Intelligent microgrid multi-modal collaborative optimization method and system

Abstract

The application discloses a smart micro-grid multi-modal collaborative optimization method and system, belonging to the field of smart grids. The method comprises the following steps: collecting electric data of a public connection point and a distributed unit, quantifying a harmonic responsibility coefficient by using an improved complex independent component analysis and a partial coherence method; constructing a dynamic price game model, and seeking an optimal incentive coefficient by using a non-cooperative game; building a multi-objective optimization model, correcting model prediction control parameters by using a digital twin training reinforcement learning intelligent agent, rolling optimization output scheduling instructions, and realizing multi-modal switching. The system comprises four layers of data acquisition, regional collaboration, cloud computing and digital twin, and intelligent optimization core. The application integrates power quality and economic dispatch, and improves the intelligent and safety of micro-grid optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smart grid and power quality analysis, in particular to a multi-modal collaborative optimization method and system for smart microgrid. BACKGROUND

[0002] With the large number of distributed power sources and loads such as photovoltaic, energy storage, and charging piles connected to the smart microgrid, the harmonic problem caused by power electronic equipment is increasingly prominent, seriously affecting the power quality and restricting the economic and reliable operation of the microgrid, which has become a difficult problem to be solved in the industry.

[0003] In the Chinese patent database, the patent with application number 202510733756.5 discloses a power grid mutual correlation harmonic control method, device, equipment and storage medium, which has the following problems in harmonic control and microgrid optimization:

[0004] First, the harmonic responsibility quantification is not comprehensive enough. The comparative document focuses on the photovoltaic subnetwork, uses wavelet packet decomposition, mutual correlation function and topological impedance matrix to calculate the harmonic responsibility proportion, but does not cover multi-element distributed units such as energy storage and charging piles. Moreover, its calculation method is difficult to update dynamically with real-time operation state of the microgrid (such as frequent load fluctuations and random start-stop of equipment), resulting in inaccurate harmonic responsibility positioning and difficulty in precise tracing.

[0005] Second, the economy and harmonic control are independent of each other. Traditional technologies rely on adding filters and adjusting equipment parameters to suppress harmonics, and the comparative document controls harmonics by optimizing the output of control equipment, but neither of them associates harmonic responsibility with economic dispatch. The comparative document constructs a double-layer game model, but only for the distribution of control equipment output, without establishing a dynamic price mechanism based on real-time harmonic responsibility. Each distributed unit lacks the economic motivation to actively reduce harmonic emission, resulting in high control cost.

[0006] Third, the optimization algorithm and multi-modal collaborative capability are not good enough. Existing technologies mostly use single algorithms such as fuzzy control and basic reinforcement learning, without fully utilizing digital twin technology to build a closed-loop optimization system of "virtual simulation-physical execution". When the microgrid switches between grid-connected economy and islanded power supply, it is difficult to balance economy, power quality and system stability, and prone to problems such as power surge and voltage fluctuation. Moreover, it is difficult to cope with uncertainties such as random changes in wind and light output and sudden increases and decreases in load, resulting in a significant reduction in the practicality of the optimization results.

[0007] Therefore, there is an urgent need for a new solution that can accurately calculate the harmonic responsibility of multi-element distributed units, reduce harmonics with dynamic pricing incentives, and realize multi-modal collaborative optimization with digital twin, to break through the current bottleneck of smart microgrid operation. SUMMARY

[0008] The application aims to provide a smart micro-grid multi-modal collaborative optimization method and system to solve the problems in the background art.

[0009] To achieve the above-mentioned purpose, the application provides the following technical scheme: a smart micro-grid multi-modal collaborative optimization method, comprising the following steps:

[0010] S1: Real-time synchronous acquisition of voltage and current data at the micro-grid public connection point and the outlet of each distributed unit;

[0011] S2: Based on the improved complex independent component analysis algorithm and the partial coherence method, the data is processed, and the real-time harmonic responsibility coefficient η of each distributed unit is calculated i ;

[0012] S3: A dynamic price game model is constructed, in which the actual settlement price of the i-th unit is associated with the basic price, the harmonic incentive coefficient and the real-time harmonic responsibility coefficient of the unit, wherein the harmonic incentive coefficient is dynamically solved by non-cooperative game;

[0013] S4: A multi-objective optimization function is established to minimize the total operating cost and total harmonic responsibility of the system, and the dynamic price game model is integrated to form an optimization scheduling model;

[0014] S5: In the digital twin environment, the optimization scheduling model is solved by using the model predictive control algorithm optimized by reinforcement learning, and the optimal scheduling instruction in the future rolling time domain is obtained;

[0015] S6: The optimal scheduling instruction is issued to each distributed unit for execution, realizing the multi-modal collaborative optimization operation of the micro-grid.

[0016] Further, the "improved complex independent component analysis algorithm and partial coherence method" in S2 specifically includes:

[0017] S21: Fast Fourier transform is performed on the collected time series data to extract each harmonic component;

[0018] S22: An observation signal model containing background harmonic voltage and all potential harmonic source currents is constructed;

[0019] S23: The improved complex independent component analysis algorithm is used to perform blind source separation on the observation signal model to decouple independent harmonic source signals;

[0020] S24: The partial coherence method is used to calculate the contribution of each independent harmonic source to the total harmonic distortion rate at the public connection point, and the accurate harmonic responsibility coefficient η of each distributed unit is obtained i .

[0021] Further, the "harmonic excitation coefficient solved by non-cooperative game dynamics" in S3 is specifically:

[0022] Each distributed unit is regarded as an independent decision-making subject, its strategy set is to adjust its own operating state to change the harmonic emission characteristics, and its revenue function is the sum of its operating cost and harmonic penalty cost;

[0023] The system operator dynamically adjusts the coefficient, iteratively simulates and calculates through the digital twin, and finally guides all unit strategies to converge to the Nash equilibrium point that minimizes the total system cost, and takes the coefficient value of this point as the optimal harmonic excitation coefficient in the current period.

[0024] Further, the "multi-objective optimization function" in S4 takes the sum of the total system operating cost and the total harmonic responsibility as the optimization objective, wherein the total system operating cost includes the electricity purchase cost, the electricity sales revenue and the equipment operation and maintenance cost, the total harmonic responsibility is the sum of the harmonic responsibility coefficients of all units, and the two are balanced through a weight coefficient.

[0025] Further, S5 specifically includes:

[0026] S51: Offline training phase: In the digital twin, a reinforcement learning agent is trained using historical data, its state space is the system history and real-time data, its action space is the instruction for adjusting the internal prediction model parameters of the model predictive control, and its reward function is the negative value of the multi-objective optimization function value;

[0027] S52: Online application phase: the reinforcement learning agent outputs actions according to the real-time system state to dynamically correct the prediction model of the model predictive control; the model predictive controller uses the corrected prediction model for rolling optimization to solve the optimization scheduling model.

[0028] Further, the "multi-modal collaborative optimization operation" in S6 means that the system automatically and smoothly switches between the grid-connected economic mode, the grid-connected low-carbon mode, the island power supply mode, the island power limiting mode and the grid support mode according to the optimization results and the grid state.

[0029] A system for performing the intelligent microgrid multi-modal collaborative optimization method, comprising:

[0030] A data acquisition and perception execution unit, including synchronous acquisition devices deployed at the point of common coupling and the outlets of each distributed unit, and intelligent terminals with a unified protocol collaboration agent built-in;

[0031] A regional collaborative control unit composed of regional controllers deployed in each region of the microgrid, for collecting local data, performing coordination control and issuing instructions;

[0032] Cloud computing and digital twin unit, including cloud server and digital twin model deployed thereon, are used for mapping physical system state and providing simulation training environment for algorithm;

[0033] Intelligent optimization core unit, including harmonic responsibility analysis module, dynamic game module and reinforcement learning-model predictive control hybrid optimization module, is used for performing optimization calculation of harmonic responsibility coefficient calculation, dynamic price game solving and optimization scheduling model solving in the method.

[0034] Further, the intelligent terminal supports plug and play function, and when any new distributed unit meeting the unified information model and communication protocol is accessed, the identity and capability parameters can be automatically registered to the system.

[0035] Further, the digital twin model keeps real-time data synchronization with the physical micro-grid, and serves as a safe sandbox for exploratory training of the reinforcement learning intelligent agent and a simulation test platform for new control strategies.

[0036] Compared with the prior art, the method has the following beneficial effects:

[0037] (1) By improving the complex independent component analysis and partial coherence method, multiple units such as photovoltaic, energy storage and charging pile are covered, and η i is updated in real time, the error is small under dynamic working conditions, and accurate harmonic tracing is realized in all scenarios.

[0038] (2) The dynamic price game model is innovated to guide the units to actively reduce harmonics by economic incentives, and the system cost is reduced by more than 15%, solving the problem of "disconnection between governance and economy".

[0039] (3) Digital twin and RL-MPC algorithm cope with uncertainty, 5 modes are smoothly switched, system availability reaches 99.98%, and stability under complex working conditions is improved. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 The method flowchart of the application;

[0041] Figure 2 The system block diagram of the application. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the application will be described clearly and completely below.

[0043] An intelligent micro-grid multi-modal collaborative optimization method, comprising the following steps:

[0044] S1. Real-time synchronous data acquisition

[0045] Synchronization acquisition devices are arranged at the micro-grid public connection point PCC and the outlets of each distributed unit to collect three-phase voltage and three-phase current data in real time. After 50 Hz notch filtering, the collected data is uploaded to the regional controller through the MQTT protocol, and the packet loss rate is less than or equal to 0.1%.

[0046] S2. Calculate real-time harmonic responsibility coefficient η i

[0047] S21. Extract harmonic components

[0048] The time series data collected by S1 is subjected to fast Fourier transform (FFT) to extract 3-25 harmonics, the sampling points are zero-padded to 1024 to improve the frequency resolution, and the harmonic voltage is calculated as formula (1):

[0049]

[0050] In the formula, U h is the hth harmonic voltage, u(n) is the voltage value of the nth sampling point, T s =1 / 2560s is the sampling interval, N=1024 is the FFT sampling point number, h is the harmonic number, f0 is the fundamental frequency of the power system, and the fundamental frequency of the power grid in China is f0=50Hz.

[0051] S22. Construct an observation signal model

[0052] The hth harmonic voltage at the PCC point and the hth harmonic current of each distributed unit extracted by S21 are used as observation variables, and the model is constructed as formula (2):

[0053] X(t)=[U PCC,h (t),I 1,h (t),I 2,h (t),…,I m,h (t)] T

[0054] In the formula, m is the number of distributed units, I i,h (t) is the hth harmonic current of the ith unit, and U PCC,h (t) is the hth harmonic voltage at the PCC point.

[0055] S23. Improve C-ICA blind source separation

[0056] ​​Discrete Fourier transform (DFT) is performed on the observation signal model x(t) constructed in S22 to obtain a frequency domain observation signal matrix X(k), where k is an iteration number index (k = 1, 2,..., 100), the dimension of X(k) is consistent with that of x(t), and X(k) represents a frequency domain observation signal set at the kth iteration. An adaptive step adjustment mechanism is introduced to optimize the C-ICA iteration process, the objective function is to maximize the signal non-Gaussianity, and the negative entropy is used for measurement, and the iteration formula is as shown in equation (3):

[0057] W(k + 1) = W(k) + μ(k) [I - E{g(Y(k)Y(k) T}]W(k)

[0058] In the formula, W is a demixing matrix, μ(k) is an adaptive step, which linearly decreases from 0.1 to 0.01 with the iteration number, Y(k) = W(k)X(k) is a separated independent harmonic source signal, g(•) is a nonlinear activation function, g(y) = tanh(ay) is adopted, a = 1.2; the error is ≤1e-6 after 100 iterations, and the independent harmonic source signal S(t) = [S1(t), S2(t),..., S m (t)]^T) is decoupled, I is a unit matrix, which is used to adjust the update direction and amplitude of the demixing matrix W, E{•} is an expectation operator, which represents the mathematical expectation of the random variable in the parentheses.

[0059] S24. Partial coherence analysis to obtain harmonic responsibility coefficient η i

[0060] The partial coherence function of the ith harmonic source S i (t) and the PCC harmonic voltage U PCC,h (t) is calculated, and other source interference is excluded, as shown in equation (4):

[0061] ,

[0062] In the formula, is the partial coherence function of the ith harmonic source and the PCC harmonic voltage U at frequency f, the value range is 0-1, the closer the value is to 1, the stronger the "net coherence" between the two (i.e., the greater the exclusive contribution of unit i to the PCC harmonic voltage U), and the closer to 0, the weaker the "net coherence", G Ui (f) is the cross power spectrum of U PCC,h (t) and S i (t), G ii (f) -1 is the matrix inverse of the self-power spectrum of the harmonic source S i (t), G iU (f) is the cross power spectrum of S i (t) and the PCC harmonic voltage U, and G Ui(f) is a reciprocal quantity. If the signal is a real signal, That is, conjugate, G UU (f) is the power spectrum of the PCC harmonic voltage U, G ii (f) is S i The self-power spectrum of (t); for Integrating over the harmonic frequency range yields the contribution of the i-th unit to the h-th harmonic, which, after normalization, becomes the real-time harmonic liability coefficient η. i ( ).

[0063] S3. Construct a dynamic electricity price game model and solve for the harmonic excitation coefficient α.

[0064] Game Theory Model Construction

[0065] Dynamic electricity price formula: P price,i =P base ±α·η i Where "+" indicates high η i The unit penalty, "-" indicates a penalty for low η. i Unit excitation; P price,i P represents the actual settlement price of the i-th distributed unit. base Let α be the time-of-use base electricity price of the power grid, such as 0.35 yuan / kWh during off-peak hours and 0.85 yuan / kWh during peak hours, and α be the harmonic excitation coefficient to be solved.

[0066] Solving for harmonic excitation coefficient α using non-cooperative game theory

[0067] Each distributed unit is an independent decision-making entity, and its strategy set is "adjusting its own operating status, such as reducing photovoltaic power and staggering charging pile peak hours". The revenue function is as shown in formula (5):

[0068] R i =-[C buy,i +C op,i +α·η i P i,use ]

[0069] In the formula, R i Let C be the profit function value of the i-th distributed unit. buy,i The electricity purchase cost for the i-th unit is given by the formula C. buy,i =P price,i ·P i,use C op,i For operation and maintenance costs (PV 0.02 yuan / kWh, energy storage 0.03 yuan / kWh), P i,use This refers to the unit's electricity consumption. The system operator aims to "minimize the total system operating cost." To target, dynamically adjust a; after each adjustment, simulate each unit strategy and system cost through digital twin, and converge to Nash equilibrium after 12 iterations, such as a = 0.45 yuan / (kWh·η), which is the current 1-hour optimal incentive coefficient.

[0070] S4. Establish a multi-objective optimization scheduling model

[0071] Multi-objective optimization function

[0072] Minimize[F cost +λ·Σ(η i )],

[0073] In the formula: F cost is the total operating cost of the system, as shown in formula (6):

[0074] ,

[0075] In the formula, C grid,buy =P base ·P buy,total , P buy,total is the total power purchase of the microgrid, C grid,sell =0.75 yuan / kWh·P sell,total , P sell,total is the total power purchase, C i =0.75 yuan / kWh·P PV,out is the total power purchase, and C bat,out =0.75 yuan / kWh·P buy,total is the total power purchase. is the total operating cost of all distributed units, with a unit of yuan.

[0076] Σ(η i ) is the sum of the harmonic responsibility coefficients of all units, λ=0.4, and after 100 simulations, the optimal weight balance of economy and power quality is verified.

[0077] Constraint conditions

[0078] Power balance: P PV,out +P bat,out +P buy,total =P load,total +P sell,total , in the formula, P PV,out is the output power of photovoltaic, from S1 data collection, P bat,out is the output power of energy storage, negative when charging, and needs to meet the safety constraint of energy storage SOC∈[20%, 90%], P buy,total is the total power purchase of the microgrid, P load,total is the total load power of the microgrid, from S1 data collection, P sell,total is the total power purchase of the microgrid, the left and right sides of the formula respectively correspond to the "power supply end" and "power consumption end" of the microgrid, and ensure that the sum of all power sources is equal to the sum of all power destinations. It is the core constraint for stable operation of the system.

[0079] Device safety: energy storage charge and discharge power ±2MW; charging pile power 0-30kW per station; PCC point THD <5% (national standard GB / T14549-1993).

[0080] S5: Solving the optimal scheduling model in the digital twin environment

[0081] Based on the model predictive control (MPC) algorithm optimized by reinforcement learning (RL), the specific steps are as follows:

[0082] S51. Offline training of RL agent

[0083] In the digital twin, using 1 year of historical data, photovoltaic irradiance, load, and harmonic data, a deep deterministic policy gradient (DDPG) agent is trained:

[0084] State space S: contains 14 types of parameters, i.e. current output and previous 1 minute output of photovoltaic, predicted output (future 15 minutes) and previous 10 minutes predicted value of photovoltaic, current value and previous 1 minute value of total load, current value and previous 5 minutes value of energy storage SOC, current value and previous 1 minute value of energy storage charge / discharge power, current value and previous 1 minute value of microgrid and main grid exchange power, current value and previous 1 minute value of PCC point voltage, current value and previous 1 minute value of PCC point current, current value and previous 5 minutes value of PCC point total harmonic distortion (THD), harmonic responsibility coefficient η of each distributed unit i average value and 15 minutes average value, current period basic electricity price and previous 1 hour basic electricity price, incentive coefficient k current value and previous 1 hour value in dynamic electricity price game model, current running mode identifier (such as 1 for grid-connected economic mode, 2 for grid-connected low-carbon mode, etc.) and previous 1 minute mode identifier, current value and 30 minutes average value of environmental temperature, each type of parameter contains "current value" and corresponding "historical characteristic value", therefore the total dimension of state space is 14x2=28.

[0085] Action space A: output 2 MPC prediction model correction parameters, A1: photovoltaic prediction confidence 0.1-1.5; A2: load prediction time constant 1-10 minutes, used for dynamic adaptation of photovoltaic output, load fluctuation, etc.

[0086] Reward function R=-[F cost +λ·Σ(η i )] consistent with the multi-objective optimization function of S4, guiding the agent to optimize in the direction of "low cost and less harmonic"; training 100,000 rounds, the agent converges to a stable strategy, providing a decision basis for online optimization.

[0087] S52. Online rolling optimization

[0088] RL agent outputs actions according to real-time system state, such as A1=0.8, A2=4.5, and corrects the MPC prediction model;

[0089] The MPC has a rolling horizon of 15 minutes and a control period of 1 minute, solves the optimization model by using an interior point method, and has a single calculation time of less than or equal to 5 seconds, and outputs optimal scheduling instructions for every 1 minute of the next 15 minutes, such as 2.8 MW of photovoltaic power and 0.5 MW of energy storage discharge.

[0090] S6. Instruction execution and multi-modal switching

[0091] Instruction issuance

[0092] The optimal scheduling instructions output by S5 are converted into device protocols by the regional collaborative control unit, photovoltaic / energy storage uses IEC61850, charging piles use Modbus, the protocol conversion logic is Modbus->JSON, and the instructions are issued to the data acquisition and sensing execution unit of S1 through Ethernet / RS485; each device responds and executes within 1 second, and the data acquisition and sensing execution unit collects actual operation data (such as actual photovoltaic output and energy storage SOC change) after execution, and feeds back to the cloud computing and digital twin unit and the regional collaborative control unit, forming a closed-loop control of "collection-analysis-optimization-execution-feedback".

[0093] Multi-modal switching

[0094] The system automatically switches the operation mode according to the PCC state (PCC voltage and fault signal collected from S1) and the optimization result of S5:

[0095] Grid-connected economic mode (PCC normal + low price valley): photovoltaic full-load and energy storage charging;

[0096] Grid-connected low-carbon mode (PCC normal + price peak + photovoltaic≥2MW): photovoltaic and energy storage selling electricity, and charging pile peak shifting;

[0097] Island power supply mode (PCC fault + SOC≥30%): disconnecting PCC, discharging energy storage, and cuttingη≥0.2 units;

[0098] Island power supply mode (PCC fault + SOC<30%): cutting unnecessary loads;

[0099] Grid support mode (PCC voltage<0.85pu): emergency discharge of energy storage to support voltage;

[0100] When switching, the power is smoothed through a first-order inertia link, and the change rate is less than or equal to 0.5 MW / s, to avoid impact.

[0101] The embodiment also provides a multi-modal collaborative optimization system of an intelligent microgrid, which is used for executing the above method, and the system structure is as follows:

[0102] 1. Data acquisition and perception execution unit

[0103] Fluke 1750 power quality analyzer at PCC for THD acquisition; EMU-800 intelligent measurement and control device at each unit outlet for voltage, current, and power acquisition;

[0104] All intelligent terminals have built-in MQTT collaboration agents, supporting "plug and play": within 10 seconds of power-on, the new unit sends a registration message containing "device type, rated parameters, and protocol version"; the regional controller compares the device whitelist and completes identity authentication within 15 seconds, without manual configuration.

[0105] 2. Regional collaborative control unit

[0106] One AR502H edge gateway from Huawei at each substation and charging station, equipped with a 4-core CPU and 16GB of memory; used for collecting local data, protocol conversion (Modbus→JSON), executing coordinated control, issuing instructions, supporting 4G / ethernet dual-link communication, and ensuring uninterrupted service in case of network interruption.

[0107] 3. Cloud computing and digital twin unit

[0108] Cloud server: Aliyun ECS, 4-core 8GB, 500GB SSD, MySQL database deployment;

[0109] Digital twin model: built based on MATLAB / Simulink, containing photovoltaic, energy storage, and load submodules with consistent parameters and physical devices; synchronized with physical data every 100ms through OPCUA protocol, with a dynamic error of ≤2%;

[0110] Security sandbox function: new strategies are first simulated in the twin body for 24 hours, including 10+ typical working conditions, and are only issued to the physical system after meeting the standards.

[0111] 4. Intelligent optimization core unit

[0112] Harmonic responsibility analysis module: based on Python SciPy / NumPy library, implementing step S2 algorithm;

[0113] Dynamic game module: based on Nashpy library, implementing step S3 game solution;

[0114] Reinforcement learning-model predictive control hybrid optimization module: based on TensorFlow (RL) and GEKKO (MPC), implementing step S5 solution;

[0115] Module interaction: the harmonic responsibility module pushes η i to other modules every 15 minutes, and the game module updates α value every 1 hour, ensuring data synchronization.

[0116] It will be apparent to those skilled in the art that the application is not limited to the details of the above-exemplified embodiments and that the present application can be implemented in other particular forms without departing from the spirit or essential characteristics of the present application. The present embodiments are, therefore, to be considered in all respects as illustrative and not restrictive, the scope of the application being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. No reference signs in the claims shall be construed as limiting the scope of the claims.

Claims

1. A multi-modal collaborative optimization method for smart microgrids, characterized in that, Includes the following steps: S1: Real-time synchronous acquisition of voltage and current data at the microgrid's common connection point and the output of each distributed unit; S2: Based on the improved complex independent component analysis algorithm and partial coherence method, the data is processed to calculate the real-time harmonic responsibility coefficient η of each of the distributed units. i ; S3: Construct a dynamic electricity price game model. In this model, the actual settlement price of the i-th unit is related to the base price, the harmonic excitation coefficient, and the real-time harmonic responsibility coefficient of the unit. The harmonic excitation coefficient is solved dynamically through non-cooperative game theory. Game theory model construction, dynamic electricity price formula: P price,i =P base ±α·η i Where "+" indicates high η i The penalty for a unit, "-" indicates a penalty for low η. i Unit excitation; P price,i P represents the actual settlement price of the i-th distributed unit. base Let α be the time-of-use base electricity price of the power grid, and α be the harmonic excitation coefficient to be solved. The "harmonic excitation coefficients are dynamically solved through non-cooperative game theory" mentioned in S3 specifically refers to: Each distributed unit is regarded as an independent decision-making entity, whose strategy set is to adjust its own operating state to change the harmonic emission characteristics, and its revenue function is the sum of its own operating cost and harmonic penalty cost. The system operator dynamically adjusts this coefficient and performs iterative simulation calculations through a digital twin to ultimately guide all unit strategies to converge to the Nash equilibrium point that minimizes the total system cost. The coefficient value at this point is then used as the optimal harmonic excitation coefficient for the current time period. S4: To minimize the total system operating cost and total harmonic liability, a multi-objective optimization function is established and integrated with the dynamic electricity price game model to form an optimized scheduling model; S5: In the digital twin environment, the optimized scheduling model is solved using a model predictive control algorithm optimized by reinforcement learning to obtain the optimal scheduling instruction in the future rolling time domain. S6: The optimal scheduling command is issued to each distributed unit for execution, so as to realize the multimodal collaborative optimization operation of the microgrid.

2. The multi-modal collaborative optimization method for smart microgrids according to claim 1, characterized in that, The "improved complex independent component analysis algorithm and partial coherence method" mentioned in S2 specifically includes: S21: Perform a fast Fourier transform on the acquired time-series data to extract each harmonic component; S22: Construct an observation signal model that includes background harmonic voltage and currents from all potential harmonic sources; S23: An improved complex independent component analysis algorithm is used to perform blind source separation on the observed signal model, decoupling the independent harmonic source signals; S24: The contribution of each independent harmonic source to the total harmonic distortion rate at the common junction point is calculated using the partial coherence method, yielding the precise harmonic liability coefficient η for each distributed unit. i .

3. The multi-modal collaborative optimization method for smart microgrids according to claim 1, characterized in that, The "multi-objective optimization function" described in S4 takes the sum of the total system operating cost and the total harmonic responsibility as the optimization objective. The total system operating cost includes the cost of purchasing electricity, the revenue from selling electricity, and the cost of equipment operation and maintenance. The total harmonic responsibility is the sum of the harmonic responsibility coefficients of all units, and the two are balanced by weighting coefficients.

4. The multi-modal collaborative optimization method for smart microgrids according to claim 1, characterized in that, S5 specifically includes: S51: Offline training phase: In the digital twin, a reinforcement learning agent is trained using historical data. Its state space is the system's historical and real-time data, its action space is the instructions for adjusting the parameters of the internal prediction model of the model prediction control, and its reward function is the negative value of the multi-objective optimization function. S52: Online application phase: The reinforcement learning agent outputs actions based on the real-time system state, dynamically correcting the prediction model of the model prediction control; the model prediction controller uses the corrected prediction model to perform rolling optimization and solve the optimized scheduling model.

5. The multi-modal collaborative optimization method for smart microgrids according to claim 1, characterized in that, The "multimodal collaborative optimization operation" mentioned in S6 refers to the system automatically and smoothly switching between grid-connected economic mode, grid-connected low-carbon mode, islanded supply guarantee mode, islanded power curtailment mode and grid support mode based on the optimization results and grid status.

6. A system for executing the multimodal collaborative optimization method for smart microgrids according to any one of claims 1-5, characterized in that, include: The data acquisition and sensing execution unit includes synchronous acquisition devices deployed at common connection points and at the exits of each distributed unit, as well as intelligent terminals with built-in unified protocol collaborative agents. The regional coordination control unit consists of regional controllers deployed in various regions of the microgrid, and is used to collect local data, perform coordinated control, and issue commands. The cloud computing and digital twin unit, including cloud servers and digital twin models deployed on them, is used to map the state of physical systems and provide a simulation training environment for algorithms; The intelligent optimization core unit includes a harmonic responsibility analysis module, a dynamic game module, and a reinforcement learning-model prediction control hybrid optimization module, which are used to perform optimization calculations in the method, such as harmonic responsibility coefficient calculation, dynamic electricity price game solution, and optimal scheduling model solution.

7. The system according to claim 6, characterized in that, The intelligent terminal supports plug-and-play functionality. When any new distributed unit that conforms to the unified information model and communication protocol is connected, it can automatically register its identity and capability parameters with the system.

8. The system according to claim 6, characterized in that, The digital twin model maintains real-time data synchronization with the physical microgrid and serves as a safe sandbox for exploratory training of the reinforcement learning agent, as well as a simulation test platform for new control strategies.

Images