominal value. The multiple scenario combinations are summarized and encapsulated to construct a random disturbance set that can cover the random fluctuations of distributed power sources and loads during the operation of the microgrid.

[0031] The worst-case scenario search is performed on the random disturbance set, and the boundary optimization is performed in combination with the multi-dimensional physical operation constraints of the microgrid to obtain the dynamic physical security constraint set.

[0032] Specifically, among various scenario combinations covered by the random disturbance set, the power distribution under each scenario is iteratively simulated using power flow calculation methods to find the extreme fluctuation scenario that causes the microgrid power flow distribution to deviate from the preset equilibrium point. The preset equilibrium point is exemplified as the ideal operating state with a nominal voltage of 11kV and a frequency of 50Hz. The worst-case scenario is determined, exemplified as the extreme combination of a sudden increase in load and a sudden decrease in photovoltaic power. Multi-dimensional physical operating constraints of the microgrid, including physical node voltage limits, branch current limits, and frequency fluctuation range limits, are obtained. Using the multidimensional physical operation constraints of the microgrid as hard constraint boundaries, the interior point method is used to perform a feasibility traversal of the active and reactive power of each node of the microgrid under the worst-case scenario. The extreme points of the allowable fluctuations of each energy unit under the premise of satisfying the multidimensional physical operation constraints of the microgrid are searched to complete the boundary optimization and finally obtain the dynamic physical security constraint set that encompasses all physical security states.

[0033] In microgrid dispatching, existing linear models or general reinforcement learning frameworks struggle to handle high-dimensional nonlinear features, easily leading to feature loss and convergence difficulties, and failing to accurately capture the temporal correlations of multi-source heterogeneous data. A more critical drawback is that, due to the lack of robust physical constraint mechanisms, dispatching commands are highly susceptible to triggering voltage or current overruns during periods of severe power fluctuations, making it difficult to meet the safety and reliability requirements of industrial-grade operation. Therefore, this invention introduces a multi-level residual mapping and dynamic physical safety constraint projection mechanism, specifically as follows: Methods for generating preliminary power allocation instructions include: Multi-level residual mapping processing is performed on the microgrid operation status data. By establishing cross-layer short-circuit connections between multiple linear processing layers, multi-dimensional correlation features in the microgrid operation status data are fused to obtain multi-dimensional feature vectors.

[0034] Specifically, the generated microgrid operation status data is input into a preset deep neural network model. Multiple cascaded linear processing layers with cross-layer short-circuit connections are used to perform feature transformation on the microgrid operation status data. The input data of the previous linear processing layer is directly accumulated into the output result of the next linear processing layer. For example, the initial value of distributed power after scaling is superimposed across layers onto the power fluctuation feature after feature transformation. Multi-dimensional correlation features are fused to output a multi-dimensional feature vector that represents the implicit logic of the current operation status of the microgrid. The pre-set process of the deep neural network model is as follows: A benchmark sample set containing historical microgrid operation status data and corresponding historical optimal scheduling instructions is retrieved to determine the input and output dimensions of the deep neural network model. For example, the input dimension corresponds to the total number of features in the microgrid operation status data. Then, an offline training algorithm is used to iteratively optimize the weight matrix and bias vector within the deep neural network model. By minimizing the mean square error between the deep neural network model output and the benchmark sample set, the deep neural network model acquires a preliminary mapping capability between distributed power source fluctuation trends and microgrid scheduling logic. When the mean square error reaches the preset accuracy requirement, and the error value is less than the example of 0.005, all level parameters of the deep neural network model are solidified. It should be noted that the deep neural network model consists of an input layer, multiple interconnected residual blocks with cross-layer short-circuit connections, and an output layer containing the action decision space. Each residual block contains two linear processing layers and a skip connection path. The input features of the residual block are added element-wise with the output features transformed by the second linear processing layer through the skip connection path to form an identity mapping channel. Multiple residual blocks are stacked according to a preset depth, exemplified by stacking four residual blocks, to perform multi-level feature abstraction from shallow to deep on the microgrid operating status data, ensuring that the multi-dimensional feature vector can simultaneously cover the instantaneous physical numerical features and long-term time evolution correlation features of the microgrid. The output layer pre-sets corresponding neuron nodes based on the adjustable power range of distributed power sources and batteries in the microgrid to construct an action decision space. For example, if the microgrid contains 3 controllable energy units, the output layer pre-sets 3 independent action dimensions, each corresponding to a power adjustment step size, thereby transforming the abstract feature vector into an action mapping basis corresponding to the actual physical devices.

[0035] By using a nonlinear mapping function to perform nonlinear dimensional transformation and feature inference on multidimensional feature vectors, the nonlinear fitting strength between scheduling features and decision logic is enhanced, and the action probability distribution is obtained.

[0036] Specifically, the multidimensional feature vector is input into the fully connected layer to perform spatial basis transformation. The abstract features are projected onto the decision score space through a linear combination of the weight matrix and the bias vector, generating the original score vector that reflects the output tendency of each energy unit. The original score vector is transformed element-wise by a nonlinear mapping function at the end of a deep neural network model. The nonlinear mapping function can be exemplified by the ReLU activation function or the Softmax normalized exponential function. By performing nonlinear truncation or exponential stretching on the components of the original score vector, the nonlinear mapping relationship between the microgrid operating state and the dispatch command is captured, and the fitting strength between the multidimensional feature vector and the dispatch action is enhanced. The original score vector after the nonlinear dimension transformation is mapped to the action decision space, and the probability of each energy unit performing different power allocation actions is calculated to obtain the action probability distribution covering all possible action combinations.

[0037] Multi-dimensional operational index data collected during the microgrid operation process are standardized, and weighted fusion is performed using preset scheduling preference weight coefficients to construct a comprehensive microgrid operation optimization target. Taking the comprehensive microgrid operation optimization target as the optimization direction, gradient evolution is performed on the action probability distribution to generate preliminary power allocation instructions.

[0038] It should be noted that the system acquires multi-dimensional operational index data, including electricity purchase cost, operating loss, and voltage deviation, during the microgrid operation process in real time and performs standardization processing. The system then uses preset scheduling preference weight coefficients to multiply and sum the multi-dimensional operational index data. The preset scheduling preference weight coefficients are set to an example of 0.6 for economic cost and 0.4 for voltage stability. This achieves weighted fusion of multiple objective attributes and constructs a comprehensive microgrid operation optimization objective that reflects the current performance of the microgrid. Using the microgrid integrated operation optimization objective as the objective function, the gradient flow of the action probability distribution relative to the microgrid integrated operation optimization objective is calculated, driving the action probability distribution to iteratively optimize along the gradient evolution direction. By adjusting the distribution parameters of the output layer of the deep neural network model, the optimal action expectation value that maximizes the microgrid integrated operation optimization objective is searched, and the physical output value corresponding to the expectation value is extracted in the action decision space, generating a preliminary power allocation instruction containing the target values ​​of active power and reactive power of each energy unit.

[0039] Methods for obtaining the optimal scheduling action include: Extract the nominal power values ​​of each energy unit from the preliminary power allocation command, construct the power vector to be checked, and transform the dynamic physical security constraint set into the geometric boundary constraints of the power space.

[0040] It should be noted that, from the initial power allocation instruction, the nominal power value of each energy unit is extracted sequentially according to the order of splicing of distributed power, battery state of charge and load demand in multi-source heterogeneous data, and the extracted nominal power values ​​of each energy unit are vectorized and combined to construct the power vector to be verified. Simultaneously, the power flow equations are used to decouple and transform the physical quantities such as node voltage, branch current and system frequency contained in the dynamic physical safety constraint set. The constraints based on physical quantities are transformed into convex space constraints with active power and reactive power as coordinate axes. Thus, the dynamic physical safety constraint set is transformed into geometric boundary constraints of the power space. The geometric boundary constraints are exemplified as a closed polygonal region formed with energy storage output and distributed power output as coordinate axes.

[0041] Determine whether the power vector to be checked is inside the geometric boundary constraint; if the power vector to be checked is outside the geometric boundary constraint, then use the projection operator to perform spatial mapping, mapping the power vector to be checked to the boundary point on the geometric boundary constraint that is closest to the power vector to be checked, and obtain the corrected power vector.

[0042] Specifically, a ray is drawn from the coordinate point of the power vector to be checked in the power space in any fixed direction. The number of intersection points between the ray and each boundary line segment defined by the geometric boundary is counted. If the number of intersection points is odd, it is determined that the power vector to be checked is within the envelope defined by the geometric boundary. If the number of intersection points is even, it is determined that the power vector to be checked is outside the envelope defined by the geometric boundary. If the judgment result indicates that the power vector to be checked is outside the envelope defined by the geometric boundary limit, then by solving for the minimum L2 norm between the power vector to be checked and each point on the geometric boundary limit, the minimum Euclidean distance of the power vector to be checked on the geometric boundary limit is calculated. Spatial mapping is then performed to project the power vector to be checked onto the boundary point on the geometric boundary limit closest to the power vector to be checked. This forces the coordinates of the power vector exceeding the branch current limit back to the edge line of the geometric boundary limit, eliminating the over-limit components in the initial power allocation command, and obtaining the corrected power vector, expressed as: ; in, It is a power vector correction. It is the power vector to be checked. Geometric boundary constraints Any feasible power point vector within and on the boundary, It is the 2-norm (Euclidean norm). It is the set of variables when the objective function takes its minimum value. It belongs to a relationship.

[0043] The corrected power vector is identified as the optimal scheduling action; if the power vector to be checked does not exceed the limit, it is identified as the optimal scheduling action.

[0044] It should be noted that when the power vector to be checked is outside the geometric boundary limit, the corrected power vector generated after mapping by the projection operator is identified as the optimal scheduling action; if the power vector to be checked is determined to be within the limit during the position detection process, that is, the coordinates of the power vector to be checked are inside the geometric boundary limit, then the power vector to be checked is directly identified as the optimal scheduling action, and the optimal scheduling action that conforms to the multi-dimensional physical operation limit of the microgrid is output. By introducing a multi-level residual mapping and a dynamic physical safety constraint projection mechanism, a deep coupling between data-driven decision-making and the laws of power physics is achieved. The residual block significantly enhances the model's ability to abstract implicit logic using identity mapping channels, ensuring accurate capture of initial instructions. By using projection operators to perform spatial verification, over-limit instructions can be forcibly corrected to within geometric boundaries. This optimizes power purchase costs and losses while eliminating the hidden dangers of branch overload and voltage exceedance, ensuring that dispatching actions always operate within the safe and feasible domain.

[0045] Methods for generating transient response waveforms include: Based on the power settings of each energy unit in the optimal scheduling action, a standardized communication command stream is generated; Standardized communication command streams are sent to the execution terminal to drive each energy unit to adjust its output state according to a preset response slope.

[0046] Specifically, the active and reactive power target values ​​in the optimal scheduling action are analyzed and used as the power setting values ​​for each energy unit. Data message encapsulation is performed using the industrial control communication protocol, and device address codes and check bits are added to generate a standardized communication command stream. The standardized communication command stream is sent to the execution terminal through the communication link, which drives the power electronic converter to perform pulse width modulation duty cycle update, so that each energy unit adjusts its output state according to the preset response slope. The preset response slope is set to 10% of the rated power per second, guiding the microgrid to smoothly transition from the current equilibrium point to the target equilibrium point.

[0047] Monitor instantaneous fluctuation data during the physical state adjustment process within the microgrid; summarize the instantaneous fluctuation data, perform time series feature alignment, and generate transient response waveforms for the dynamic process of power balance regulation in the microgrid.

[0048] Specifically, a micro phasor measurement unit is used to capture the rapid changes in active power, reactive power, frequency, and voltage amplitude during the state adjustment process within the microgrid, forming instantaneous fluctuation data. The instantaneous fluctuation data is then aligned with time series features using a unified timestamp, and a linear interpolation algorithm is used to fill in sampling gaps to eliminate feature misalignment. Finally, the aligned data is mapped to coordinates and connected to curves in chronological order to generate transient response waveforms.

[0049] Methods for obtaining runtime feedback and scheduling gains include: Calculate the sliding standard deviation of the transient response waveform within the sampling window; when the sliding standard deviation is continuously within the preset fluctuation threshold range, the microgrid is determined to have entered steady state, and the voltage of each physical node and the power flow distribution of each branch are extracted to generate a steady-state operation profile.

[0050] It should be noted that real-time sampling data within a specific time period (exemplarily, a time interval containing the most recent 10 consecutive sampling points, with the total length of the specific time period set to 50ms) of the transient response waveform is extracted, and the standard deviation of the active power and frequency sampling values ​​within that time period is calculated to obtain the sliding standard deviation. The sliding standard deviation is then compared with a preset fluctuation threshold in real-time. If the sliding standard deviation is less than the preset fluctuation threshold within a preset continuous time period, it is determined that the microgrid has completed the power balance adjustment process and entered steady-state operation. An example of the preset fluctuation threshold is 0.01% of the 50Hz frequency value at the equilibrium point. The energy management unit reads and records the measured values ​​of the voltage amplitude of each physical node and the power flow distribution of each branch within the microgrid. The voltage of each physical node and the power flow distribution of each branch are encapsulated spatially according to the topology to generate a steady-state operation profile reflecting the final physical balance state of the microgrid after adjustment.

[0051] Transient performance indicators are extracted from the transient response waveform and combined with the steady-state operation section to construct an operation feedback that includes dynamic and static characteristics.

[0052] Specifically, transient performance indicators reflecting the dynamic regulation quality of the microgrid are identified and extracted from the transient response waveform. Examples of transient performance indicators include maximum overshoot, regulation time, and frequency change rate. The extracted transient performance indicators are then combined with the voltage of each physical node and the power flow distribution of each branch contained in the generated steady-state operation section to perform multi-dimensional data splicing. This allows the transient performance indicators to serve as dynamic features characterizing the evolution trajectory of the regulation process, while the steady-state operation section serves as static features characterizing the final state of the regulation result. The dynamic and static features are combined and merged using a structured encapsulation operator to construct the operation feedback.

[0053] The initial economic and safety indicators of the microgrid before executing the optimal scheduling action are obtained. These indicators are compared with the real-time data of multi-dimensional indicators in the operation feedback to obtain the improvement of operation indicators. Multi-dimensional indicator weight aggregation is performed to quantitatively evaluate the degree of improvement of the microgrid operation performance by the optimal scheduling action and obtain the scheduling gain.

[0054] It should be noted that the operation data of the microgrid before receiving the standardized communication command stream is retrieved to determine the initial economic indicators and initial safety indicators before the microgrid performs the optimal scheduling action. The initial economic indicator is exemplified by the instantaneous cost under the real-time electricity purchase price, and the initial safety indicator is exemplified by the initial deviation of the voltage of each physical node from the ideal value.

[0055] The initial economic indicators and initial safety indicators are respectively compared with the real-time data of multi-dimensional indicators such as electricity purchase cost, operation loss and voltage deviation extracted from the operation feedback to obtain the optimization and improvement of the operation indicators. By using preset scheduling preference weight coefficients to perform weighted summation on the optimization and improvement of various operating indicators, the change score of the comprehensive score is calculated through the aggregation of multi-dimensional indicator weights, and the degree of improvement of the microgrid operating performance by the optimal scheduling action is quantitatively evaluated to obtain the scheduling gain.

[0056] In the field of microgrid dispatching, existing technologies often focus only on single economic benefits or static indicators when updating strategies, lacking effective constraints on the transient stability of microgrids during dynamic regulation. Traditional algorithms, when performing gradient optimization, typically determine the direction of the gradient flow solely by a single reward function, failing to fully consider the sensitivity relationship between control parameters and multidimensional physical constraints. This results in slow strategy iteration speeds and a high likelihood of physical constraint violations when facing random disturbances from distributed generation sources, making it difficult to achieve a balance between safety and efficiency. Therefore, this invention introduces a spatiotemporal feedback sequence and an associated gradient sensitivity matrix, as detailed below: Methods for constructing scheduling gain criteria include: The runtime feedback is time-stamped and combined with the scheduling gain to construct a spatiotemporal feedback sequence; and the deviation of the scheduling gain from the expected value of the baseline gain is determined.

[0057] Specifically, a unified timestamp is used to perform time-series marking on the operational feedback containing both dynamic and static characteristics. The operational feedback with time-series marking is dimensionally correlated with the calculated scheduling gain. Data is arranged according to the order of the control cycle to construct a spatiotemporal feedback sequence that can simultaneously reflect the spatial state distribution and temporal evolution effect of the microgrid.

[0058] The baseline expected gain is exemplified by the average historical average performance improvement. The algebraic difference between the current scheduling gain and the baseline expected gain is calculated to determine the deviation of the scheduling gain from the baseline expected gain, which reflects the quality of the current scheduling scheme.

[0059] The deviation magnitude is corrected using transient performance indicators in the spatiotemporal feedback sequence to implement a punitive reduction when the transient performance indicators exceed a preset safety threshold.

[0060] Specifically, transient performance indicators, including maximum overshoot, settling time, and frequency change rate, are extracted from the spatiotemporal feedback sequence, and preset safety thresholds are obtained for each transient performance indicator. It is then determined whether the extracted transient performance indicators exceed the preset safety thresholds. If the result indicates that the transient performance indicators exceed the preset safety thresholds, for example, the frequency change rate exceeds 0.5 Hz per second, a nonlinear penalty factor is used to penalize the deviation amplitude. By reducing the value of the deviation amplitude, the positive contribution of the unsafe adjustment process to the evaluation result is weakened, thus achieving a safety correction process for the deviation amplitude. If the transient performance indicators are within the preset safety threshold range, the original value of the deviation amplitude remains unchanged.

[0061] By summarizing the deviation magnitude and spatiotemporal feedback sequence after safety correction, a scheduling gain criterion is established with the evaluation orientation of maximizing scheduling gain and minimizing physical constraint default rate.

[0062] Specifically, the deviation magnitude after safety correction is aggregated with the spatiotemporal feedback sequence containing dynamic and static characteristics to establish a scheduling gain criterion. The scheduling gain criterion takes maximizing scheduling gain as the performance optimization guideline to ensure that scheduling actions can effectively improve economic and safety indicators. At the same time, it takes minimizing the physical constraint default rate as the safety guarantee guideline. By analyzing whether the voltage of each physical node and the power flow distribution of each branch in the spatiotemporal feedback sequence touch the boundary of the dynamic physical safety constraint set, the default risk in the microgrid operation process is quantitatively assessed, and the scheduling gain criterion is formed.

[0063] Methods for obtaining the optimization strategy mapping logic include: Retrieve the internal correlation parameters corresponding to each processing level in the nonlinear mapping derivation, and establish the correlation gradient sensitivity matrix of the internal correlation parameters relative to the scheduling gain criterion.

[0064] Specifically, the weight matrix and bias vector contained in each fully connected layer and residual block that performs nonlinear dimensional transformation and feature inference are retrieved from the deep neural network model as internal correlation parameters. The established scheduling gain criterion is used as the objective evaluation function. The partial derivative of the scheduling gain criterion with respect to each internal correlation parameter is calculated. The partial derivative values ​​are used to quantify the influence of parameter fluctuations at each processing level on the scheduling gain. All partial derivatives are arranged and combined according to the topological position of the internal correlation parameters in the deep neural network model to establish the correlation gradient sensitivity matrix of the internal correlation parameters with respect to the scheduling gain criterion.

[0065] By using the correlation gradient sensitivity matrix to perform weight reshaping and direction correction on the gradient flow generated by the scheduling gain criterion, the feedback iteration direction of the internal correlation parameters is determined; error backpropagation is performed according to the feedback iteration direction to realize the iterative update of the internal correlation parameters and obtain the optimization strategy mapping logic.

[0066] Specifically, the original gradient flow generated by the scheduling gain criterion is obtained, matrix multiplication is performed between the associated gradient sensitivity matrix and the original gradient flow, the magnitude of the original gradient flow is weighted and scaled by introducing the deviation magnitude after safety correction, and the gradient vector direction is projected and corrected in combination with the physical constraint default rate, thereby completing the weight reshaping and direction correction of the gradient flow and determining the feedback iteration direction of the internal associated parameters. Subsequently, following the feedback iteration direction, the corrected gradient information is propagated backward from the output layer of the deep neural network model to each residual block and input layer using the chain rule, and the error is propagated backward. The internal correlation parameters are updated by numerical accumulation according to the preset step size parameters, and the internal correlation parameters are updated iteratively. Finally, the optimization strategy mapping logic that enables the microgrid operation performance to be continuously optimized is obtained. By introducing spatiotemporal feedback sequences and associated gradient sensitivity matrices, precise correction and efficient iteration of the scheduling strategy are achieved. Transient performance indicators are used to perform safety corrections on scheduling deviations, and a nonlinear penalty mechanism effectively suppresses the evaluation contribution of unsafe adjustment actions. Combined with the associated gradient sensitivity matrix, weight reshaping and direction correction of the gradient flow are performed, ensuring that the feedback iteration of internal associated parameters always points to the optimal direction that maximizes scheduling gain and minimizes default risk. This design not only significantly improves the convergence stability of the strategy mapping logic but also achieves continuous optimization of microgrid operating performance under a dynamic safety closed loop while ensuring adaptive energy flow control.

[0067] Methods for rolling adaptive control of energy flow in continuous scheduling cycles include: At the start of each control step, acquire the latest multi-source heterogeneous data and update the microgrid operating status data for the current step. Input the microgrid operating status data into the optimization strategy mapping logic to deduce and generate the preliminary power allocation command for the current step.

[0068] Specifically, at the beginning of each control step, the latest multi-source heterogeneous data is collected using a power quality monitor and an energy management unit. The microgrid operating status data for the current step is updated using a normalization operator and a sliding window operator. The microgrid operating status data is then input into the optimization strategy mapping logic. The residual block is used to perform residual mapping processing, and feature deduction is performed in conjunction with a fully connected layer and a nonlinear mapping function. The action probability distribution is output, and the physical output value is extracted in the action decision space to generate the preliminary power allocation command for the current step.

[0069] The initial power allocation command is mapped to the synchronously refreshed dynamic physical security constraint set using the projection operator, and the optimal scheduling action for the current step size is determined and issued for execution.

[0070] Specifically, the power flow calculation method and interior point method are used to perform a feasibility traversal of the random disturbance set of the current control step, and the dynamic physical safety constraint set is generated synchronously. The nominal power value in the preliminary power allocation instruction is extracted according to the splicing order to construct the power vector to be checked, and it is determined whether the power vector to be checked is inside the geometric boundary limit. If the limit is exceeded, the projection operator is used to calculate the minimum value of the L2 norm between the power vector to be checked and the geometric boundary limit, and the power vector to be checked is mapped to the nearest boundary point on the geometric boundary limit to eliminate the limit-exceeding component. The optimal scheduling action for the current step is determined and issued for execution.

[0071] After the current step size control is completed, the time window is slid forward by one control step size, and the state update, instruction deduction and constraint correction actions are executed cyclically to achieve rolling adaptive control of energy flow in continuous scheduling cycles.

[0072] Specifically, after the current step-size command is executed and the microgrid enters a steady state, the time window is shifted backward by one control step using a sliding window operator, and the predicted data for the next time scale is loaded. Based on the shifted time window, the latest microgrid operating status data update, preliminary power allocation command deduction, and constraint correction actions based on the dynamic physical security constraint set are triggered cyclically. By repeatedly executing the logical closed loop of acquisition, decision-making, and verification on the time axis, random fluctuation interference is offset, and rolling adaptive control of energy flow for continuous scheduling cycles is achieved.

[0073] Methods for synchronously updating the dynamic physical security constraint set include: Within each control step, real-time measurements of distributed power and load demand are acquired and the historical operating samples are incrementally updated. The updated historical operating samples are then used to update the uncertainty distribution parameters and reconstruct the random disturbance set.

[0074] Specifically, at the start of each control step, the power quality monitor and energy management unit acquire real-time measurements of distributed power and load demand, and add them to the historical operating samples for incremental updates. For the updated historical operating samples, the probability density fitting is re-executed to update the uncertainty distribution parameters such as mean and variance that reflect the fluctuation characteristics. Using the updated uncertainty distribution parameters, equal probability interval division and hierarchical sampling are re-executed to generate multiple scenario combinations containing the latest fluctuation characteristics, and a random disturbance set covering the fluctuations of the current period is reconstructed.

[0075] The current topology connection status of the microgrid and the latest multidimensional physical operation constraints of the microgrid are monitored as optimization constraints. The worst-case scenario search and boundary optimization are performed on the random disturbance set to obtain a new envelope optimization boundary and confirm it as the dynamic physical security constraint set of the current step size, so as to complete the synchronization update.

[0076] Specifically, it involves real-time monitoring of the on / off positions of circuit breakers and disconnectors to determine the current topology connection status, and obtaining multi-dimensional physical operation constraints of the microgrid, including physical node voltage limits, branch current limits, and frequency fluctuation range limits. Using the topology connection state and the multidimensional physical operation constraints of the microgrid as hard constraint boundaries, the worst-case scenario that causes the power flow distribution to deviate from the preset equilibrium point is located in the random disturbance set using power flow calculation methods. The interior point method is used to perform a feasibility traversal on the active and reactive power of each energy unit in the worst-case scenario to search for the extreme points that satisfy the multidimensional physical operation constraints of the microgrid. A new envelope optimization boundary is obtained and confirmed as the dynamic physical safety constraint set of the current step size, and the synchronization is refreshed.

[0077] This embodiment also provides a computer device applicable to the microgrid energy optimization scheduling method based on deep reinforcement learning, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the microgrid energy optimization scheduling method based on deep reinforcement learning as proposed in the above embodiment.

[0078] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0079] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the microgrid energy optimization scheduling method based on deep reinforcement learning as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0080] In summary, this invention addresses the technical challenge of traditional AI scheduling commands easily violating physical constraints under complex random disturbances by establishing a deep reinforcement learning architecture with physical boundary awareness and dynamic constraint self-verification capabilities, thus achieving deep coupling between algorithm decision-making logic and power system safety boundaries. Furthermore, it utilizes a feedback mechanism that fuses transient and steady-state multi-dimensional features to guide the adaptive evolution of strategies, effectively overcoming the shortcomings of single-scale scheduling in complex operating conditions. While ensuring the "red line" of full-cycle physical safety operation of microgrids, this invention significantly improves the economic efficiency and dynamic response quality of energy scheduling, achieving a dual leap in system operational reliability and intelligent scheduling capabilities.

[0081] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the present invention.

Claims

1. A microgrid energy optimization scheduling method based on deep reinforcement learning, characterized in that, include: S1, acquire microgrid operating status data, and construct a dynamic physical security constraint set based on the uncertainty distribution of historical operating samples; S2 extracts multi-dimensional features from the microgrid operating status data and performs nonlinear mapping deduction. With the microgrid comprehensive operation optimization objective as the goal, it generates preliminary power allocation instructions and maps them into the dynamic physical security constraint set to obtain the optimal scheduling action. The method for generating preliminary power allocation instructions includes: Multi-level residual mapping processing is performed on the microgrid operation status data. By establishing cross-layer short-circuit connections between multiple linear processing layers, multi-dimensional correlation features in the microgrid operation status data are fused to obtain a multi-dimensional feature vector. By using a nonlinear mapping function to perform nonlinear dimensional transformation and feature inference on multidimensional feature vectors, the nonlinear fitting strength between scheduling features and decision logic is enhanced, and the action probability distribution is obtained. Multi-dimensional operational index data collected during microgrid operation are standardized and then weighted and fused using preset scheduling preference weight coefficients to construct a comprehensive microgrid operation optimization objective. Taking the comprehensive microgrid operation optimization objective as the optimization direction, gradient evolution is performed on the action probability distribution to generate preliminary power allocation instructions. S3 executes the control command corresponding to the optimal scheduling action, generates a transient response waveform, and obtains operation feedback and scheduling gain by combining the node voltage and branch power flow distribution. S4. Based on the operation feedback and scheduling gain, a scheduling gain criterion is constructed, and the internal correlation parameters in the nonlinear mapping derivation are updated iteratively according to the parameter feedback to obtain the optimization strategy mapping logic. S5 utilizes the optimization strategy mapping logic to perform rolling adaptive control of energy flow in continuous scheduling cycles and synchronously refreshes the dynamic physical safety constraint set.

2. The microgrid energy optimization scheduling method based on deep reinforcement learning as described in claim 1, characterized in that, The method for constructing a dynamic physical security constraint set includes: Collect multi-source heterogeneous data including distributed power generation, load demand, battery state of charge and tiered electricity pricing, and generate microgrid operation status data through normalization and time series feature enhancement processing. Based on historical operating samples, the probability distribution characteristics of distributed power generation output and load demand are analyzed to determine the uncertainty distribution parameters and construct the corresponding random disturbance set; The worst-case scenario search is performed on the random disturbance set, and the boundary optimization is performed in combination with the multi-dimensional physical operation constraints of the microgrid to obtain the dynamic physical security constraint set.

3. The microgrid energy optimization scheduling method based on deep reinforcement learning as described in claim 1, characterized in that, The method for obtaining the optimal scheduling action includes: Extract the nominal power values ​​of each energy unit in the preliminary power allocation command, construct the power vector to be checked, and transform the dynamic physical security constraint set into the geometric boundary constraints of the power space; Determine whether the power vector to be checked is inside the geometric boundary constraint; if the power vector to be checked is outside the geometric boundary constraint, then use the projection operator to perform spatial mapping to map the power vector to be checked to the boundary point on the geometric boundary constraint that is closest to the power vector to be checked, and obtain the corrected power vector; The corrected power vector is identified as the optimal scheduling action; if the power vector to be checked does not exceed the limit, it is identified as the optimal scheduling action.

4. The microgrid energy optimization scheduling method based on deep reinforcement learning as described in claim 1, characterized in that, The method for generating transient response waveforms includes: Based on the power settings of each energy unit in the optimal scheduling action, a standardized communication command stream is generated; Standardized communication command streams are sent to the execution terminal to drive each energy unit to adjust its output state according to a preset response slope; Monitor instantaneous fluctuation data during the physical state adjustment process within the microgrid; summarize the instantaneous fluctuation data, perform time series feature alignment, and generate transient response waveforms for the dynamic process of power balance regulation in the microgrid.

5. The microgrid energy optimization scheduling method based on deep reinforcement learning as described in claim 4, characterized in that, The methods for obtaining operational feedback and scheduling gains include: Calculate the sliding standard deviation of the transient response waveform within the sampling window; when the sliding standard deviation is continuously within the preset fluctuation threshold range, the microgrid is determined to have entered steady state, and the voltage of each physical node and the power flow distribution of each branch are extracted to generate a steady-state operation profile. Extract transient performance indicators from transient response waveforms and construct operational feedback that includes both dynamic and static features by combining steady-state operational sections; The initial economic and safety indicators of the microgrid before executing the optimal scheduling action are obtained. These indicators are compared with the real-time data of multi-dimensional indicators in the operation feedback to obtain the improvement of operation indicators. Multi-dimensional indicator weight aggregation is performed to quantitatively evaluate the degree of improvement of the microgrid operation performance by the optimal scheduling action and obtain the scheduling gain.

6. The microgrid energy optimization scheduling method based on deep reinforcement learning as described in claim 1, characterized in that, The method for constructing the scheduling gain criterion includes: The runtime feedback is time-stamped and combined with the scheduling gain to construct a spatiotemporal feedback sequence; and the deviation of the scheduling gain from the expected value of the baseline gain is determined. The deviation magnitude is corrected using transient performance indicators in the spatiotemporal feedback sequence to implement a punitive reduction when the transient performance indicators exceed a preset safety threshold. By summarizing the deviation magnitude and spatiotemporal feedback sequence after safety correction, a scheduling gain criterion is established with the evaluation orientation of maximizing scheduling gain and minimizing physical constraint default rate.

7. The microgrid energy optimization scheduling method based on deep reinforcement learning as described in claim 6, characterized in that, The method for obtaining the optimized strategy mapping logic includes: Retrieve the internal correlation parameters corresponding to each processing level in the nonlinear mapping derivation, and establish the correlation gradient sensitivity matrix of the internal correlation parameters relative to the scheduling gain criterion. By using the correlation gradient sensitivity matrix to perform weight reshaping and direction correction on the gradient flow generated by the scheduling gain criterion, the feedback iteration direction of the internal correlation parameters is determined; error backpropagation is performed according to the feedback iteration direction to realize the iterative update of the internal correlation parameters and obtain the optimization strategy mapping logic.

8. The microgrid energy optimization scheduling method based on deep reinforcement learning as described in claim 7, characterized in that, The method for rolling adaptive control of energy flow in continuous scheduling cycles includes: At the start of each control step, acquire the latest multi-source heterogeneous data and update the microgrid operating status data for the current step; input the microgrid operating status data into the optimization strategy mapping logic to deduce and generate the preliminary power allocation command for the current step. The projection operator is used to map the initial power allocation command to the synchronously refreshed dynamic physical security constraint set, determine the optimal scheduling action for the current step size, and issue it for execution. After the current step size control is completed, the time window is slid forward by one control step size, and the state update, instruction deduction and constraint correction actions are executed cyclically to achieve rolling adaptive control of energy flow in continuous scheduling cycles.

9. The microgrid energy optimization scheduling method based on deep reinforcement learning as described in claim 8, characterized in that: The method for synchronously refreshing the dynamic physical security constraint set includes: Within each control step, real-time measurements of distributed power and load demand are acquired and the historical operating samples are incrementally updated. The updated historical operating samples are then used to update the uncertainty distribution parameters and reconstruct the random disturbance set. The current topology connection status of the microgrid and the latest multidimensional physical operation constraints of the microgrid are monitored as optimization constraints. The worst-case scenario search and boundary optimization are performed on the random disturbance set to obtain a new envelope optimization boundary and confirm it as the dynamic physical security constraint set of the current step size, so as to complete the synchronization update.

Images