An Adaptive Coordinated Control System and Method for Source-Storage-Load in Islanded Microgrids
The hierarchical coordination control system solves the voltage and frequency stability problem of isolated microgrids when new energy sources and load changes or communication failures occur. It realizes multi-energy flow economic dispatch and flexible control of the training platform, and enhances the continuity of training and engineering verification.
Patent Information
- Application Number
- CN202611241263.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-17
- Publication Date
- 2026-11-17
AI Technical Summary
Existing isolated microgrid platforms struggle to maintain stable bus voltage and frequency, enable fault-tolerant succession of multiple energy storage systems, and facilitate economic dispatch of multiple energy flows when there are changes in renewable energy output, load, and available energy storage capacity, or when energy storage converters and communication links fail. Furthermore, the control logic of teaching and training platforms is rigid, making it difficult to conduct basic understanding, dynamic control observation, and engineering optimization on the same platform.
An islanded microgrid source-storage-load adaptive coordinated control system is adopted, including a new energy power generation unit, multiple energy storage converter units, load units, a common bus, a data acquisition and execution module, and a hierarchical coordinated control module. It utilizes a bottom-level adaptive fault-tolerant heterogeneous virtual synchronous generator control, a middle-level multi-energy flow adaptive robust model predictive control scheduling, and an upper-level digital twin-driven virtual-real two-way closed-loop intelligent decision-making system to achieve voltage and frequency stability and global optimization.
Coordinated control at different time scales maintains the stability and flexibility of islanded operation, reduces the risk of single-point failure in fixed master-slave structures, enhances the continuity of training content and engineering verification, and achieves stable voltage frequency and global scheduling.
Smart Images

Figure CN123121802A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent microgrid control technology, and in particular to an adaptive coordinated control system and method for source-storage-load in an islanded microgrid. Background Technology
[0002] Islanded microgrids consist of renewable energy generation units, energy storage units, and load units. After being disconnected from the main grid, they rely on internal power sources and energy storage to maintain power balance. Droop control, virtual synchronous generator control, master-slave control, model predictive control, and reinforcement learning control are each suitable for different time scales and operational objectives. Existing islanded microgrid platforms typically employ a fixed, single control architecture. When renewable energy output, load, and available energy storage capacity change, or when energy storage converters and communication links fail, it is difficult to simultaneously ensure bus voltage and frequency stability, fault-tolerant succession of multiple energy storage units, and economical dispatch of multiple energy flows.
[0003] Existing teaching and training platforms mostly employ pure simulation or fixed source-storage-load control programs. The control algorithms, parameter open ranges, and safety constraints cannot be adjusted according to the training tasks. Therefore, it is difficult to continuously conduct basic learning, dynamic control observation, engineering optimization, and algorithm development on the same platform, and it is also difficult to directly apply teaching results to engineering verification under real power conditions. Therefore, it is necessary to construct a coordinated control system capable of collaborative control at different time scales, maintaining islanded operation during communication or equipment failures, and safely switching control complexity according to the task. Summary of the Invention
[0004] To address the challenges of achieving transient stability, global optimization, and fault tolerance in existing isolated microgrids using a single control architecture, as well as the problem of rigid control logic in training platforms, this invention provides an adaptive coordinated control system and method for source-storage-load in isolated microgrids.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: an islanded microgrid source-storage-load adaptive coordinated control system, comprising a new energy power generation unit, multiple energy storage converter units, a load unit, a common bus, a data acquisition and execution module, and a hierarchical coordinated control module. The data acquisition and execution module is used to collect operating status and execute control commands. The hierarchical coordinated control module includes: The bottom layer is an adaptive fault-tolerant heterogeneous virtual synchronous generator control layer, used to allocate support weights according to the energy storage state of charge and equipment health status to maintain the voltage and frequency of the common bus; the middle layer is a multi-energy flow adaptive robust model predictive control scheduling layer, used to correct scheduling weights and constraint margins according to the new energy prediction deviation, and to perform rolling optimization of new energy power generation, energy storage power and load power; the upper layer is a digital twin driven virtual-real two-way closed-loop intelligent decision layer, used to generate control strategies in the digital twin environment, distribute the control strategies processed by safety constraints, and correct the digital twin model and control strategies according to the returned physical operation data; The rapid control prototype interface module is used to transmit state variables, constraint parameters, and control commands between each control layer and the data acquisition and execution module; the mode configuration module is used to load control algorithms, adjustable parameter ranges, and safety constraints according to the running task and user permissions; when the hierarchical coordination control module encounters communication failures in the middle or upper layers, it enables the bottom adaptive fault-tolerant heterogeneous virtual synchronous generator control layer to maintain local autonomous operation.
[0006] Furthermore, the multiple energy storage converter units are connected in parallel to the common bus in an equal manner, and are respectively configured with a virtual synchronous generator voltage support channel and a PQ power follower channel; the underlying adaptive fault-tolerant heterogeneous virtual synchronous generator control layer includes a nonlinear droop calculation unit, which is used to calculate the support weight based on the energy storage state of charge and the equipment health status; the underlying adaptive fault-tolerant heterogeneous virtual synchronous generator control layer switches or combines the outputs of the two channels according to the support requirements and the equipment health status, and resets the support weight of any energy storage converter unit to zero when it fails, disconnects from the grid, or exits, and renormalizes the support weights of the remaining energy storage converter units.
[0007] Furthermore, the nonlinear droop calculation unit calculates the discharge capability factor and charging capability factor of the i-th energy storage converter unit according to the following formulas:
[0008]
[0009] And in accordance with Calculate the charging or discharging support weights; under discharging conditions, according to Corrected active power droop coefficient; where, The energy storage state of charge of the i-th energy storage converter unit. and These are the upper and lower limits of the state of charge, respectively. The health indicator is set to 1 when the device is healthy and 0 when it is faulty or out of service. and These are the nonlinear indices for discharging and charging, respectively, and N is the number of energy storage converter units. The baseline active power droop coefficient is 0 < ε < 1; when the sum of all capability factors is zero, all support weights are reset to zero and power deficit processing is triggered.
[0010] Furthermore, the mid-level multi-energy flow adaptive robust model predictive control scheduling layer is also connected to at least one of the heat pump thermal storage unit and the electrolysis hydrogen production unit, and incorporates the new energy power generation unit, the energy storage converter unit, the connected heat pump thermal storage unit or electrolysis hydrogen production unit, rigid loads and flexible loads into a unified state-space model; the mid-level multi-energy flow adaptive robust model predictive control scheduling layer calculates the prediction deviation according to the following formula: ; And in accordance with Adjust the robust constraint weights and the renewable energy consumption weights; in, , and These represent the actual power, predicted power, and rated power of the new energy power generation unit r, respectively. As a preset positive number, This is the error gain coefficient; when the prediction deviation increases, the energy storage reserve margin and constraint penalty weight are increased, and when the prediction deviation decreases, the weight of new energy consumption is increased.
[0011] Furthermore, the upper-layer digital twin-driven virtual-real two-way closed-loop intelligent decision-making layer includes a working condition configuration library, a digital twin model, a hierarchical reinforcement learning decision-making unit, a model calibration unit, and a safety action projection unit; the hierarchical reinforcement learning decision-making unit trains the control strategy in the digital twin model, the safety action projection unit restricts the control actions to the constraints of equipment power, state of charge, voltage, frequency, and switching actions, and the model calibration unit updates the model parameters according to the error between the physical output and the digital twin output.
[0012] Furthermore, the mode configuration module sets four levels of control permissions: the first level allows fixed master-slave control and traditional droop control; the second level allows adaptive droop control and virtual synchronous generator control parameters; the third level allows model predictive control parameters and multi-energy flow constraints; and the fourth level allows reinforcement learning strategies, model predictive control models, and distributed collaborative interfaces. Each level locks the device protection parameters and performs value range, rate of change, and interlock condition verification before issuing instructions.
[0013] Furthermore, it also includes a multi-level islanding self-healing safety protection module, which includes: a simulation prediction unit for predicting voltage and frequency over-limit trends and freezing or limiting power commands; an algorithm protection unit for limiting energy storage charging and discharging power and cutting off flexible loads according to load priority; and a hardware isolation unit for isolating faulty branches according to fault signals.
[0014] Another objective of this invention is to provide an adaptive coordinated control method for source-storage-load in an islanded microgrid, the method comprising the following steps: S1. Collect the operating status of the new energy power generation unit, multiple energy storage converter units and load units, and load control algorithms and safety constraints according to the operating tasks and user permissions; S2. Allocate support weights based on the energy storage charge status and equipment health status, use the underlying adaptive fault-tolerant heterogeneous virtual synchronous generator to control and maintain the voltage and frequency of the common bus, and reallocate support weights when the energy storage converter unit fails, disconnects from the grid, or is taken out. S3. Based on the scheduling weights and constraint margins of the new energy prediction deviation correction model prediction control, perform rolling optimization of new energy power generation, energy storage power and load power. S4. Generate control strategies in the digital twin environment, distribute the control strategies processed with safety constraints, and use the returned physical operation data to correct the digital twin model and control strategies. S5. Execute control commands and monitor communication and operational status, and maintain the local autonomous operation of the underlying layer when communication between the middle or upper layers is abnormal. S6. Return the execution results and status data, and update the support weights, scheduling constraints and digital twin model parameters for subsequent control cycles.
[0015] Further, in step S2, the charging capacity factor or discharging capacity factor of the healthy energy storage converter unit is normalized to obtain the support weight, and the virtual inertia, damping coefficient and droop coefficient are corrected accordingly; in step S3, the rolling optimization problem is solved using the energy storage state of charge and the thermal storage state or hydrogen storage state of the connected equipment as state variables, and the corresponding quantities among the energy storage charging and discharging power, new energy limited power, heat pump power, electrolyzer power and flexible load shedding amount as control variables; after communication is restored and state synchronization is completed, the middle layer rolling optimization operation is smoothly restored.
[0016] Further, step S4 includes: rehearsing in the digital twin environment conditions such as sudden drop in new energy, load impact, bus voltage disturbance, branch fault, or energy storage converter unit shutdown; issuing control strategies or restricted test commands that have been verified by safety action projection and interlock; and updating the digital twin model, model predictive control prediction model, and reinforcement learning reward function based on the deviation between the returned data and the digital twin output. Step S1 restricts the loadable control algorithms and modifiable parameters according to the four-level control authority; Step S5 sequentially performs over-limit trend prediction, algorithm amplitude limiting, flexible load shedding, and fault branch hardware isolation.
[0017] By adopting the above technical solution, the present invention has the following beneficial effects: 1. By utilizing the bottom-level millisecond-level virtual synchronous generator control, the middle-level minute-level rolling optimization, and the upper-level longer-time-scale digital twin decision-making to form a hierarchical collaboration, voltage and frequency stability and global scheduling can be taken into account without relying on a single algorithm to undertake all control objectives. 2. By using equipment health indicators, nonlinear weighting of energy storage state of charge, and communication degradation logic, support tasks are reallocated when energy storage units are decommissioned or upper-layer communication is interrupted, thereby reducing the risk of single-point failure in a fixed master-slave structure. 3. By forming a two-way closed loop through virtual pre-simulation, physical execution, data feedback and model correction, the simulation model is continuously aligned with the actual equipment status, and risks are limited by safety action projection and instruction verification before the policy is issued. 4. By hierarchically opening up control algorithms and parameters, it is possible to organize basic control, dynamic observation of virtual synchronous generators, multi-energy flow engineering optimization, and intelligent algorithm development on the same real power platform, thereby enhancing the continuity between training content and engineering verification. Attached Figure Description
[0018] Figure 1 This is a schematic diagram of the overall flow of the adaptive coordinated control system for source-storage-load of an islanded microgrid provided in an embodiment of the present invention; Figure 2 A schematic diagram of a three-layer hybrid islanded collaborative control architecture provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the islanded local control principle of an adaptive fault-tolerant heterogeneous virtual synchronous generator provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the dynamic scheduling principle of multi-energy flow adaptive robust model predictive control provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of a digital twin virtual-real two-way closed-loop adaptive intelligent decision-making system provided in an embodiment of the present invention. Figure 6 This is a schematic diagram illustrating the switching of a hierarchical adaptive teaching mode provided in an embodiment of the present invention. Detailed Implementation
[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments. The following parameters are used to illustrate an implementable configuration that can be adjusted within safety boundaries according to the rated capacity of the equipment, the allowable voltage-frequency deviation, and the training task. Unless otherwise specified, the active power of the energy storage converter is positive when discharging to the common bus, and the power generation of new energy sources and the load power both adopt the same per-unit standard.
[0020] In this embodiment, the Rapid Control Prototype (RCP) is used to realize the real-time operation of the control algorithm and data exchange with physical objects; the Virtual Synchronous Generator (VSG) is used to form voltage and frequency support; Model Predictive Control (MPC) is used for rolling scheduling; Deep Reinforcement Learning (DRL) is used for upper-level policy training; the Power Storage Converter System (PCS) is referred to as the Power Storage Converter Unit in this document; and PQ control is power following control that specifies active and reactive power reference values. All the above terms and abbreviations are used throughout this document.
[0021] like Figure 1 As shown, this invention provides an adaptive coordinated control system for source-storage-load in an islanded microgrid. The system consists of a digital twin simulation environment, an upper-level RCP controller, a middle-level RCP controller, a lower-level RCP controller, and actual islanded microgrid hardware objects. The digital twin simulation environment establishes a control model corresponding to the physical topology. The upper-level RCP controller forms a global strategy, the middle-level RCP controller forms rolling power commands, and the lower-level RCP controller forms converter modulation reference values. The physical operating status is sequentially fed back for model parameter correction and strategy iteration. Figure 1 The simulation prediction protection, algorithm adaptive amplitude limiting, and hardware segmentation isolation shown at the bottom correspond to three levels of security protection.
[0022] The hardware components of a practical islanded microgrid include at least photovoltaic or wind power generation units, two or more energy storage converter units, AC or DC loads, a common bus, measurement devices, switching devices, and a communication network. Optionally, heat pump thermal storage units and electrolysis hydrogen production units can be connected to form an electric-thermal-hydrogen multi-energy flow. Each RCP controller is implemented by a processor with deterministic task scheduling capabilities, a real-time operating system, communication interfaces, and input / output interfaces; lower-level control tasks are executed on a millisecond-level cycle, mid-level rolling optimization is executed on a second-to-minute-level cycle, and upper-level model correction and strategy evaluation are executed on a minute-to-hour-level cycle. All time scales are timestamped and aligned; the upper-level control layer only sends reference values and constraints to adjacent lower-level control layers and does not replace the closed-loop regulation of the lower-level control layer.
[0023] like Figure 2As shown, the three-layer control architecture employs bidirectional transmission of state feedback and control commands. The bottom layer obtains voltage, frequency, active power, reactive power, energy storage state of charge, and health indicators locally from each energy storage converter unit; the middle layer receives the aggregated state from the bottom layer and new energy power predictions to solve the finite-time-domain scheduling problem; the top layer receives statistics from the middle layer and physical operation data, and performs condition simulation, model correction, and reinforcement learning strategy evaluation in a digital twin environment. When any upper-layer control layer exits, its most recent valid reference value is maintained by the adjacent lower-layer control layer within the valid time limit, while the bottom layer still uses local measurements to maintain the common bus. Therefore, inter-layer collaboration does not constitute a single point of dependence for the bottom-layer stability control.
[0024] In one embodiment, the underlying adaptive fault-tolerant heterogeneous virtual synchronous generator control is implemented in the following manner.
[0025] like Figure 3 As shown, multiple energy storage converter units are connected in parallel on an islanded three-phase common AC bus in an equal manner. Each energy storage converter unit is simultaneously configured with a virtual synchronous generator voltage support channel and a PQ power follower channel. The status acquisition unit acquires U, f, P, Q, SOC, and health indicators in each low-level control cycle; the nonlinear droop calculation unit generates support weights; the mode switching unit selects the virtual synchronous generator channel, PQ channel, or a continuous participation factor of both channels based on whether voltage source support is needed, whether the upper-level power reference is valid, and the health status of the equipment; the modulation output unit generates SPWM or SVPWM modulation references accordingly.
[0026] For the i-th energy storage converter unit, the discharge capability factor and the charging capability factor are calculated as follows:
[0027]
[0028] In the formula, This is a health indicator for the equipment; it is set to 1 when the equipment is healthy and allows participation, and to 0 when the equipment is locked, disconnected from the network due to a fault, or manually withdrawn. and The upper and lower limits of the permissible state of charge; and This is a nonlinear exponent; a value greater than or equal to 1 can enhance the weight change near the boundary of the charged state. The charge and discharge support weights are calculated as follows:
[0029] In the formula, q represents the charging or discharging state. Discharging support weights are used when the system has a power deficit, and charging support weights are used when the system has a power surplus. Normalization is performed only on healthy units where the sum of capacity factors is greater than zero; when the sum of capacity factors is zero, no division operation is performed, all support weights are reset to zero, and power deficit handling is triggered by the middle layer or the flexible load is cut off by the bottom layer according to load priority. After a unit exits, the remaining units are re-normalized in the next bottom-layer control cycle.
[0030] Taking three healthy energy storage converter units as an example, their SOCs are 0.80, 0.60, and 0.40 respectively, and their SOC_min is 0.20. It is 0.90. When the value is 2, the discharge capability factors are approximately 0.7347, 0.3265, and 0.0816, and the normalized discharge support weights are approximately 0.643, 0.286, and 0.071; after the second energy storage converter unit exits, the weights of the remaining two units are approximately 0.90 and 0.10. This calculation is updated based on the latest SOC and health indicators in each underlying control cycle.
[0031] Under discharge conditions, the support weight can be mapped to the active droop coefficient according to the following formula:
[0032] A larger support weight results in a smaller droop coefficient, allowing for more active power to be handled under the same frequency deviation. Charging states are processed using the same mapping direction for charging support weights. To balance multiple energy storage states of charge, additional weights are also added. Equalized power components in the form of, where The weighted average state of charge of the healthy energy storage unit. To balance the gain, cells above the average state of charge increase power during discharge, while cells below the average state of charge decrease power. The balancing component is limited by the rated power and state of charge boundaries.
[0033] The active frequency loop of the virtual synchronous generator is implemented using the following formula:
[0034] in, For follow Virtual inertia adjusted with health margin, For virtual damping, Angular frequency, The rated angular frequency, This is an active power reference that includes the mid-level reference value and the SOC equalization component. This represents the actual active power. The reactive voltage loop uses:
[0035] in, For voltage amplitude reference, Rated voltage, This is the reactive power droop factor. and These are the actual reactive power and the reactive power reference, respectively. This is a multi-machine voltage balancing component. , , and First, the settings are determined based on the converter's rated power, allowable frequency change rate, and closed-loop damping requirements. Then, the SOC-supported weights are adjusted within preset upper and lower limits. All adjustments are processed by a saturator and a rate limiter to avoid step changes during control mode switching.
[0036] The energy storage state of charge is updated according to the law of conservation of energy:
[0037] in, and These represent the non-negative discharge power and charging power, respectively. and These are the discharge efficiency and the charging efficiency, respectively, both greater than 0 and less than or equal to 1. Rated energy, The sampling period; at the same time, mutual exclusion constraints are used to make... and At most one of them is greater than zero. This updated formula conforms to the energy conservation law and is used in conjunction with SOC measurement correction and upper and lower limit constraints for rolling optimization and low-level control.
[0038] Equipment health indicators are determined by converter protection interlock signals, communication heartbeats, DC side voltage, temperature, and switch positions. When a fault indicator is valid, the fault unit's modulation output is first blocked and χ_i is set to zero, then the weights of the remaining units are recalculated; bus voltage and frequency are maintained by healthy units still in the virtual synchronous generator support channel. After fault recovery, once the heartbeat and electrical quantities have been continuously and stably maintained for a preset period, the system is first connected in zero-power PQ mode, and then a ramp is used to increase the participation factor to suppress circulating current during the recovery process.
[0039] In one embodiment, mid-level multi-energy flow adaptive robust model predictive control scheduling is implemented in the following manner.
[0040] like Figure 4As shown, the mid-level scheduling topology may include photovoltaic units, wind turbine units, hybrid energy storage, heat pump thermal storage, electrolytic hydrogen production, rigid electrical loads, and flexible cold, heat, and electrical loads. The data acquisition module performs time alignment on each power, energy state, and prediction data, and uses the current measurement value as the initial state for rolling optimization. The state vector x includes at least each energy storage SOC; when heat pump thermal storage and electrolytic hydrogen production are connected, it also includes the amount of heat storage and hydrogen storage. The control vector u includes the energy storage charging and discharging power, renewable energy curtailment power, heat pump power, electrolyzer power, and tiered flexible load shedding amount; the disturbance vector d includes the renewable energy prediction power and the rigid load prediction power. The discrete prediction model is expressed as:
[0041] Matrices A, B, and E can be established using the equipment energy conservation model. Dynamic coefficients that cannot be determined from nameplate parameters can be identified using step tests or historical operating data. The heat storage capacity and hydrogen storage capacity are updated discretely according to the following formulas:
[0042]
[0043] in, To store heat, The self-loss rate of thermal storage within a scheduling cycle. The coefficient of performance (COP) of a heat pump. For the heat pump's electrical power, For heating capacity; For hydrogen storage capacity, For the efficiency of the electrolytic cell, The conversion factor for converting electrical energy into hydrogen mass is... The power of the electrolytic cell. Let be the hydrogen mass flow rate. The heat storage capacity and hydrogen storage capacity are constrained by the upper and lower limits of the equipment capacity, respectively. For the set of new energy power generation units R, the prediction deviation is calculated as follows:
[0044] Prediction bias is used to adjust the robust constraint weights and the renewable energy consumption weights:
[0045] In the formula, A preset positive number with the same dimension as power is used to avoid zero denominator when renewable energy rated power data is missing. When the prediction deviation increases, the upper and lower reserve margins of energy storage are increased, the range of achievable renewable energy power is narrowed, and the penalties for violating SOC, voltage, and power boundaries are increased; when the prediction deviation decreases, the conservative margin is reduced and the renewable energy absorption weight is increased. The error gain k_e is calibrated by the historical prediction error quantiles, ensuring that the reserve margin corresponding to 95% of the historical error does not exceed the available energy storage adjustment range.
[0046] Within the prediction time domain H, the rolling optimization objective can be written as:
[0047] In the formula, This is the power imbalance quantity. To limit the power generation of new energy sources, Power cut-off for flexible loads, This represents the control increment between adjacent scheduling times. Weight , , , and All are non-negative numbers; Penalty or tightening factor used for SOC, voltage, power, and reserve margin constraints. Used for penalties related to restrictions on new energy vehicle emissions. When prediction deviation increases, the penalty is increased. and reduce This prioritizes maintaining a safety margin in scheduling; as prediction deviations decrease, it improves the absorption of renewable energy. In the training mode, adjustments to these weights are only permitted within the limits set by the instruction verifier. Each optimization cycle satisfies power balance:
[0048] Simultaneously satisfying constraints on energy storage SOC upper and lower limits and mutual exclusion of charge and discharge power, thermal and hydrogen storage capacity limits, heat pump and electrolyzer ramp rate limits, rigid load non-cutoff constraints, flexible load graded cutoff constraints, and bus power reserve constraints. Equipment safety boundaries, rigid load power supply, and protection interlocks are treated as hard constraints and are not relaxed due to solution difficulties; flexible load comfort constraints can be set with relaxation variables with high penalty coefficients. The solver executes only the first control variable of the optimal control sequence each time, and resolves with the new measurement state in the next cycle.
[0049] When the rolling optimization problem is infeasible or no feasible solution is found within the allowed solution time, only the first unexecuted control variable from the previous cycle's feasible control sequence is executed within the valid time limit. If the valid time limit is exceeded or the control variable violates the latest hard constraints, the middle layer stops updating the power reference and sends a degradation flag. The bottom layer VSG maintains the common bus based on local measurements, and if the power is still insufficient, flexible loads are cut off according to load priority. This avoids distributing infeasible solutions or expired control variables to physical objects.
[0050] The communication fault-tolerant switching unit judges the message timestamp, checksum, and heartbeat. When no valid aggregation status is received for three consecutive middle-layer scheduling cycles, the middle layer sends a degradation flag to the lower layer and stops updating the power reference. Within the validity period of the most recent valid reference, the lower layer limits and maintains the reference; after the validity period expires, it switches to local VSG autonomous control. When communication is restored and data from three consecutive valid cycles is received, the middle layer first resets the prediction model with the current measured state and completes state synchronization, then restores rolling optimization through the ramp reference to avoid reference mutations. The above-mentioned number of cycles and validity period are parameters of the embodiment and can be configured according to the communication cycle and the allowable rate of change of the equipment.
[0051] In one embodiment, the upper-layer digital twin virtual-real two-way closed-loop intelligent decision-making is achieved in the following way.
[0052] like Figure 5 As shown, the upper layer consists of a digital twin RCP simulation terminal and physical islanded microgrid hardware objects. The digital twin RCP simulation terminal includes a condition configuration library, a digital twin model, a hierarchical reinforcement learning decision unit, a model calibration unit, and a safety action projection unit. The model calibration unit includes an MPC prediction model calibration subunit and a reward function correction subunit. The safety action projection unit is used to perform safety constraint processing on the control strategies issued to the physical objects. The condition configuration library includes at least the following conditions: sudden drop in photovoltaic output, wind turbine shutdown, three-phase short circuit or branch fault, bus voltage disturbance, high-power load impact, and phased shutdown of multiple energy storage units. Each condition is first executed within the digital twin model. For the physical objects, only control strategies processed and interlocked by the safety action projection unit or restricted test commands corresponding to the condition are issued. Short-circuit conditions are implemented through a dedicated isolated fault injection module under the condition that relay protection is effective, or by injecting an equivalent signal that does not generate actual short-circuit current, without bypassing hardware protection.
[0053] The digital twin model parameter vector θ includes the equivalent line impedance, available energy storage capacity, converter efficiency, equivalent load parameters, and dynamic parameters of the heat pump or electrolyzer. Calculations are performed on the physical output and digital twin output at the same time stamp:
[0054] Model calibration uses weighted errors within a sliding time window and updates parameters according to the following formula:
[0055] Before model calibration, timestamp alignment, quality bit verification, and outlier removal are performed on the physical data, and the sensitivity matrix is checked. The condition number; online updates are not performed when data is invalid or parameters are unidentifiable. The allowed parameter set Θ is determined by the device nameplate, offline identification results, and the physically feasible range. Projection operations prevent updated parameters from violating energy conservation or exceeding device capabilities. Update step size. A fixed small step size can be used, or it can be gradually reduced as the verification error continues to decrease. If the verification window error does not decrease after the update, the parameter update is canceled and the previous version of the model is retained. This forms a feasible closed loop for back-correcting the simulation model using physical data.
[0056] The state variables of the hierarchical reinforcement learning decision-making unit include available renewable energy power, load demand, SOC of each energy storage unit, thermal and hydrogen storage status, bus voltage-frequency deviation, and equipment health indicators; the action variables include MPC weight correction, energy storage reserve ratio, flexible load level, and permissible operating condition trigger sequence. The reward function can be expressed as:
[0057] in, For energy loss or operating costs, For voltage, frequency, SOC, or device power exceeding limits, , c_l and The weights are non-negative. For the frequency deviation, voltage deviation, energy loss, or load shedding, the [number]th [weight] is [value]. The reward weight for each performance metric can be updated according to the following formula:
[0058] in, The upper horizontal line represents the mean of the physical performance index within the correction window. For the target value, For the weight update step size, Proj indicates that the weights are limited to preset upper and lower bounds. This ensures safe weights do not exceed these limits. The weights are fixed at high levels to prioritize suppressing hard constraint violations, and are not reduced using this formula. The updated reward weights are first evaluated in independent validation scenarios; they are only used as weights for the next training cycle if the cumulative reward improves and there are no hard constraint violations; otherwise, the previous version is reverted. Training is conducted in a digital twin environment, generating multiple rounds based on randomized renewable energy sources, loads, and equipment parameters, stopping when the cumulative reward converges and there are no hard constraint violations in the validation scenarios. During physical operation, the reinforcement learning output does not form the converter modulation; instead, it is first constrained by the safety action projection unit and then used as intermediate layer weights or reference values, preserving the independence of the underlying stability loop and hardware protection.
[0059] In one embodiment, the tiered teaching model and security protection are implemented in the following ways.
[0060] like Figure 6 As shown, the platform switches control complexity based on teaching tasks and authorization levels. The lower-grade cognitive mode and the middle-grade general education mode both fall under Level 1 permissions, allowing fixed master-slave control and traditional droop control, while locking device protection parameters. The middle-grade advanced practical mode corresponds to Level 2 permissions, allowing adaptive droop control and VSG parameter observation and limited modification. The upper-grade engineering development mode corresponds to Level 3 permissions, allowing MPC objective function weights, prediction time domain, and multi-energy flow constraints. The graduate research mode corresponds to Level 4 permissions, allowing DRL strategies, MPC models, distributed collaborative interfaces, and approved secondary development programs. Therefore, Figure 6 The five types of teaching scenarios shown are implemented through four levels of technical permission configuration to avoid confusion between the number of modes and the permission level.
[0061] Each permission configuration includes an algorithm package identifier, parameter whitelist, allowed rate of change, switch interlock rules, and fault injection range. After the user selects a mode, the mode configuration module first performs a configuration self-check in the digital twin environment, and then the instruction verifier completes the value range verification, rate of change verification, and interlock verification. Disabling hardware overcurrent protection, bypassing SOC boundaries, simultaneously closing mutual exclusion switches, and paralleling voltage sources under asynchronous conditions are prohibited operations and are not enabled by permission upgrades.
[0062] The three-tiered safety protection system operates in a sequence from prediction to execution. The first layer uses a digital twin model to predict future voltage and frequency changes in the time domain; when a limit is exceeded, it freezes or limits input and output power and issues a warning. The second layer uses bottom and middle-layer algorithms to monitor State of Charge (SOC) and power margin; when the SOC approaches its upper limit, it restricts charging and reduces renewable energy output; when the SOC approaches its lower limit, it restricts discharging and sequentially disconnects interruptible loads in the order of tier 3, tier 2, and tier 1, with tier 1 critical loads being retained last. The third layer uses relay protection and sectionalizing switches to perform hardware isolation based on short-circuit, overcurrent, insulation, or branch fault signals. Each layer uses independent trigger signals, with hardware protection actions having the highest priority.
[0063] Specific operational examples: Example 1: Adaptive Scheduling for Industrial or Commercial Islanding. The system collects voltage, frequency, power, environmental parameters, SOC, and equipment health status at a 10ms low-level cycle and sends them to the real-time control link and data logging link, respectively. For systems with a grid-connected test interface, the tie switch status or external islanding command serves as the primary criterion for islanding transition, while PCC voltage exceeding the rated value by ±10% or frequency exceeding the range of 49Hz to 51Hz serves as an auxiliary anomaly criterion. After the transition conditions are met, the tie switch is disconnected and its position is confirmed, and then the low-level VSG control takes over the common bus. For systems that always operate independently, the system directly enters the low-level control from the islanding startup state without performing grid-connected transition judgment.
[0064] The data fusion module calculates the total available power of new energy sources. Available power of energy storage and total load demand .when Greater than Furthermore, when the State of Charge (SOC) is below 90%, charging commands are prioritized for energy storage; once the SOC reaches 90%, the output of controllable renewable energy sources is reduced sequentially while the transferable load is increased. and When basic balance is allowed within the dead zone, energy storage maintains low-power frequency regulation standby. Less than Furthermore, when the State of Charge (SOC) is higher than 20%, energy storage fills the gap according to the discharge support weight; when the SOC reaches 20%, the third-level interruptible load is cut off first, followed by the second-level flexible load, while maintaining the first-level critical load. All thresholds are generated from the equipment's allowable SOC range and load classification table, and verified by the mode configuration module.
[0065] After receiving instructions, the execution layer completes converter power regulation and circuit breaker operation. The controller records the action timestamp, trigger conditions, control parameters, and response curves. The monitoring interface synchronously displays the industrial operation view and the training analysis view. Faults can be alarmed via dual channels of sound and light and remote communication. When the external power grid is restored, the voltage, frequency, and phase angle synchronization conditions are checked first, and then the tie switch is closed; if the system is only used for islanded operation, this step is not performed. After the operation is completed, an operating condition report containing curves, parameters, scheduling logic, and protection actions is automatically generated.
[0066] Example 2: Lightweight teaching platform from 0.5kW to 3kW. The hardware employs a modular photovoltaic simulation source or small wind source, two or more energy storage converters, and tiered loads. First-level or second-level permissions are pre-configured; the bottom layer executes traditional droop or adaptive VSG control, the middle layer uses MPC with a shortened prediction time domain and locks power and SOC boundaries, and the upper layer provides virtual-real linkage visualization. Operators can conduct topology construction, islanding startup, load switching, power sharing, and single energy storage deactivation training. The system automatically blocks operations that violate interlocks and rated power.
[0067] Example 3: 3kW to 6kW Integrated Engineering Training Platform. Hardware configuration includes wind, solar, and energy storage modules; heat pump thermal storage; and electrolysis hydrogen production modules. The middle layer fully operates an electro-thermal-hydrogen multi-energy flow (MPC) system. The upper layer allows configuration of new energy surges, load surges, and multiple energy storage shutdown scenarios within a digital twin environment. Third-level users can adjust the objective function weights and security constraints within permissible limits; fourth-level users can train DRL strategies and modify MPC prediction models, but strategy deployment still requires security action projection and command verification. In the event of a communication anomaly, the system automatically degrades to autonomous operation at the lower-level VSG, fully saving waveforms and parameters for course design and engineering verification.
[0068] The system of this invention can be configured with annotated control programs, training manuals, and online resources, and sets up three levels of accounts: general operation, engineering debugging, and system management. Account permissions are used to access control programs and runtime data, but do not replace the aforementioned four-level control permission configuration; the same account can only load control permissions authorized by the task. Runtime data is separated into industrial control links and teaching record links, critical control links are isolated from display links, and fault information is output through local audio-visual and remote communication dual channels.
[0069] Deployment can be executed in six stages: requirements analysis and topology confirmation, equipment parameter input, digital twin modeling, RCP code generation and closed-loop testing, physical integration and debugging, and safety acceptance. During continuous trial operation, voltage and frequency deviations, SOC boundary triggers, communication degradation times, protection actions, and model correction errors are recorded; a 30-day continuous trial operation can serve as an acceptance configuration. This process can support applications such as university-enterprise R&D, electrical training, and independent power supply for industrial parks, but the specific delivery cycle and supporting resources do not constitute a limitation on the scope of protection of this invention.
[0070] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. An islanded microgrid source-storage-load adaptive coordinated control system, comprising a new energy generation unit, multiple energy storage converter units, a load unit, a common bus, a data acquisition and execution module, and a hierarchical coordinated control module, wherein the data acquisition and execution module is used to collect operating status and execute control commands, characterized in that, The hierarchical coordination control module includes: The bottom layer is an adaptive fault-tolerant heterogeneous virtual synchronous generator control layer, used to allocate support weights according to the energy storage state of charge and equipment health status to maintain the voltage and frequency of the common bus; the middle layer is a multi-energy flow adaptive robust model predictive control scheduling layer, used to correct scheduling weights and constraint margins according to the new energy prediction deviation, and to perform rolling optimization of new energy power generation, energy storage power and load power; the upper layer is a digital twin driven virtual-real two-way closed-loop intelligent decision layer, used to generate control strategies in the digital twin environment, distribute the control strategies processed by safety constraints, and correct the digital twin model and control strategies according to the returned physical operation data; The rapid control prototype interface module is used to transmit state variables, constraint parameters, and control commands between each control layer and the data acquisition and execution module; the mode configuration module is used to load control algorithms, adjustable parameter ranges, and safety constraints according to the running task and user permissions; when the hierarchical coordination control module encounters communication failures in the middle or upper layers, it enables the bottom adaptive fault-tolerant heterogeneous virtual synchronous generator control layer to maintain local autonomous operation.
2. The islanded microgrid source-storage-load adaptive coordinated control system according to claim 1, characterized in that, The multiple energy storage converter units are connected in parallel to the common bus in an equal manner, and are respectively configured with a virtual synchronous generator voltage support channel and a PQ power follower channel; the underlying adaptive fault-tolerant heterogeneous virtual synchronous generator control layer includes a nonlinear droop calculation unit, which is used to calculate the support weight based on the energy storage state of charge and the equipment health status; the underlying adaptive fault-tolerant heterogeneous virtual synchronous generator control layer switches or combines the outputs of the two channels according to the support requirements and the equipment health status, and resets the support weight of any energy storage converter unit to zero when it fails, disconnects from the grid, or exits, and renormalizes the support weights of the remaining energy storage converter units.
3. The islanded microgrid source-storage-load adaptive coordinated control system according to claim 2, characterized in that, The nonlinear droop calculation unit calculates the discharge capability factor and charging capability factor of the i-th energy storage converter unit according to the following formulas: And in accordance with Calculate the charging or discharging support weights; According to the discharge state Corrected active power droop coefficient; where, For the i-th energy storage converter unit, the energy storage state of charge is... and These are the upper and lower limits of the state of charge, respectively. The health indicator is set to 1 when the device is healthy and 0 when it is faulty or out of service. and These are the nonlinear indices for discharging and charging, respectively, and N is the number of energy storage converter units. The baseline active power droop coefficient is 0 < ε < 1; When the sum of all capability factors is zero, all support weights are reset to zero and power deficit processing is triggered.
4. The islanded microgrid source-storage-load adaptive coordinated control system according to claim 1, characterized in that, The mid-level multi-energy flow adaptive robust model predictive control scheduling layer is also connected to at least one of the heat pump thermal storage unit and the electrolysis hydrogen production unit, and incorporates the new energy power generation unit, energy storage converter unit, the connected heat pump thermal storage unit or electrolysis hydrogen production unit, rigid load and flexible load into a unified state space model; the mid-level multi-energy flow adaptive robust model predictive control scheduling layer calculates the prediction deviation according to the following formula: ; And in accordance with Adjust the robust constraint weights and renewable energy consumption weights; in, , and These represent the actual power, predicted power, and rated power of the new energy power generation unit r, respectively. As a preset positive number, This is the error gain coefficient; when the prediction deviation increases, the energy storage reserve margin and constraint penalty weight are increased, and when the prediction deviation decreases, the weight of new energy consumption is increased.
5. The islanded microgrid source-storage-load adaptive coordinated control system according to claim 1, characterized in that, The upper-layer digital twin-driven virtual-real two-way closed-loop intelligent decision-making layer includes a working condition configuration library, a digital twin model, a hierarchical reinforcement learning decision-making unit, a model calibration unit, and a safety action projection unit. The hierarchical reinforcement learning decision-making unit trains the control strategy in the digital twin model. The safety action projection unit restricts the control actions to the constraints of equipment power, state of charge, voltage, frequency, and switching actions. The model calibration unit updates the model parameters based on the error between the physical output and the digital twin output.
6. The islanded microgrid source-storage-load adaptive coordinated control system according to claim 1, characterized in that, The mode configuration module sets four levels of control permissions: Level 1 allows fixed master-slave control and traditional droop control; Level 2 allows adaptive droop control and virtual synchronous generator control parameters; Level 3 allows model predictive control parameters and multi-energy flow constraints; Level 4 allows reinforcement learning strategies, model predictive control models, and distributed collaborative interfaces; Each level locks the device protection parameters and performs value range, rate of change, and interlock condition verification before issuing instructions.
7. The islanded microgrid source-storage-load adaptive coordinated control system according to claim 1, characterized in that, It also includes a multi-level islanding self-healing safety protection module, which includes: a simulation prediction unit for predicting voltage and frequency over-limit trends and freezing or limiting power commands; an algorithm protection unit for limiting energy storage charging and discharging power and cutting off flexible loads according to load priority; and a hardware isolation unit for isolating faulty branches according to fault signals.
8. A source-storage-load adaptive coordinated control method for an islanded microgrid, employing the source-storage-load adaptive coordinated control system for an islanded microgrid as described in any one of claims 1 to 7, characterized in that, The method includes the following steps: S1. Collect the operating status of the new energy power generation unit, multiple energy storage converter units and load units, and load control algorithms and safety constraints according to the operating tasks and user permissions; S2. Allocate support weights based on the energy storage charge status and equipment health status, use the underlying adaptive fault-tolerant heterogeneous virtual synchronous generator to control and maintain the voltage and frequency of the common bus, and reallocate support weights when the energy storage converter unit fails, disconnects from the grid, or is taken out. S3. Based on the scheduling weights and constraint margins of the new energy prediction deviation correction model prediction control, perform rolling optimization of new energy power generation, energy storage power and load power. S4. Generate control strategies in the digital twin environment, distribute the control strategies processed with safety constraints, and use the returned physical operation data to correct the digital twin model and control strategies. S5. Execute control commands and monitor communication and operational status, and maintain the local autonomous operation of the underlying layer when communication between the middle or upper layers is abnormal. S6. Return the execution results and status data, and update the support weights, scheduling constraints and digital twin model parameters for subsequent control cycles.
9. The adaptive coordinated control method for source-storage-load in an islanded microgrid according to claim 8, characterized in that, In step S2, the charging capacity factor or discharging capacity factor of the healthy energy storage converter unit is normalized to obtain the support weight, and the virtual inertia, damping coefficient and droop coefficient are corrected accordingly. In step S3, the state of charge of the energy storage and the state of thermal storage or hydrogen storage of the connected equipment are used as state variables, and the corresponding quantities among the energy storage charging and discharging power, new energy power limiting power, heat pump power, electrolyzer power and flexible load shedding amount are used as control variables to solve the rolling optimization problem; after communication is restored and state synchronization is completed, the middle layer rolling optimization operation is smoothly restored.
10. The adaptive coordinated control method for source-storage-load in an islanded microgrid according to claim 8, characterized in that, Step S4 includes: rehearsing in the digital twin environment conditions such as sudden drop in new energy, load impact, bus voltage disturbance, branch fault, or energy storage converter unit shutdown; issuing control strategies or restricted test commands that have been verified by safety action projection and interlock; and updating the digital twin model, model predictive control prediction model, and reinforcement learning reward function based on the deviation between the returned data and the digital twin output. Step S1 restricts the loadable control algorithms and modifiable parameters according to the four-level control authority; Step S5 sequentially performs over-limit trend prediction, algorithm amplitude limiting, flexible load shedding, and fault branch hardware isolation.
