Multi-energy complementary operation control system

Abstract

The invention discloses a multi-energy complementary operation control system, and relates to the technical field of electric power systems and automation thereof. A multi-micro-grid parallel system structure comprising various renewable energy sources such as wind power, photovoltaic energy and energy storage is introduced, and a distributed controller is configured at each micro-grid end; the graph neural network is utilized to perform unified modeling on the network topology and the real-time power operation state, and the power oscillation risk index and the target output power share correction can be output at the same time, so that the power distribution strategy is upgraded from the traditional static rule to the risk-oriented dynamic distribution. The power swing of a high-risk channel is actively weakened on the premise that the total power requirement and multi-energy complementary utilization are met; in combination with energy storage charging and discharging coordination control based on oscillation risk and power share, an energy storage unit provides effective damping support for power oscillation while ensuring a power distribution target.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power systems and their automation, in particular to a multi-energy complementary operation control system. BACKGROUND

[0002] In the new energy base and distribution network scenarios, a large number of distributed power sources such as wind power, photovoltaic power and energy storage are connected in parallel to the public bus through inverters or transformers to form a multi-energy complementary parallel system. Various types of power sources supply power to the load in the same network, and through reasonable power distribution, the renewable energy consumption level can be improved and the output and operation cost of conventional power sources can be reduced. However, with the increase of renewable energy penetration, the random fluctuation of power output and the interaction effect brought by multi-point parallel connection make it more difficult to control the power sharing and voltage and frequency stability of the parallel system.

[0003] In the process of multi-microgrid interconnection or multi-power parallel operation, local power disturbance is easy to propagate along the interconnection channel and the public bus. For example, when a wind turbine in a certain microgrid suddenly trips out or the photovoltaic output suddenly drops, a power gap will be generated on the local bus. If there is no effective power distribution and coordination control, the power disturbance will be transmitted to other nodes of the parallel system, causing power sharing imbalance among multiple power sources and system-level power oscillation, and further causing fluctuations in bus voltage and frequency. The existing control methods mostly use centralized or hierarchical control architecture, which collects the operation state of each power source through the communication network and issues unified power instructions. In the case of communication delay, data packet loss or inaccurate power source model, power adjustment lag, oscillation amplitude amplification or frequency deviation accumulation may occur, and local faults may even evolve into global instability.

[0004] In order to improve the operation performance of the multi-energy complementary system, some distributed power coordination schemes have appeared in recent years, such as using distributed optimization or predictive control methods to realize power coordination among parallel power sources through local calculation and limited information interaction, reducing the dependence on centralized control and high-speed communication. However, due to the complex dynamic coupling relationship between multi-energy systems, and the difficulty of accurately reflecting the real-time characteristics under different working conditions by the prediction model, in the case of rapid output change or large disturbance, such schemes may still have response lag and limited suppression ability to power oscillation propagation, and have not considered power distribution rationality and oscillation suppression effect from the system level. Therefore, it is necessary to provide a multi-energy complementary operation control system which can ensure reasonable power sharing of multiple power sources and effectively suppress the propagation of power oscillation in multi-microgrid or multi-power parallel system. SUMMARY

[0005] In view of the above existing problems, the present application is proposed.

[0006] This invention provides a multi-energy complementary operation control system to address the problem that local power disturbances in multi-energy complementary operation under multi-microgrid interconnection scenarios are prone to propagate into system oscillations in the network, and that existing centralized or conventional distributed control systems have insufficient oscillation suppression capabilities under conditions of communication delay and model error.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] This invention provides a multi-energy complementary operation control system, which includes:

[0009] Multiple microgrids, each microgrid including at least one renewable energy generation unit, at least one energy storage unit and at least one load unit, wherein the renewable energy generation unit and the energy storage unit are connected to the local microgrid bus via a grid-connected converter;

[0010] Interconnecting lines are used to connect the local microgrid buses of the multiple microgrids in parallel or to interconnect them via a common collection bus to form a network topology of a multi-microgrid parallel system.

[0011] A distributed controller deployed in each of the microgrids is configured to: collect the output active power, reactive power, and frequency operation of the local microgrid; process real-time power data of the network topology, including local operation and operation of adjacent microgrids coupled through the interconnecting lines, based on a graph neural network model; determine the target output power share of the local microgrid under the total power demand; generate power control commands for adjusting the output of the grid-connected converter; and realize the output power distribution among the multiple microgrids and suppress the propagation of power oscillations in the network topology.

[0012] As a preferred embodiment of the multi-energy complementary operation control system of the present invention, the renewable energy power generation unit includes at least one of wind power generation unit and photovoltaic power generation unit, the energy storage unit includes battery energy storage device, and the load unit includes at least one of industrial load and residential load.

[0013] As a preferred embodiment of the multi-energy complementary operation control system of the present invention, the graph neural network model includes at least one graph convolutional layer, which is used to extract the dynamic coupling characteristics of each microgrid and interconnection line based on the adjacency relationship and node characteristics of the network topology, and output the power oscillation risk index for each microgrid and the correction amount of the target output power share, so as to realize the collaborative optimization of power oscillation risk prediction and power allocation based on the graph convolutional layer.

[0014] In a preferred embodiment of the multi-energy complementary operation control system described in this invention, the distributed controller is further configured as follows:

[0015] Based on the power oscillation risk index and the target output power share, the charging and discharging power and timing of the energy storage unit are adjusted so that the energy storage unit can dampen power oscillations while ensuring the output power distribution target of the multi-microgrid.

[0016] As a preferred embodiment of the multi-energy complementary operation control system of the present invention, the distributed controller includes a local communication module, which is used to exchange limited state information, including local output power, frequency deviation and energy storage status, between adjacent microgrids, and adopts an event triggering mechanism to initiate the exchange of limited state information only when the local output power deviation is detected to exceed a preset threshold and / or the power oscillation risk index exceeds a risk threshold.

[0017] As a preferred embodiment of the multi-energy complementary operation control system described in this invention, the local communication module adopts an asynchronous update mechanism, allowing each microgrid to independently update its timestamp and control variables after receiving the latest status information from its neighboring microgrids.

[0018] In a preferred embodiment of the multi-energy complementary operation control system described in this invention, the distributed controller is further configured as follows:

[0019] The frequency deviation of the local microgrid is monitored in real time, and the power control command is adaptively adjusted based on the frequency deviation, so that the output power distribution can maintain the power sharing coordination between microgrids while suppressing frequency fluctuations.

[0020] As a preferred embodiment of the multi-energy complementary operation control system described in this invention, the adaptive adjustment includes:

[0021] A sliding window is used to record the frequency deviation sequence within a preset time period, and the power control parameters are iteratively optimized based on the frequency deviation sequence within the sliding window.

[0022] As a preferred embodiment of the multi-energy complementary operation control system of the present invention, the power control parameters include active power droop coefficient and / or reactive power droop coefficient and / or energy storage output power control gain, and the droop coefficient and / or energy storage output power control gain are dynamically adjusted according to the frequency deviation sequence during the iterative optimization process.

[0023] As a preferred embodiment of the multi-energy complementary operation control system described in this invention, the graph neural network model is obtained through pre-training offline. The offline training is based on historical operating data and / or simulation-generated data to construct a training set containing power oscillation samples under various operating conditions, and after the system is put into operation, it is periodically retrained or the parameters are fine-tuned based on the data collected online.

[0024] The beneficial effects of this invention are as follows: Addressing the technical problem that local power disturbances in multi-energy complementary operation scenarios involving interconnected microgrids can easily propagate into system oscillations within the network, and that existing centralized or conventional distributed control systems lack sufficient oscillation suppression capabilities under conditions of communication delay and model error, this invention proposes a multi-energy complementary operation control system based on graph neural networks and adaptive parameter tuning. By introducing a multi-microgrid parallel system structure encompassing multiple renewable energy sources such as wind power, photovoltaics, and energy storage, and configuring distributed controllers at each microgrid terminal, a graph neural network is used to uniformly model the network topology and real-time power operation status. This allows for the simultaneous output of power oscillation risk indicators and target output power share corrections, upgrading the power allocation strategy from traditional static rules to risk-oriented dynamic allocation. Under the premise of meeting total power demand and multi-energy complementary utilization, it proactively weakens power oscillations in high-risk channels. Combined with energy storage charging and discharging coordination control based on oscillation risk and power share, the energy storage units provide effective damping support for power oscillations while ensuring the power allocation target. The system employs an event-triggered and asynchronously updated distributed communication mechanism, exchanging necessary status information only when power deviation or risk levels exceed thresholds. This reduces reliance on high-bandwidth, highly synchronous communication conditions and improves robustness and feasibility under engineering constraints such as communication delays, packet loss, and clock asynchrony. Furthermore, an evaluation function is constructed using a sliding window frequency deviation sequence, and gradient-based iteration is employed to adaptively optimize power control parameters such as the droop coefficient and energy storage power control gain online. This allows parameters to dynamically adjust with system changes, automatically enhancing frequency support and oscillation suppression capabilities when frequency fluctuations intensify, and gradually returning to the baseline level when the system stabilizes. This balances stability and control smoothness during long-term operation.

[0025] In summary, this invention achieves a collaborative design of multi-energy complementary power allocation, power oscillation risk prediction and suppression, and adaptive enhancement of frequency support capability within the same framework. Compared with existing solutions, it can better suppress the propagation of power oscillations in multi-microgrid networks, improve the voltage and frequency stability and operational economy of multi-microgrid interconnection systems under conditions of high new energy penetration, and provide an operation and control system solution with intelligent decision-making capabilities and engineering feasibility for large-scale new energy bases and coordinated operation of source-grid-load-storage systems. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0027] Figure 1 This is a schematic diagram of the framework of the multi-energy complementary operation control system in the embodiment. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0029] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0030] For example, the terms “first” and “second” used in this application are only used to distinguish and describe similar objects, to differentiate the first object from another object, and are not used to describe a specific order or sequence, nor should they be interpreted as indicating or implying relative importance.

[0031] This application proposes a multi-energy complementary operation and control system, combining... Figure 1 As shown, the system includes:

[0032] Multiple microgrids, each microgrid including at least one renewable energy generation unit, at least one energy storage unit and at least one load unit, the renewable energy generation unit and the energy storage unit being connected to the local microgrid bus via a grid-connected converter;

[0033] Interconnection lines are used to connect the local microgrid buses of multiple microgrids in parallel or to interconnect them via a common collection bus to form a network topology of a multi-microgrid parallel system.

[0034] A distributed controller is deployed in each microgrid. The distributed controller is configured to: collect the output active power, reactive power and frequency of the local microgrid, and process the real-time power data of the network topology, including the local operating data and the operating data of adjacent microgrids coupled by interconnecting lines, based on a graph neural network model, determine the target output power share of the local microgrid under the total power demand, generate power control commands for adjusting the output of the grid-connected converter, and realize the output power distribution among multiple microgrids and suppress the propagation of power oscillations in the network topology;

[0035] In this embodiment, the output active and reactive power of the renewable energy generation and energy storage units of each microgrid can be obtained through digital signal processing from the voltage and current detection channels inside the grid-connected converter. The frequency operation quantity can be calculated based on the voltage phase change rate through the frequency detection link on the local microgrid bus side. The distributed controller periodically receives the above operation data through existing measurement and communication channels. To ensure sufficient time resolution under dynamic operating conditions while balancing computational and communication burdens, the sampling period for output active power, reactive power, and frequency operation quantities can be set to a default range of 20 milliseconds to 100 milliseconds. In engineering applications, the specific values ​​can be adjusted by operation and maintenance personnel according to the system inertia level and power fluctuation rate. Under typical configuration, the number of microgrids in the area can be 10 to 50. The total power demand can be given based on upper-level dispatch instructions or load forecast results. The initial target output power share can be calculated based on the installed capacity or expected economic ratio, and then refined and adjusted by the correction quantity output by the subsequent graph neural network. Optionally, when a short-term fault or communication interruption occurs in an individual microgrid measurement device, the distributed controller can temporarily maintain the target output power share of the microgrid at the effective value of the previous control cycle, and apply amplitude limiting and quality marking to the operating quantity of the microgrid, thereby avoiding affecting the overall power distribution and oscillation suppression effect by ignoring obviously abnormal data points.

[0036] In one embodiment, the renewable energy power generation unit includes at least one of wind power generation unit and photovoltaic power generation unit, the energy storage unit includes battery energy storage device, and the load unit includes at least one of industrial load and residential load;

[0037] In one embodiment, the graph neural network model includes at least one graph convolutional layer, which is used to extract the dynamic coupling characteristics of each microgrid and interconnection line based on the adjacency relationship and node characteristics of the network topology, and output the power oscillation risk index and the correction amount of the target output power share for each microgrid, so as to realize the collaborative optimization of power oscillation risk prediction and power allocation based on graph convolutional layer.

[0038] The collaborative optimization approach for power oscillation risk prediction and power allocation based on graph convolutional layers includes:

[0039] Step a, at each discrete control time , targeting In a multi-microgrid parallel system consisting of several microgrids, the distributed controller obtains operational quantities such as output active power, output reactive power, frequency deviation, and energy storage state of charge of each microgrid from the limited state information exchanged between local measurements and neighboring microgrids. These operational quantities are combined into a fixed-length node feature vector and arranged sequentially according to the microgrid index to form an input feature matrix. The input feature matrix is ​​used as the zeroth layer node in subsequent graph convolution operations to characterize the spatial distribution characteristics of the system in terms of power distribution, frequency level and energy storage state at the current moment.

[0040] Specifically, the output active power, output reactive power, frequency deviation, and energy storage state of charge contained in the node feature vector can all be obtained through existing measurements and state estimations of the local controller. The output active power and output reactive power can be the average of the most recent control cycle, the frequency deviation can be calculated as the difference between the current detection frequency and the rated frequency, and the energy storage state of charge can be provided by the battery management algorithm based on battery voltage, current, and historical charge / discharge records. To account for the contributions of physical quantities with different dimensions in the graph neural network, this embodiment can normalize the node feature vector according to its range or standard deviation, ensuring that the normalized feature values ​​fall within the range of 0-1 or -1-1. This facilitates network training convergence and reduces the excessive dominance of any single physical quantity on the training results. For example, when the number of nodes in a multi-microgrid parallel system varies between 10 and 100, the number of rows in the input feature matrix corresponds to the number of microgrids, and the number of columns can be set between 4 and 10, with the specific values ​​determined by the designer based on the types of physical quantities required. Even when the number of nodes changes but the feature dimension of a single node remains constant, the graph neural network structure remains unchanged, allowing compatibility with systems of different scales. Optionally, when a physical quantity changes very little over a long period of time and contributes little to the prediction of oscillation risk, the designer can remove the physical quantity from the feature vector while keeping the dimension of the node feature vector no less than the requirements of the target task, so as to reduce the complexity of subsequent calculations.

[0041] Step b: Based on the physical connection relationship and line impedance of the interconnecting lines in the multi-microgrid parallel system, assign weights inversely proportional to the line impedance to each pair of microgrid nodes with interconnecting lines, and construct an adjacency matrix. Building upon this, to characterize the information retained by each microgrid regarding its own state, a self-loop is introduced on the diagonal of the adjacency matrix, resulting in an adjacency matrix with self-loops. ; then according to The rows and the corresponding degree matrix are constructed. This is used to normalize the aggregation intensity of adjacent information in graph convolution updates;

[0042] For example, the impedance of interconnecting lines can be calculated from line parameters during the engineering design phase, or verified through system identification or operation records after commissioning. When constructing the adjacency matrix, the connection strength can be taken as the reciprocal of the impedance or its normalized form, so that lines with lower impedance have higher weights during graph convolution aggregation. In this embodiment, if some interconnecting line impedance parameters are incomplete, the designer can use typical values ​​or equivalent impedances estimated based on line length and voltage level as substitutes, and correct these estimated values ​​based on the measured voltage and current responses during subsequent operation. To avoid individual extremely small impedance values ​​causing excessively large values ​​during aggregation, an upper limit can be set on the weights in the adjacency matrix to ensure they do not exceed a preset maximum value. The typical upper limit can be between 1 and 10, and the specific value can be tuned according to the system scale and numerical stability requirements. Optionally, in power distribution scenarios where impedance information is difficult to obtain, a zero-one adjacency matrix based solely on topological connections can be used, and degree matrix normalization can be used to weaken the impact of node degree differences on the aggregation results to a certain extent, thereby maintaining the effectiveness of graph convolution operations even with limited parameter information.

[0043] Step c: Based on the input feature matrix and adjacency matrix obtained in steps a and b, the graph convolutional layer propagates and aggregates the node information of the multi-microgrid parallel system to simultaneously output the power oscillation risk index and the target output power share correction; this can be represented in the following unified form:

[0044]

[0045] in, and This represents the index of a microgrid node in a multi-microgrid parallel system, with a value range of [value missing]. to , This indicates the number of microgrids in a multi-microgrid parallel system. This represents the set of edges formed by pairs of microgrid nodes connected by interconnecting lines. Representing the adjacency matrix The Line number Column elements are used to describe the first The microgrid and the first Topological coupling strength between microgrids Indicates the connection of the first The microgrid and the first The impedance scalar of the interconnecting lines of a microgrid Indicates by each The adjacency matrix formed express An identity matrix of order 1. In the adjacency matrix The adjacency matrix obtained by adding self-loops to the basic structure Indicates and The corresponding degree matrix has diagonal elements as follows: The sum of elements in each row is used to normalize and weight the adjacency relationships in the graph convolution. Indicates the discrete control time index. Indicates at time The input feature matrix, formed by stacking the characteristics of each microgrid node in rows, includes operational variables such as output active power, output reactive power, frequency deviation, and energy storage state of charge. Indicates at time The initial node representation matrix is ​​fed into the graph convolutional network. Indicates at time After the first The node representation matrix after layer graph convolution operation. This represents the graph convolutional layer index, with a value range of [value range missing]. to , This indicates the total number of convolutional layers in the graph. Indicates the first The trainable weight matrix of the layer is used to complete the linear mapping after feature aggregation. This represents an element-wise nonlinear activation function, used to introduce nonlinear representation capabilities. Indicates at time go through The high-dimensional representation matrix of nodes obtained after layer graph convolution operation. Indicates at time For each microgrid's power oscillation risk index vector, each component corresponds to the power oscillation risk level of that microgrid. This means mapping the high-dimensional representation of nodes to a weight vector of power oscillation risk indicators. Indicates and Bias vector for dimension matching, Indicates at time For the correction vector of the target output power share of each microgrid, each component corresponds to a correction amount for the target output power share of a microgrid. This means mapping the high-dimensional representation of nodes to a weight vector of power share corrections. Indicates and Bias vector for dimension matching, This represents the initial target output power share vector calculated based on the total power demand and preset allocation principles. This represents the final target output power share vector after considering the power oscillation risk and power share correction, which is used to generate the power control command output by the distributed controller.

[0046] Through the above update rules, the graph convolutional network completes the spatial correlation modeling of power oscillation risk at each time step using the operating characteristics and network topology, and outputs the quantitative results that can be directly used to correct the target output power share of the local microgrid, thereby realizing the synergistic optimization of power oscillation risk prediction and power allocation.

[0047] Furthermore, the power oscillation risk index vector can be normalized to a dimensionless value between zero and one, where values ​​close to zero correspond to microgrid nodes with lower oscillation risk, and values ​​close to one correspond to microgrid nodes with higher oscillation risk. The risk threshold can be adjusted by operators based on the system's allowable power fluctuations and frequency deviation requirements, with typical values ​​between 0.6 and 0.8. The components of the target output power share vector can be constrained between 0 and 1, ensuring that the sum of all components is one. Multiplying the total power demand by the target output power share yields the target active power setpoint for each node. The correction vector can vary between negative and positive values, but its absolute value should not be too large to prevent the target share from deviating too much from the initial allocation principle. In this embodiment, a regularization term can be set during training, or the correction amount can be limited during inference to ensure that the corrected share remains within a reasonable range. The number of layers and the dimension of node representation in each layer of the graph convolutional network can be selected based on the system scale and task complexity. For example, the number of layers can be set to 2-4, and the dimension of node representation in each layer can be set to 16-128. Designers can choose a configuration that strikes a balance between prediction accuracy and computational complexity through offline verification. Alternatively, when the number of microgrids is small or the oscillation mode is relatively simple, a shallower network structure and a lower node representation dimension can be selected to reduce the hardware resource consumption of the field controller.

[0048] Specifically, the original collaborative optimization formulation is expanded here into an achievable modeling and computational process:

[0049] On the one hand, by collecting operational quantities such as active power, reactive power, frequency deviation and energy storage status from each microgrid, a unified node feature matrix is ​​constructed, enabling the system to have a learnable input representation in both the time domain and spatial structure.

[0050] On the other hand, based on the physical connection relationship and line impedance of the interconnecting lines, an adjacency matrix with self-loops and a corresponding degree matrix are formed, so that the node characteristics can reflect the electrical coupling strength when propagating on the graph. The graph convolutional network uses normalized adjacency relationships to aggregate neighborhood information, and obtains a high-dimensional node representation that combines local operating state and network structure information. Then, through the linear mapping module, the power oscillation risk index and power share correction amount are output respectively, and superimposed on the original power allocation scheme to realize risk-aware power allocation adjustment. This design enables multiple microgrids to meet the overall power demand while taking into account the oscillation risk suppression requirements, which is conducive to improving the dynamic stability and coordinated scheduling capability of multi-energy complementary operation, and provides a unified data-driven input basis for subsequent energy storage scheduling and frequency control strategies.

[0051] Similarly, the training of the graph neural network model can be completed offline before the system is put into operation. Training data can come from historical operation records, disturbance conditions generated by the simulation platform, and typical fault analyses conducted during the planning phase. The training samples should cover common power fluctuations, power flow transfers in interconnected lines, and energy storage charging and discharging switching scenarios to improve the model's ability to identify oscillation risks under different operating conditions. In this embodiment, the model can be retrained or its parameters fine-tuned at a fixed period or when the system topology changes or the installed capacity changes significantly. The retraining period can be set to several weeks to several months, determined by the operation and maintenance unit based on the frequency of system upgrades and operational stability requirements. To ensure the real-time performance of online inference, the model parameters can be fixed and downloaded to each distributed controller after training. The time overhead of a single forward computation should be less than one control cycle. On resource-constrained controllers, the computation time can be further reduced by reducing the number of network layers or using low-precision arithmetic. Optionally, when the online system detects that the model output is significantly abnormal or exceeds the expected range, it can temporarily revert to the backup control strategy based on the initial power allocation principle, while recording relevant operating data for subsequent analysis and model correction, so as to ensure that the system still has basic controllability and security in extreme cases.

[0052] In one embodiment, the distributed controller is further configured as follows:

[0053] Based on the power oscillation risk index and the target output power share, the charging and discharging power and charging and discharging sequence of the energy storage unit are adjusted so that the energy storage unit can dampen power oscillations while ensuring the output power distribution target of multiple microgrids.

[0054] In one embodiment, the distributed controller includes a local communication module, which is used to exchange limited state information, including local output power, frequency deviation and energy storage status, between adjacent microgrids, and adopts an event triggering mechanism to initiate the exchange of limited state information only when the local output power deviation is detected to exceed a preset threshold and / or the power oscillation risk indicator exceeds a risk threshold.

[0055] In one embodiment, the local communication module adopts an asynchronous update mechanism, which allows each microgrid to independently update its timestamp and control variables after receiving the latest status information from its neighboring microgrids, thereby reducing the impact of communication delay and clock asynchrony on output power distribution and oscillation suppression.

[0056] In one embodiment, the distributed controller is further configured as follows:

[0057] Real-time monitoring of local microgrid frequency deviation and adaptive adjustment of power control commands based on frequency deviation, so that output power distribution can maintain power sharing coordination between microgrids while suppressing frequency fluctuations.

[0058] In one embodiment, adaptive adjustment includes:

[0059] A sliding window is used to record the frequency deviation sequence within a preset time period. The power control parameters are iteratively optimized based on the frequency deviation sequence within the sliding window to achieve online updates of the power control parameters as the system dynamically changes.

[0060] The process of iteratively optimizing the power control parameters based on the frequency deviation sequence within the sliding window is as follows:

[0061] Step d, at the discrete control time index is At that moment, there are a total of [number] microgrids in parallel system. Each microgrid node, with its distributed controller, collects the frequency deviation of each microgrid at every moment and records the result. A microgrid at any time frequency deviation is Select window length ,exist The moment, the most recent The frequency deviation samples of all microgrids within a certain time period are used to form a sliding window frequency deviation sequence, which is used to assess the overall level and trend of system frequency fluctuations over a current period.

[0062] In this embodiment, the frequency deviation can be calculated using a frequency detection device on the local microgrid bus side with the same control period as the power sampling. The sliding window length can be selected based on the time scale the system wants to focus on, typically ranging from tens to hundreds of sampling points. For example, when the control period is 50 milliseconds, a window length of 20 sampling points corresponds to a time range of approximately one second, suitable for quickly detecting short-term oscillations; a window length of 200 sampling points corresponds to a time range of approximately ten seconds, suitable for evaluating frequency offset trends over a medium duration. In engineering practice, operators can prioritize selecting an appropriate window length based on the system's allowed frequency recovery time and inertia level, and adjust it during operation in conjunction with actual frequency fluctuations. Optionally, when frequency data for individual microgrids is missing or there are significant outliers at certain times, the outlier samples can be skipped or interpolated values ​​from adjacent times can be used to replace them when constructing the sliding window frequency deviation sequence, to avoid single-point errors excessively affecting the evaluation results.

[0063] Step e, for the current moment Using the sliding window frequency deviation sequence obtained in step d, an evaluation function with the sum of squared frequency deviations as the main term is constructed so that the evaluation function value can reflect the strength of frequency fluctuations within the window. On this basis, a penalty term for the power control parameters deviating from the reference value is superimposed to suppress large jumps in parameters. This evaluation function serves as the basis for the cost function of subsequent parameter iteration updates, providing a quantitative target for the online adaptive adjustment of power control parameters.

[0064] Optionally, the parameter deviation penalty trade-off coefficient in the evaluation function can be configured according to the different emphases of the system on frequency stability and parameter smoothness. When more attention is paid to quickly suppressing frequency deviation, the trade-off coefficient can be taken to a smaller value, making the evaluation function more biased towards the sum of squared frequency deviations. When more attention is paid to the smoothness of control parameter changes and actuator lifespan, the trade-off coefficient can be taken to a larger value to limit the power control parameter vector from deviating too much from the reference value. In this embodiment, the typical value of this trade-off coefficient can be between 0.01 and 1, and is tuned by the operator based on experience with the system frequency quality indicators and the frequency of control actions. The reference parameter vector can be determined based on offline tuning before commissioning or engineering experience. Usually, a parameter combination that has been proven to have basic stability under traditional droop control or fixed energy storage control gain is selected as the reference to ensure that adaptive adjustment is fine-tuned near a safe and acceptable starting point. For the numerical calculation of the evaluation function, in order to avoid numerical overflow or accuracy loss, the frequency deviation can be appropriately normalized in the implementation to keep the magnitudes involved in the calculation within a range that is easy to process numerically.

[0065] Step f, based on steps d and e, uses a gradient-based iterative approach to link the sliding window frequency deviation sequence with the power control parameter vector. The evaluation function and iterative update rule can be expressed as:

[0066]

[0067] in, Indicates at time The corresponding set of sliding window time indices is from arrive The set of discrete moments, This indicates the index of the current discrete control time, which is consistent with the sampling time of the aforementioned multi-microgrid operation parameters. This represents the length of the sliding window, which is the number of historical moments included in the evaluation. Indicates at time The evaluation function value is constructed based on the frequency deviation sequence and the power control parameter vector. Indicates at time The power control parameter vector can include components such as the active power droop coefficient, reactive power droop coefficient, and energy storage output power control gain for each microgrid. This represents the index of a historical moment within the sliding window, and iterates through the collection. At various moments in the process, This represents the index of a microgrid in a multi-microgrid parallel system, starting from... arrive Numbered sequentially, This represents the number of microgrids in a multi-microgrid parallel system, consistent with the number of nodes in the aforementioned graph neural network model. Indicates the first A microgrid at any time The frequency deviation is the deviation of the microgrid frequency from the rated frequency. This represents the tradeoff coefficient for the parameter deviation penalty term in the evaluation function, used to adjust the tradeoff between the frequency deviation suppression objective and the parameter smoothing objective. The reference vector representing the power control parameters can be set based on empirical tuning results or offline optimization results. The 2-norm of a vector is used to measure the Euclidean distance between parameter vectors. Indicates at time Updated power control parameter vector, This represents the step size coefficient for gradient-based iterations, used to control the magnitude of each parameter update. In the parameter vector Evaluation function Regarding parameter vectors The gradient vector is used to give the descent direction of the evaluation function;

[0068] Under this rule, the distributed controller updates the evaluation function based on the sliding window frequency deviation at each control moment and iteratively corrects the power control parameter vector along the gradient descent direction, so that the power control parameters can be updated online as the system frequency changes dynamically, taking into account both the frequency deviation suppression effect and the smoothness of parameter changes.

[0069] Specifically, by aggregating frequency deviation samples within a certain time range, the main term of the evaluation function characterizes the overall frequency fluctuation level within that time period, thereby introducing historical information of the operating status into the parameter optimization process; the penalty term is used to limit the deviation of the parameter vector from the preset benchmark, to avoid excessive changes in parameters in a short period of time, and to reduce the impact on the device execution layer.

[0070] Meanwhile, a gradient-type iterative update rule is introduced, which enables the sliding window data in each control cycle to correct the control parameters at the next moment, forming a closed loop of window evaluation-gradient descent-parameter update in the algorithm structure. This mechanism enables the power control parameters to be adjusted in a rolling manner according to the operating conditions. When the system frequency deviation continues to increase, the parameters are automatically adjusted in a direction that is conducive to enhancing frequency support and oscillation damping. When the frequency deviation falls back, the parameters gradually return to the benchmark level to maintain the stability and robustness of long-term operation, and provide a dynamic matching control parameter basis for the subsequent droop control and energy storage regulation links.

[0071] Furthermore, the step size coefficient of the gradient-type iteration can be tuned according to the system's tolerance for convergence speed and parameter oscillation. An excessively large step size may cause the power control parameters to change too drastically between adjacent control cycles, or even oscillate; an excessively small step size may cause the parameters to converge too slowly, making it difficult to reflect changes in operating conditions in a timely manner. In this embodiment, the step size coefficient can be initially set between 0.001 and 0.1, and fine-tuned during the trial operation phase by observing the frequency deviation convergence speed and the control parameter change curves to achieve a better compromise. To prevent the parameter vector from being pushed out of the safe range under extreme disturbances or large measurement noise, upper and lower limits can be set for the power control parameters. For example, the active power droop coefficient and reactive power droop coefficient can be limited to between pre-tuned minimum and maximum values, and the energy storage output power control gain can be limited to ensure that the charging and discharging power does not exceed the equipment's rated value. When the update result exceeds the boundary, it can be truncated to the boundary value. Optionally, when the frequency deviation is detected to be at a very low level and the parameter change amplitude is lower than a preset threshold within multiple consecutive control cycles, the parameter update process can be temporarily frozen to reduce unnecessary computational overhead, and the adaptive update mechanism can be reactivated when a new significant frequency deviation or change in operating conditions occurs.

[0072] In one embodiment, the power control parameters include active power droop coefficient and / or reactive power droop coefficient and / or energy storage output power control gain. During the iterative optimization process, the droop coefficient and / or energy storage output power control gain are dynamically adjusted according to the frequency deviation sequence to smooth the active power output trajectory of each microgrid and reduce the power oscillation amplitude.

[0073] In one embodiment, the graph neural network model is obtained through pre-training offline. Offline training constructs a training set containing power oscillation samples under various operating conditions based on historical operating data and / or simulation-generated data. After the system is put into operation, it is periodically retrained or the parameters are fine-tuned based on the data collected online to improve the accuracy of power oscillation risk prediction and the adaptability to new operating conditions.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0075] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A multi-energy complementary operation control system, characterized in that, include: Multiple microgrids, each microgrid including at least one renewable energy generation unit, at least one energy storage unit and at least one load unit, wherein the renewable energy generation unit and the energy storage unit are connected to the local microgrid bus via a grid-connected converter; Interconnecting lines are used to connect the local microgrid buses of the multiple microgrids in parallel or to interconnect them via a common collection bus to form a network topology of a multi-microgrid parallel system. A distributed controller deployed in each of the microgrids is configured to: collect the output active power, reactive power, and frequency operation of the local microgrid; process real-time power data of the network topology, including local operation and operation of adjacent microgrids coupled through the interconnecting lines, based on a graph neural network model; determine the target output power share of the local microgrid under the total power demand; generate power control commands for adjusting the output of the grid-connected converter; and realize the output power distribution among the multiple microgrids and suppress the propagation of power oscillations in the network topology.

2. The multi-energy complementary operation control system as described in claim 1, characterized in that, The renewable energy power generation unit includes at least one of wind power generation unit and photovoltaic power generation unit, the energy storage unit includes battery energy storage device, and the load unit includes at least one of industrial load and residential load.

3. A multi-energy complementary operation control system as described in claim 1 or 2, characterized in that, The graph neural network model includes at least one graph convolutional layer, which is used to extract the dynamic coupling characteristics of each microgrid and interconnection line based on the adjacency relationship and node characteristics of the network topology, and output the power oscillation risk index for each microgrid and the correction amount of the target output power share, so as to realize the collaborative optimization of power oscillation risk prediction and power allocation based on the graph convolutional layer.

4. The multi-energy complementary operation control system as described in claim 3, characterized in that, The distributed controller is also configured to: Based on the power oscillation risk index and the target output power share, the charging and discharging power and timing of the energy storage unit are adjusted so that the energy storage unit can dampen power oscillations while ensuring the output power distribution target of the multi-microgrid.

5. A multi-energy complementary operation control system as described in any one of claims 1 to 4, characterized in that, The distributed controller includes a local communication module, which is used to exchange limited state information, including local output power, frequency deviation and energy storage status, between adjacent microgrids. It adopts an event triggering mechanism to initiate the exchange of limited state information only when the local output power deviation exceeds a preset threshold and / or the power oscillation risk index exceeds a risk threshold.

6. A multi-energy complementary operation control system as described in claim 5, characterized in that, The local communication module adopts an asynchronous update mechanism, which allows each microgrid to independently update its timestamp and control variables after receiving the latest status information from its neighboring microgrids.

7. A multi-energy complementary operation control system as described in any one of claims 1 to 6, characterized in that, The distributed controller is also configured to: The frequency deviation of the local microgrid is monitored in real time, and the power control command is adaptively adjusted based on the frequency deviation, so that the output power distribution can maintain the power sharing coordination between microgrids while suppressing frequency fluctuations.

8. A multi-energy complementary operation control system as described in claim 7, characterized in that, The adaptive adjustment includes: A sliding window is used to record the frequency deviation sequence within a preset time period, and the power control parameters are iteratively optimized based on the frequency deviation sequence within the sliding window.

9. A multi-energy complementary operation control system as described in claim 8, characterized in that, The power control parameters include active power droop coefficient and / or reactive power droop coefficient and / or energy storage output power control gain. During the iterative optimization process, the droop coefficient and / or the energy storage output power control gain are dynamically adjusted according to the frequency deviation sequence.

10. A multi-energy complementary operation control system as described in any one of claims 3 to 9, characterized in that, The graph neural network model is obtained through pre-training offline. The offline training is based on historical operating data and / or simulation-generated data to construct a training set containing power oscillation samples under various operating conditions. After the system is put into operation, it is periodically retrained or the parameters are fine-tuned based on the data collected online.

Images