Multi-micro-grid layered collaborative optimization scheduling method for park-level virtual power plant

Abstract

The invention discloses a park-level virtual power plant-oriented multi-microgrid hierarchical collaborative optimization scheduling method, and relates to the technical field of microgrid optimization scheduling, and the method comprises the steps: carrying out the multi-source photovoltaic data fusion and high-confidence power generation prediction; carrying out layered progressive load decomposition and demand baseline establishment; carrying out local multi-target self-optimization; real-time coordination of dynamic excitation is realized; micro-grid response and VPP polymerization clearing are carried out; and executing evaluation and model evolution. According to the method, the fundamental problem that a complete and feasible technical solution cannot be provided for the transverse aggregation type park under the condition of lack of physical conditions in the prior art, and efficient, reliable and sustainable resource aggregation cannot be realized is solved.

Description

A hierarchical collaborative optimization scheduling method for multi-microgrids in a park-level virtual power plant Technical Field

[0001] This invention relates to the field of microgrid optimization scheduling technology, specifically to a hierarchical collaborative optimization scheduling method for multi-microgrids oriented towards park-level virtual power plants. Background Technology

[0002] Currently, microgrids and virtual power plants (VPPs) have emerged as two key technological forms, becoming important pillars for building new power systems. A microgrid is a small-scale autonomous power generation and distribution system that achieves self-control, protection, and management by coordinating and controlling internal distributed power sources, energy storage systems, and controllable loads. It can operate in parallel with the main grid or in islanded mode during grid failures, significantly improving the power supply reliability and renewable energy absorption capacity of local areas.

[0003] A Virtual Power Plant (VPP) is a technological system that aggregates and coordinates a large number of dispersed distributed power sources, energy storage systems, and controllable loads in the power grid through advanced information and communication technologies, intelligent metering, and coordinated control software. Its core lies in the word "virtual," meaning it is not a physical power plant, but through aggregation and control, it can participate in grid operation and the electricity market as a special, equivalent "power plant," providing ancillary services such as peak shaving and valley filling, frequency regulation, and voltage regulation. It is a key enabling technology for improving power system flexibility, promoting supply and demand interaction, and facilitating the consumption of renewable energy.

[0004] In existing technologies and academic research, the most typical and advanced solution can be summarized as "a centralized-distributed two-layer optimization scheduling architecture based on model predictive control (MPC)". However, existing technologies have the following problems: (1) resource status perception is extensive, and scheduling is disconnected from physical characteristics; (2) the coordination mechanism is based on rigid instructions, which makes it difficult to stimulate voluntary response; (3) there is a lack of dynamic transformation models that connect physical status and economic incentives; (4) the decision-making algorithm is simple and it is difficult to cope with multiple complex uncertainties; (5) there is a lack of internal benefit quantification evaluation, which makes it impossible to form an optimization closed loop.

[0005] In summary, the fundamental flaw of existing centralized-distributed solutions based on model predictive control (MPC) lies in the forced coordination logic of "rigid power commands," which conflicts with the fundamental demands of each microgrid within the park as an independent entity. This results in passive, inefficient, and unsustainable coordination responses. Furthermore, its "black box aggregation" model cannot cope with the high heterogeneity of assets such as photovoltaics and energy storage, and dispatch commands often deviate from physical reality. These structural defects make it difficult for this solution to achieve efficient operation of the park's energy system when inter-plot photovoltaic power dispatch is not possible due to physical conditions. Summary of the Invention

[0006] In view of the above-mentioned shortcomings in the prior art, the present invention provides a hierarchical collaborative optimization scheduling method for multi-microgrids in a park-level virtual power plant, which solves the fundamental problem that the prior art cannot achieve efficient, reliable and sustainable resource aggregation.

[0007] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: a multi-microgrid hierarchical collaborative optimization scheduling method for park-level virtual power plants, comprising the following steps: S1: Each microgrid controller collects multi-source photovoltaic data from each photovoltaic subarray, and dynamically calculates the instantaneous confidence weight of each photovoltaic type through a multi-source data fusion confidence algorithm to generate a fused high-confidence photovoltaic total output prediction curve; S2: A three-layer decomposition architecture is used to perform refined decomposition of the park's total load, and a high-confidence load component baseline database is constructed, including building-level load files, adjustable load equipment-level lists, and charging pile cluster aggregation potential curves; S3: Based on the high-confidence photovoltaic total output prediction curve, the high-confidence load component baseline database, energy storage charge status, and electricity price signals, the high-confidence day-ahead baseline plan is obtained with the objective of minimizing the total net cost within the local microgrid scheduling cycle; S4: The VPP coordination center receives external scheduling instructions and high-confidence day-ahead baselines reported by each microgrid controller. Based on the online plan and real-time weather and electricity price data, a hybrid intelligent decision engine integrating time series forecasting, Monte Carlo simulation, optimal stopping theory, and queuing theory is used to determine the system's physical requirements and corrected risk indicators, which are then passed to the dynamic excitation signal conversion model to generate dynamic excitation signals. S5: After receiving the dynamic excitation signals, each microgrid controller initiates rapid secondary optimization based on the high-confidence day-ahead baseline plan, autonomously deciding on the available power adjustment amount and reporting it to the VPP coordination center. S6: The VPP coordination center aggregates and clears the responses of all microgrids. If the external scheduling requirements are met, the execution plan is confirmed and proceeds to step S7; otherwise, it returns to step S4 to adjust the excitation parameters and regenerate the dynamic excitation signals. S7: The VPP coordination center quantitatively evaluates the collaborative effect by comparing the actual operating data before and after collaborative scheduling, and uses the evaluation results to adjust the key parameters of the core algorithm model, completing the hierarchical collaborative optimization scheduling of multiple microgrids for park-level virtual power plants.

[0008] Furthermore, the instantaneous reliable weight The calculation formula is: in, and For reliability indicators based on device online rate and communication quality, and As an accuracy metric based on recent prediction errors, For time, This refers to the number of photovoltaic technology types.

[0009] Furthermore, the total load power of the park is defined as... This can be represented as a set of specific adjustable load powers through a three-level decomposition: in, For the first The first of the buildings A load in The state of the cut at any given moment. ∈{0,1}, This represents the characteristic power of the load under switching conditions. To decompose the total error, including unidentified loads and measurement noise, For the number of buildings, For the first The number of adjustable loads monitored within each building; the three-layer decomposition architecture includes: top-level decomposition: a non-negative matrix decomposition algorithm based on physical topology constraints, which decomposes the total load of the park into the load of each building area; middle-level decomposition: a power allocation model based on the power distribution network topology, which decomposes the total load of the building into the load of each key node; bottom-level decomposition: high-frequency voltage and current waveform data are collected through non-intrusive load monitoring devices, and after event detection, feature extraction and power decomposition, the load of key nodes is decomposed into the real-time power and switching status of specific adjustable loads.

[0010] Furthermore, the calculation formula for minimizing the total net cost within the local microgrid dispatch cycle is as follows: in, Total net cost, For electricity purchase costs, For equipment depreciation costs, For electricity sales revenue, Penalize users for their comfort. For electricity purchase price, For the power purchase capacity, For scheduling time intervals, The unit charge / discharge loss cost coefficient for energy storage. The charging power for energy storage, The discharge power of the stored energy. For electricity sales price, For electricity sales capacity, The penalty coefficient is... For users Expected charging completion time, The predicted charging completion time based on the current scheduling plan. A set of electric vehicle charging stations under scheduling. The scheduling period is [number].

[0011] Furthermore, the calculation formula for the dynamic excitation signal conversion model is as follows: in, The converted dynamic excitation signal, Based on the basic compensation unit price, and To adjust the parameters of excitation intensity and sensitivity, This is the risk aversion coefficient. For physical needs, The management parameters for the physical asset configuration and business operation objectives of a VPP coordination center in a campus are typically set during system initialization and can be fine-tuned based on operational experience. The risk premium derived from the Monte Carlo simulation.

[0012] Furthermore, the objective function of the fast quadratic optimization is to maximize the expected net benefit of the response to the excitation signal, as shown in the formula: in, For expected net income, The amount of power regulation that a microgrid can provide. This refers to the increase in internal costs resulting from the adjustment of the plan.

[0013] Furthermore, the synergistic effect is quantitatively assessed by increasing the percentage of photovoltaic power consumption, using the following formula: in, To increase the percentage of photovoltaic power consumption, To contribute to the actual grid connection of each photovoltaic array, This represents the theoretical power wasted without the coordination of the VPP (Vibration Power Producer) coordination center. To contribute to the maximum theoretical potential of photovoltaics.

[0014] Furthermore, the collaborative workflow of the four algorithms—time series prediction, Monte Carlo simulation, optimal stopping theory, and queuing theory—is as follows: time series prediction is executed first, providing the foundational power gap prediction for all subsequent algorithms. Its uncertainty range; Monte Carlo simulation receives the prediction results and uncertainty model, and generates risk quantification indicators. Optimal stopping theory and queuing theory, as two parallel resource schedulers, share the received risk indicators. The two schedulers adjust their respective scheduling strategies accordingly; the outputs and risk indicators of the two schedulers. All inputs are fed into the integrated module of the hybrid intelligent decision engine for multi-objective trade-offs, ultimately determining the system's physical requirements. and risk-corrected .

[0015] Furthermore, the percentage of photovoltaic power consumption will be increased. The parameters of the dynamic excitation signal conversion model are iteratively updated according to the following rules: in, For the updated parameters, The parameters before the update. The preset performance target value, For a small learning rate, we set it to 0.01.

[0016] The beneficial effects of the present invention are: (1) Mechanism innovation: The present invention pioneered the "dynamic excitation signal conversion model", which converts the physical demand of the system (such as power deficit) into a real-time economic excitation signal, guiding each microgrid and internal resources (such as charging piles) to respond proactively and voluntarily from their own interests.

[0017] (2) Technology Integration: Systematic Improvement of Multi-Algorithm Collaborative Decision-Making and Closed-Loop Evaluation. Existing technologies mostly employ single optimization algorithms, which are difficult to systematically address issues such as strong uncertainty in source load, equipment heterogeneity, and multi-objective conflicts. This invention constructs a hybrid intelligent decision engine that integrates multiple algorithms such as time series prediction, Monte Carlo simulation, optimal stopping theory, and queuing theory to collaboratively handle problems at different time scales and decision dimensions. At the same time, it introduces a quantitative evaluation model such as the percentage increase in photovoltaic power consumption, forming a complete closed loop of "perception-decision-incentive-execution-evaluation-optimization," which promotes the continuous self-evolution of the system and significantly improves scheduling accuracy, reliability, and long-term operational benefits. Attached Figure Description

[0018] Figure 1 is a flowchart of a hierarchical collaborative optimization scheduling method for multi-microgrids oriented towards park-level virtual power plants. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0020] As shown in Figure 1, a hierarchical collaborative optimization scheduling method for multi-microgrids oriented towards park-level virtual power plants includes the following steps: Stage 1: High confidence day-ahead autonomous goal: Each microgrid pursues its own optimal operation plan based on complete information.

[0021] S1: Each microgrid controller collects multi-source photovoltaic data from each photovoltaic subarray and dynamically calculates the instantaneous confidence weight of each photovoltaic type through a multi-source data fusion confidence algorithm to generate a high-confidence photovoltaic total output prediction curve after fusion. The multi-source photovoltaic data includes the real-time output, historical power, irradiance, backsheet temperature, and module efficiency degradation curve of each photovoltaic subarray (cadmium telluride, heterojunction, monocrystalline silicon, perovskite, etc.).

[0022] The instantaneous credible weight The calculation formula is: in, and For reliability indicators based on device online rate and communication quality, and As an accuracy metric based on recent prediction errors, For time, This refers to the number of photovoltaic technology types.

[0023] S2: A three-layer decomposition architecture is adopted to perform a fine decomposition of the total load of the park, and a high-confidence load component baseline database is constructed, which includes building-level load profiles, adjustable load equipment-level lists, and charging pile cluster aggregation potential curves. Data sources include total electricity meter data, building / floor sub-metering data, and data from AI identification devices deployed at key nodes (such as data centers and large workshops).

[0024] Define the total load power of the park as This can be represented as a set of specific adjustable load powers through a three-level decomposition: in, For the first The first of the buildings A load in The state of the cut at any given moment. ∈{0,1}, This represents the characteristic power of the load under switching conditions. To decompose the total error, including unidentified loads and measurement noise, For the number of buildings, For the first The method identifies the number of adjustable loads monitored within each building; this step is used to achieve refined identification from the total load of the park to specific adjustable loads under low cost. The method adopts a three-layer architecture combining "top-down" and "bottom-up" approaches, decomposing the data layer by layer and increasing the accuracy at each layer.

[0025] The three-layer decomposition architecture includes: Top-level decomposition: a non-negative matrix factorization (NMF) algorithm based on physical topology constraints, which decomposes the total load of the park into the load of each building area; the physical topology constraint NMF algorithm is as follows: in, ∈R T×F The data is time-series data collected from the smart meters of the main incoming line of the park. T represents the number of time points, and F represents the feature dimensions (such as active power P, reactive power Q, power factor, harmonics, etc.). The coefficient matrix, As a basis matrix, , is a coefficient.

[0026] The basis matrix W and coefficient matrix H are solved iteratively using a multiplicative update rule.

[0027] The first matrix H List That is, the first The load characteristic time-series curves of each building, after their amplitudes are calibrated, yield estimated load power values ​​for each building. .

[0028] The decomposition error of this layer is controlled within 5%.

[0029] Mid-level decomposition: A power allocation model based on the distribution network topology, which decomposes the total load of a building into the loads of each key node; Input: Sub-metering data from each key distribution node within the building. .

[0030] The power allocation model based on the distribution network topology is as follows: For the first... The total load of each building is equal to the sum of the loads of its key nodes: in, This refers to the number of key nodes monitored within the building. This is due to measurement errors and unmonitored branch loads.

[0031] Bottom-level decomposition: High-frequency voltage and current waveform data are collected through non-intrusive load monitoring devices. Through event detection, feature extraction and power decomposition, the load of key nodes is decomposed into the real-time power and switching status of specific adjustable loads.

[0032] Input: High-frequency voltage and current waveform data collected by non-intrusive load monitoring (NILM) devices installed at selected key nodes (such as electric vehicle charging areas and central air conditioning unit cabinets).

[0033] Method: Based on the refined decomposition of event detection and artificial intelligence recognition, including: (1) Event detection uses a sliding window to calculate the standard deviation of the rate of change of active power. When the adaptive threshold is exceeded When this occurs, it is marked as a switching event.

[0034] in, For at any time The calculated power change rate fluctuation index, i.e., the standard deviation of the power change rate within the sliding window, This is the function for calculating standard deviation. Total active power at the node The derivative with respect to time (i.e., the rate of change of power) reflects the instantaneous velocity of power change. This refers to the total active power measured at key power distribution nodes (such as electric vehicle charging areas and central air conditioning unit cabinets). For the current moment Centered on, with a width of 2 Symmetrical sliding time window, The width of the sliding window represents the length of the time window (the unit is the same as the sampling interval, such as seconds). For at any time The dynamically calculated adaptive event detection threshold, when > At that moment, it was determined that a load switching event had occurred. A series of calculations obtained within a historical time period (such as the past 15 minutes) The median of the values ​​represents a typical value for the level of volatility. A series of calculations obtained within the same historical time period The average value. 3 and 0.1 are empirical coefficients used to adjust the sensitivity of the threshold to fluctuation statistics.

[0035] (2) Feature extraction and load identification: For each event window, extract transient and steady-state feature vectors f and input them into the pre-trained deep learning model to identify the load type.

[0036] Model architecture: Combining one-dimensional CNN (extracting local waveform features), LSTM (capturing temporal dependencies) and attention mechanism (focusing on key segments).

[0037] Output: Identify the load type j that is being switched and its characteristic power. .

[0038] (3) Power decomposition decomposes the total power of the nodes into the sum of the power of each identified load. The state of each load is determined by solving the following optimization problem. : in, For Given decision variables, seek solutions that minimize the objective function value. This represents the total number of all adjustable load devices to be identified at this critical node. For the first An adjustable load device in The switching state variable at time t is a binary variable. For the first The characteristic active power (unit: kW) of an adjustable load device is a known quantity obtained from the load characteristic library, representing the typical power consumption of the device in the on-state (operational) state.

[0039] This problem can be solved efficiently using the Viterbi algorithm, combined with load state transition probabilities.

[0040] Output: The real-time power of each specific adjustable load (such as an air conditioner or a charging pile) under this node. And the state of casting / cutting.

[0041] Through a three-level progressive decomposition, a high-confidence load component baseline database is constructed, containing the following: building-level load profiles: obtaining the load profiles for each building / functional area ( = 1, 2, ..., The baseline load power curve of (=1,2,...,M) within the scheduling period .

[0042] Adjustable load equipment level list: For the adjustable load under each monitoring node (such as the first...) The first in the building One load, = 1, 2, ..., ), to obtain: the identified load type (e.g., charging pile, air conditioner), characteristic power Its adjustable range, and the predicted future switching state sequence .

[0043] Charging pile cluster aggregation potential curve: based on device-level inventory, through (in The total baseline power curve and its aggregation adjustment potential of the 800+ charging pile cluster (belonging to the charging pile set) are calculated and used as the key input for subsequent dynamic incentive scheduling.

[0044] The output concretizes the "load component baseline" into a structured dataset, providing clear equipment-level adjustable parameters and potential data for local optimization in step S3 and collaborative decision-making by the VPP coordination center in step S4, without introducing new symbol definitions.

[0045] S3: Based on the high-confidence photovoltaic total output prediction curve, the high-confidence load component baseline database, energy storage state of charge, and electricity price signals, the high-confidence day-ahead baseline plan is obtained with the objective of minimizing the total net cost within the local microgrid dispatch cycle; the calculation formula for minimizing the total net cost within the local microgrid dispatch cycle is as follows: in, Total net cost, For electricity purchase costs, For equipment depreciation costs, For electricity sales revenue, Penalize users for their comfort. For electricity purchase price, For the power purchase capacity, For scheduling time intervals, The unit charge / discharge loss cost coefficient for energy storage. The charging power for energy storage, The discharge power of the stored energy. For electricity sales price, For electricity sales capacity, The penalty coefficient is... For users Expected charging completion time, The predicted charging completion time based on the current scheduling plan. A set of electric vehicle charging stations under scheduling. The scheduling period is [number].

[0046] Under the premise of satisfying local electrical constraints (power balance, energy storage SOC, equipment power upper and lower limits, etc.), the above optimization problem is solved to obtain the high-confidence day-ahead baseline plan of the microgrid, which includes the energy storage charging and discharging plan for every moment in the next 24 hours. Transaction plan with the mainnet , And the default scheduling plan for flexible loads (such as charging piles).

[0047] Phase Two: Dynamic Incentives for Real-Time Collaboration Objective: The VPP Coordination Center intelligently guides each microgrid to adjust its baseline plan to achieve aggregated response.

[0048] S4: The VPP coordination center receives external dispatch instructions, high-confidence day-ahead baseline plans reported by each microgrid controller, and real-time weather and electricity price data. Through a hybrid intelligent decision engine that integrates time series forecasting, Monte Carlo simulation, optimal stopping theory, and queuing theory, it determines the system's physical demand and corrected risk indicators, and transmits them to the dynamic excitation signal conversion model to generate dynamic excitation signals. Time series forecasting: Rolling forecasts of the net load gap for the next 15-30 minutes. .

[0049] Inputs include: historical net load series ,in = tN, t-N+1, ..., t-1, where N is the historical window length; future ultra-short-term weather forecast sequence ,in Irradiance, For ambient temperature, For the prediction step size (corresponding to 15-30 minutes); the non-adjustable portion of the "high confidence day-ahead baseline plan" reported by each microgrid. .

[0050] Implementation details: A Long Short-Term Memory (LSTM) network model is used. The model input is the historical net load sequence. And meteorological sequences, output as future Net load forecast for the step And its prediction interval (such as 90% confidence interval).

[0051] Forecasted net load gap The calculation is as follows: in, For the number of microgrids, This represents the sum of the unadjustable power of all microgrids at time t+h (the determined portion in the baseline plan). This prediction provides a benchmark and range of uncertainties for subsequent risk analysis.

[0052] Output: Future Net load gap forecast at time point h=1,...,H. Confidence intervals of the predicted values. .

[0053] Monte Carlo simulation: Based on the uncertainty of photovoltaic forecasts (incorporating confidence levels) and load fluctuations, it simulates thousands of possible scenarios and assesses the probability of default risk under different response schemes.

[0054] The input includes: a photovoltaic prediction error distribution model from the multi-source confidence fusion module: for the ... Photovoltaic-like technology, its prediction error Follows a mean of 0 and a standard deviation of The distribution of .

[0055] Distribution of net load forecasting errors from time series forecasts: forecasting errors Follows a mean of 0 and a standard deviation of The distribution of .

[0056] A set of candidate response schemes Each scheme The power setting values ​​for each adjustable resource (energy storage, charging piles, etc.) are specified.

[0057] in, For the first Photovoltaic Predicted output value at any given time (point prediction). For the first In the simulated scenario, the first Photovoltaic Simulated output value at time, For the first In the first scenario, the... Photovoltaic Prediction error at time, For the first Photovoltaic The standard deviation of the prediction error at time , For the park in Net load forecast at any given time (point forecast). For the first In the simulated scenarios, the park is Simulated net load value at time 10:00 For the first In this scenario, the net load is Prediction error at time, For net load at The standard deviation of the prediction error at time , express Follows a normal distribution. express It follows a normal distribution.

[0058] Solution evaluation: For each candidate solution In each scene Next, calculate its actual available response power. This requires taking into account the physical constraints of adjustable resources (such as energy storage SOC and charging station status).

[0059] Risk Calculation: Solution Default risk probability Defined as the proportion of scenarios in which the actual response power fails to meet the target requirement across all simulated scenarios: in, For the plan The target response power at time t+h As the allowable error tolerance, This is an indicator function (1 if the condition is true, 0 otherwise). A collection of scenes.

[0060] Output: Each candidate solution Default risk probability .

[0061] Optimal stopping theory: deciding when to activate energy storage devices with different response characteristics, such as submerged energy storage and solid-state energy storage, in order to balance immediate benefits with future uncertainties ("Use it now or save it for later?").

[0062] Inputs include: various energy storage devices Real-time state: State of charge Health status , Maximum charging and discharging power, etc.; current and future incentive price forecasts. Risk probability (From Monte Carlo simulation).

[0063] Specific implementation: Establish an optimal stopping problem model for each type of energy storage device (such as power-type solid-state batteries and energy-type immersion liquid-cooled energy storage).

[0064] Define state variables: for energy storage devices Its state is .

[0065] Define the benefit of immediately invoking (stopping): if at time... When the energy storage is activated, the discharge power is (Positive value), then the immediate benefit is: in, This indicates the depreciation cost of the equipment incurred due to its use.

[0066] Define the value of continuing to wait: waiting until the next moment. The expected discounted value of acting according to the optimal strategy is .

[0067] Bellman equations: Among them, the expected term The estimation depends on the probability of risk. . The higher the value, the greater the uncertainty about the future, and the greater the uncertainty about the expected value.

[0068] The above equations can be solved using dynamic programming or approximate dynamic programming (such as value function approximation) to obtain the optimal stopping rule (i.e., the timing of the call).

[0069] Output: For each energy storage device Recommendation: Optimal call time and suggested power usage .

[0070] Queuing theory: For 800+ charging piles, model them as "service desks" and optimize the order of issuing response instructions and the intensity of incentives to minimize the dissatisfaction of car owners with waiting time.

[0071] Inputs include: real-time status of the charging pile cluster: number of vehicles waiting for service Q(t), and charging demand of each vehicle v. Waiting time Maximum acceptable delay and car owner preference tags (such as "economically sensitive"); risk probability The VPP coordination center needs to provide the total power regulation from the charging pile cluster. (Positive values ​​require a reduction in charging power).

[0072] Implementation details: The charging pile cluster is modeled as a multi-server queuing system. Among them, the number of service counters This represents the number of charging stations that can be accepted for dispatch.

[0073] Vehicle arrival follows a Poisson process, and the arrival rate λ(t) can be predicted based on historical data.

[0074] Service time (charging duration) is subject to incentive pricing. Impact: Higher prices increase the likelihood that car owners will accept slower charging (lower power, longer service time). Let the average service rate of a single service station (charging pile) be μ(t), and let μ(t) be related to... Negative correlation.

[0075] Optimization objective: To meet the power regulation requirements of the VPP coordination center. Under the constraints, minimize the average waiting time for all car owners. or maximum waiting time .in, This is a system-level requirement (constraint), which is the total task that the charging pile cluster needs to complete, derived from the global decision by the VPP coordination center. It is an input value. The device-level execution instructions (decision variables) are the specific execution plans calculated by the queuing theory model for each vehicle in order to complete the above overall task. They are the output values ​​of the model.

[0076] Scheduling strategy: based on Dynamically adjust scheduling priorities.

[0077] like High (high risk) adopts a reliability-first strategy: prioritize scheduling vehicles labeled "interruptible" or "economically sensitive", as these vehicles are more responsive to incentives and have a lower risk of default.

[0078] like Low (low risk), adopt an efficiency-first strategy: with the goal of minimizing the average waiting time, the shortest job first (SJF) or first-come first-served (FIFO) rules can be approximated.

[0079] Mathematical Model: A Markov Decision Process (MDP) can be established, with the system queue state as the state, the action as the charging power and incentive intensity allocated to each vehicle, the reward function as the negative waiting time cost, and the optimal policy can be solved through value iteration or policy iteration.

[0080] Output: Charging station scheduling sequence: Specifies which charging stations (vehicles) will be called up at what time, and the recommended charging power at the time of call. and incentive intensity .

[0081] In the above algorithms, time series prediction is executed first, providing the foundation for power gap prediction for all subsequent algorithms. Its uncertainty range; Monte Carlo simulation receives the prediction results and uncertainty model, and generates risk quantification indicators. Optimal stopping theory and queuing theory serve as two specialized resource schedulers, handling heterogeneous energy storage devices and homogeneous charging pile clusters, respectively. They share the received risk indicators. Based on this, they adjust their respective scheduling strategies (energy storage call timing, charging pile scheduling priority). The outputs of the two schedulers (energy storage call suggestions, charging pile scheduling sequences) and risk indicators... These inputs are then fed into the integrated module of the hybrid intelligent decision engine. This module performs multi-objective trade-offs to ultimately determine the system's physical requirements. and risk-corrected The data is then passed to the dynamic stimulus signal generation module. These four algorithms form a tight closed loop of "prediction-evaluation-scheduling." Prediction provides a benchmark, evaluation quantifies risk, scheduling formulates differentiated resource allocation strategies based on risk, and finally outputs a risk-adaptive collaborative solution.

[0082] The calculation formula for the dynamic excitation signal conversion model is as follows: in, The converted dynamic excitation signal, Based on the basic compensation unit price, and To adjust the parameters of excitation intensity and sensitivity, This is the risk aversion coefficient. For physical needs, The management parameters for the physical asset configuration and business operation objectives of a VPP coordination center in a campus are typically set during system initialization and can be fine-tuned based on operational experience. The risk premium derived from the Monte Carlo simulation.

[0083] S5: After receiving the dynamic excitation signal, each microgrid controller initiates rapid secondary optimization based on the high-confidence day-ahead baseline plan, autonomously decides the amount of power adjustment that can be provided and reports it to the VPP coordination center; this optimization is solved within a very short time window (seconds), with the goal of maximizing the expected net benefit that can be obtained by responding to this excitation signal while satisfying all local equipment physical and operational constraints.

[0084] The core objective is to determine the amount of power regulation that the microgrid can provide within the future timeframe covered by the excitation signal (e.g., the next 15 minutes). .definition >0 indicates that additional power is supplied to the grid (such as energy storage discharge or load reduction). <0 indicates that more power is drawn from the grid (such as energy storage charging or increased load).

[0085] The objective function of the fast quadratic optimization is to maximize the expected net benefit of responding to the excitation signal, and the formula is: in, For expected net income, The amount of power regulation that a microgrid can provide. This refers to the internal cost increments resulting from the adjustment of plans, such as the cost of additional cycle losses in energy storage, comfort penalties due to flexible load adjustments, or changes in operating costs.

[0086] Optimization must be performed strictly within the local physical boundaries, including power balance constraints.

[0087] The upper and lower limits of the state of charge (SOC) and power limits of energy storage devices.

[0088] The power adjustment range of adjustable loads (such as charging piles) is subject to user contract constraints.

[0089] Ensure that the adjusted operating plan does not jeopardize the power supply security of local critical loads.

[0090] After solving the optimization problem, the microgrid controller generates a "response curve" that includes the amount of power adjustment it can voluntarily provide at each time step. And its corresponding expected unit revenue (i.e., its net required return after assessing its internal costs). This response curve was immediately reported to the VPP coordination center.

[0091] The core of this process lies in the fact that each microgrid makes autonomous decisions entirely based on its own interests, and the incentive signals from the VPP coordination center serve only as price parameter inputs, not as mandatory commands. The microgrid, through optimization calculations, determines whether to respond, how much to respond, and at what cost to respond.

[0092] S6: The VPP coordination center aggregates and clears the responses of all microgrids. If the external scheduling requirements are met, the execution plan is confirmed and the process proceeds to step S7. If not, the process returns to step S4 to adjust the excitation parameters and regenerate the dynamic excitation signal. Stage 3: Execution Evaluation and Model Evolution S7: The VPP coordination center compares the actual operating data before and after the collaborative scheduling, quantifies the collaborative effect, and uses the evaluation results to adjust the key parameters of the core algorithm model to complete the hierarchical collaborative optimization scheduling of multi-microgrids for park-level virtual power plants.

[0093] The evaluation is performed automatically after each coordinated scheduling event. The core objective is to calculate the key indicator of "percentage increase in photovoltaic power consumption." The specific process is as follows: a. Data Collection: Within the time period T covered by the scheduling event, two sets of actual data are collected simultaneously: Actual grid-connected output of each photovoltaic array (measured by smart meters).

[0094] Actual total load of the park.

[0095] b. Baseline Scenario Simulation: Using the "High Confidence Day-ahead Baseline Plan" established by each microgrid, the theoretical operating state of the system within the same time period is simulated under the assumption of no coordinated intervention from the VPP coordination center. The core task is to calculate the theoretical curtailed power under this baseline scenario. (That is, the difference between the actual power that photovoltaics can generate and the power allowed to be fed into the grid in the baseline plan).

[0096] c. Benefit Calculation: The synergistic effect is quantitatively evaluated by the percentage increase in photovoltaic power consumption, using the following formula: in, To increase the percentage of photovoltaic power consumption, To contribute to the actual grid connection of each photovoltaic array, This represents the theoretical power wasted without the coordination of the VPP (Vibration Power Producer) coordination center. The numerator represents the additional photovoltaic power absorbed through coordinated scheduling, contributing to the maximum theoretical output of photovoltaic power.

[0097] Based on the above quantitative assessment results As a feedback signal, the parameters of the core algorithm model are adjusted in a targeted manner, forming a closed-loop optimization: if the current scheduling If the results exceed expectations, it indicates that the incentive strategy is effective and the strategy can be appropriately "strengthened"; if the results fall short of expectations, the strategy needs to be "corrected".

[0098] Example of parameter adjustment (taking a dynamic excitation model as an example): The key parameter α (excitation intensity coefficient) of the dynamic excitation signal conversion model is iteratively updated according to the following rules: in, For the updated parameters, The parameters before the update. The preset performance target value, For a small learning rate (e.g., 0.01), this formula means that when the actual absorption effect... When the response is better than the target, the system will slightly increase the future incentive intensity (α increases) to further encourage the response; otherwise, it will slightly decrease the incentive intensity to avoid wasting costs.

[0099] The adjusted parameters will take effect immediately in the next round of collaborative scheduling decisions. Through continuous "evaluation-adjustment" cycles, the system can automatically adapt to changes in external conditions and gradually approach the optimal operating strategy.

[0100] This implementation method uses a specific physical index. The system quantifies collaborative value and establishes a mathematical feedback relationship between this indicator and core economic model parameters (α, β, γ, etc.), thereby enabling the system to achieve data-driven self-evolution.

[0101] In one embodiment of the present invention, a high-tech industrial park is used as an example to verify the effectiveness of the solution.

[0102] (1) The basic information of the park is as follows: Scale: It covers an area of ​​900 mu and contains 34 plots, including 33 independently operated manufacturing / office units (microgrids MG1~MG33) and 1 centralized R&D center (microgrids MG34).

[0103] Core energy assets: Photovoltaics: Total installed capacity of 9.88MWp, with technology routes covering various modules such as monocrystalline silicon (PERC / TOPCon), heterojunction (HJT), and cadmium telluride (CdTe) thin film.

[0104] Energy storage: The total power of the energy storage systems configured in each microgrid is 8.6MW, and the total capacity is 17.17MWh.

[0105] Flexible load: The park has a total of 819 smart charging piles, distributed in various parking areas.

[0106] Physical limitations: Due to physical constraints, the thirty-four microgrids cannot achieve inter-plot photovoltaic power dispatch.

[0107] (2) Experimental design: Control group: The closest existing technology is adopted, namely the centralized-distributed optimization architecture based on model predictive control (MPC), with the central controller issuing rigid power commands.

[0108] Experimental group: The collaborative optimization system based on dynamic incentives proposed in this invention is used.

[0109] Experimental period: 30 consecutive days of operation, covering various weather conditions such as sunny and rainy days.

[0110] (3) Evaluation indicators: Photovoltaic absorption rate: the proportion of actual photovoltaic power generation that is absorbed locally.

[0111] Overall operating cost: Total electricity purchase cost of the park minus revenue from electricity sales and ancillary services.

[0112] Coordinated response reliability: The success rate of the VPP coordination center in responding to the dispatch instructions from the superior power grid.

[0113] Charging station user satisfaction: based on the average waiting time not exceeding the promised time.

[0114] (4) The experimental results are compared in Table 1: Table 1 Comparison of experimental results (5) Typical scenario effect analysis (taking a sunny day at noon as an example) Scenario: 12:00-13:00, photovoltaic power generation is high (instantaneous output exceeds 8.5MWp), the total load of the park is low, and the upper-level power grid requires the VPP coordination center to absorb the excess power.

[0115] Control group (MPC): After calculation, the central controller forces the energy storage of more than ten microgrids to charge and requires some microgrids to reduce load.

[0116] Many microgrids refused or were unable to fully execute their production plans or energy storage SOCs due to reasons such as high production schedules and high SOCs, resulting in an actual response rate of only 60% of demand and a large amount of curtailment of solar power.

[0117] Experimental Group (Invention): The VPP coordination center generates a negative incentive signal (i.e., internal preferential electricity price) based on a hybrid intelligent decision engine. = -0.12 yuan / kWh).

[0118] Each microgrid autonomously optimizes itself based on this signal: multiple manufacturing units adjust flexible production lines to increase electricity consumption; energy storage in the R&D center and some office areas starts charging; at the same time, the system guides more than 350 electric vehicles to charge at preferential rates during this period.

[0119] Results: The collaborative power consumption reached 105% of the demand, and photovoltaic curtailment was nearly zero during this period. All participating parties benefited by enjoying low-priced electricity or receiving incentive compensation.

[0120] In summary, this embodiment, based on a large-scale, complex asset configuration in a park setting, demonstrates through experimental data that, compared to the traditional rigid command model, the dynamic incentive-based collaborative optimization system proposed in this invention can effectively stimulate the participation of numerous independent microgrids. The system exhibits significant and stable advantages in several key indicators, including improving clean energy consumption, reducing overall operating costs, ensuring response reliability, and enhancing user experience, thus validating the advanced nature, scalability, and practical value of its technical solution.

[0121] Unlike the traditional "centralized optimization-command issuance" control architecture, this invention pioneers a two-layer collaborative architecture of "upper-layer dynamic incentive-lower-layer autonomous response." The upper-layer virtual power plant coordinator does not directly issue power commands, but instead generates and broadcasts dynamic economic incentive signals; based on these signals, each independent microgrid at the lower layer autonomously makes local optimization decisions and reports its response capabilities.

[0122] To address the heterogeneous characteristics of various photovoltaic technologies and energy storage methods within the industrial park, including cadmium telluride, heterojunction, and perovskite, this invention proposes a dynamic weighted multi-source data fusion algorithm. By dynamically calculating confidence weights through real-time evaluation of the reliability and prediction accuracy of each data source, a highly reliable global state awareness result is generated, laying a data foundation for precise scheduling.

[0123] The core algorithmic innovation of this invention lies in constructing a system that automatically and smoothly maps real-time physical demands and risk assessments into dynamic incentive price signals. The mathematical model typically employs nonlinear functions (such as those combined with the tanh function) to ensure smooth and bounded signals, and introduces a risk premium.

[0124] To collaboratively address complex issues such as forecasting, risk assessment, energy storage call timing, and charging pile scheduling, this invention organically integrates four algorithms—time series forecasting, Monte Carlo simulation, optimal stopping theory, and queuing theory—to form a hybrid intelligent decision engine. Each algorithm has a clearly defined role, and their outputs serve as inputs to each other, collectively generating globally optimal or near-optimal collaborative strategies.

[0125] To achieve low-cost and high-precision exploitation of the regulation potential of massive flexible loads (such as hundreds of charging piles), this invention proposes a hierarchical progressive identification method based on "total load - component loads - key nodes". At key nodes, a non-intrusive load monitoring (NILM) AI model is employed to achieve accurate decomposition of load components, providing a basis for refined and differentiated incentive scheduling of flexible loads.

[0126] This invention also proposes a quantitative evaluation model for the percentage increase in photovoltaic power consumption. The evaluation results are fed back to the parameters of the prediction model and the incentive model for self-learning, forming a complete data-driven closed loop of "strategy-execution-evaluation-optimization".

[0127] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of the invention.

Claims

1. A hierarchical collaborative optimization scheduling method for multi-microgrids in a park-level virtual power plant, characterized in that, Includes the following steps: S1: Each microgrid controller collects multi-source photovoltaic data from each photovoltaic subarray and dynamically calculates the instantaneous confidence weight of each photovoltaic type through a multi-source data fusion confidence algorithm to generate a fused high-confidence photovoltaic total output prediction curve; S2: A three-layer decomposition architecture is adopted to perform refined decomposition of the total load of the park and to build a high-confidence load component baseline database that includes building-level load files, adjustable load equipment-level lists, and charging pile cluster aggregation potential curves; S3: Based on the high-confidence photovoltaic total output prediction curve, the high-confidence load component baseline database, the energy storage state of charge and electricity price signals, the high-confidence day-ahead baseline plan is obtained with the objective of minimizing the total net cost within the local microgrid dispatch cycle; S4: The VPP coordination center receives external dispatch instructions, high-confidence day-ahead baseline plans reported by each microgrid controller, and real-time weather and electricity price data. Using a hybrid intelligent decision engine that integrates time series forecasting, Monte Carlo simulation, optimal stopping theory, and queuing theory, it determines the system's physical requirements and corrected risk indicators, and transmits them to the dynamic excitation signal conversion model to generate dynamic excitation signals. S5: After receiving the dynamic excitation signals, each microgrid controller initiates rapid secondary optimization based on the high-confidence day-ahead baseline plans, autonomously deciding on the available power adjustment amount and reporting it to the VPP coordination center. S6: The VPP coordination center aggregates and clears the responses of all microgrids. If the external dispatch requirements are met, the execution plan is confirmed and proceeds to step S7. If not, it returns to step S4 to adjust the excitation parameters and regenerate the dynamic excitation signal. S7: The VPP coordination center quantitatively evaluates the collaborative effect by comparing the actual operating data before and after collaborative dispatch, and uses the evaluation results to adjust the key parameters of the core algorithm model, completing the hierarchical collaborative optimization dispatch of multiple microgrids for a park-level virtual power plant.

2. The hierarchical collaborative optimization scheduling method for multi-microgrids oriented towards park-level virtual power plants according to claim 1, characterized in that, The instantaneous credible weight The calculation formula is: in, and For reliability indicators based on device online rate and communication quality, and As an accuracy metric based on recent prediction errors, For time, This refers to the number of photovoltaic technology types.

3. The hierarchical collaborative optimization scheduling method for multi-microgrids oriented towards park-level virtual power plants according to claim 2, characterized in that, Define the total load power of the park as This can be represented as a set of specific adjustable load powers through a three-level decomposition: in, For the first The first of the buildings A load in The state of the cut at any given moment. ∈{0,1}， This represents the characteristic power of the load under switching conditions. To decompose the total error, including unidentified loads and measurement noise, For the number of buildings, For the first The number of adjustable loads monitored within each building; the three-layer decomposition architecture includes: top-level decomposition: a non-negative matrix decomposition algorithm based on physical topology constraints, which decomposes the total load of the park into the load of each building area; middle-level decomposition: a power allocation model based on the power distribution network topology, which decomposes the total load of the building into the load of each key node; bottom-level decomposition: high-frequency voltage and current waveform data are collected through non-intrusive load monitoring devices, and after event detection, feature extraction and power decomposition, the load of key nodes is decomposed into the real-time power and switching status of specific adjustable loads.

4. The hierarchical collaborative optimization scheduling method for multi-microgrids oriented towards park-level virtual power plants according to claim 3, characterized in that, The calculation formula, which aims to minimize the total net cost within the local microgrid dispatch cycle, is as follows: in, Total net cost, For electricity purchase costs, For equipment depreciation costs, For electricity sales revenue, Penalize users for their comfort. For the electricity purchase price, For the power purchase capacity, For scheduling time intervals, The unit charge / discharge loss cost coefficient for energy storage. The charging power for energy storage, The discharge power of the stored energy. For electricity sales price, For electricity sales capacity, The penalty coefficient is... For users Expected charging completion time, The predicted charging completion time based on the current scheduling plan. A set of electric vehicle charging stations under scheduling. The scheduling period is [number].

5. The hierarchical collaborative optimization scheduling method for multi-microgrids oriented towards park-level virtual power plants according to claim 4, characterized in that, The calculation formula for the dynamic excitation signal conversion model is as follows: in, This is the converted dynamic excitation signal. Based on the basic compensation unit price, and To adjust the parameters of excitation intensity and sensitivity, This is the risk aversion coefficient. For physical needs, The management parameters for the physical asset configuration and business operation objectives of a VPP coordination center in a campus are typically set during system initialization and can be fine-tuned based on operational experience. The risk premium derived from the Monte Carlo simulation.

6. The hierarchical collaborative optimization scheduling method for multi-microgrids oriented towards park-level virtual power plants according to claim 5, characterized in that, The objective function of the fast quadratic optimization is to maximize the expected net benefit of responding to the excitation signal, and the formula is: in, For expected net income, The amount of power regulation that a microgrid can provide. This refers to the increase in internal costs resulting from the adjustment of the plan.

7. The hierarchical collaborative optimization scheduling method for multi-microgrids oriented towards park-level virtual power plants according to claim 6, characterized in that, The synergistic effect is quantitatively assessed by increasing the percentage of photovoltaic power consumption, using the following formula: in, To increase the percentage of photovoltaic power consumption, To contribute to the actual grid connection of each photovoltaic array, This represents the theoretical power wasted without the coordination of the VPP coordination center. To contribute to the maximum theoretical potential of photovoltaics.

8. The hierarchical collaborative optimization scheduling method for multi-microgrids oriented towards park-level virtual power plants according to claim 1, characterized in that, The collaborative workflow of the four algorithms—time series prediction, Monte Carlo simulation, optimal stopping theory, and queuing theory—is as follows: time series prediction is executed first, providing the foundational power gap prediction for all subsequent algorithms. Its uncertainty range; Monte Carlo simulation receives the prediction results and uncertainty model, and generates risk quantification indicators. Optimal stopping theory and queuing theory, as two parallel resource schedulers, share the received risk indicators. The two schedulers adjust their respective scheduling strategies accordingly; the outputs and risk indicators of the two schedulers. All inputs are fed into the integrated module of the hybrid intelligent decision engine for multi-objective trade-offs, ultimately determining the system's physical requirements. and risk-corrected 。 9. The hierarchical collaborative optimization scheduling method for multi-microgrids oriented towards park-level virtual power plants according to claim 7, characterized in that, Increase the percentage by utilizing photovoltaic power consumption The parameters of the dynamic excitation signal conversion model are iteratively updated according to the following rules: in, For the updated parameters, The parameters before the update. The preset performance target value, For a small learning rate, we set it to 0.01.

Images