Micro-grid source and load prediction and optimal dispatching system

Abstract

The present application relates to the technical field of power dispatching system, and particularly relates to a micro-grid source-load prediction and optimal dispatching system, which comprises a sample generation module, a generated synthetic transient risk precursor waveform sample; a model training module, a transient risk identification model is trained and constructed using the enhanced training set; a benchmark planning module, a state of charge reference value sequence of energy storage in a first preset time scale is generated; a risk identification module, a transient risk index is output with a second preset time scale as an update cycle; a rolling optimization module, real-time power instructions for controlling the energy storage system are generated and issued; through the generative adversarial network enhanced transient risk identification model and the double-layer collaborative dispatching architecture based on dynamic weight, the structural contradiction between high state of charge of energy storage and renewable energy consumption capacity under a single economic optimization strategy is solved, and the dynamic balance of economy and transient safety in micro-grid operation is realized.

Description

Technical Field

[0001] This invention relates to the technical field of power dispatching systems, and more particularly to a microgrid source-load prediction and optimization dispatching system. Background Technology

[0002] As a crucial component of new power systems, the stable and economical operation of microgrids is paramount. In specific industrial settings such as ports and mines, there are numerous regenerative loads, such as the descent of heavy loads on port cranes and the downhill braking of mining vehicles. These loads convert kinetic or potential energy into electrical energy and feed it back to the microgrid. The resulting regenerative energy spikes are characterized by large power amplitudes and rapid rise times, significantly impacting the voltage stability and power quality of the microgrid. Energy storage systems, due to their rapid power response and bidirectional energy throughput, have become key devices for mitigating these impacts and improving the economic efficiency of system operation.

[0003] Existing microgrid dispatching methods typically employ a single-objective optimization strategy, aiming to minimize microgrid operating costs. When formulating charging and discharging plans for energy storage systems, this approach tends to maintain the system's state of charge (SOC) at a high level to maximize profits during peak electricity price periods. However, this economically optimal strategy creates a structural contradiction with the safety requirements for absorbing peak regenerative feedback: when the energy storage system is at a high SOC, its available capacity to absorb instantaneous high-power charging is severely insufficient, leading to ineffective absorption of regenerative feedback energy and even safety issues such as DC bus voltage exceeding limits. Long-term implementation of this strategy bias will result in a trade-off between operational safety and economic efficiency, making it difficult to achieve the expected operational benefits. Summary of the Invention

[0004] To address the technical problems existing in the background art, this invention proposes a microgrid source-load prediction and optimal scheduling system, the specific scheme of which is as follows: A microgrid source-load prediction and optimal scheduling system includes: The sample generation module is used to acquire historical operating data of the microgrid and divide it into a normal operating condition dataset and a key event dataset. A generative adversarial network is used to learn the distribution characteristics of the precursor waveforms in the key event dataset and generate synthetic transient risk precursor waveform samples. The model training module is used to construct an enhanced training set based on the conventional working condition dataset and the synthetic transient risk precursor waveform samples, and to train and construct an instantaneous risk identification model using the enhanced training set. The baseline planning module is used during online operation. Based on a preset mathematical optimization model, it performs long-term economic optimization according to the acquired electricity price and load forecast data, and generates a series of energy storage state of charge reference values ​​within the first preset time scale. The risk identification module is used to acquire high-frequency measurement data of the microgrid in real time, input the instantaneous risk identification model, and output the transient risk index with the second preset time scale as the update period. The rolling optimization module is used to construct a multi-objective optimization problem with dynamic weights based on the energy storage state of charge reference value sequence and the transient risk index, and solve it in a rolling manner to generate and issue real-time power commands for controlling the energy storage system.

[0005] Furthermore, in the sample generation module, the normal operating condition dataset is composed of sampled data from continuous operating segments in the historical operating data that do not include regenerative feedback energy peak events; The key event dataset consists of event window data containing regenerative feedback energy peak events in the historical operating data, and the event window data includes at least time-series data within a first preset time period before the peak event occurs.

[0006] Furthermore, the regenerative feedback energy spike events are identified in the following way: When the detected power feedback amplitude exceeds a preset power feedback threshold and the duration of exceeding the power feedback threshold is greater than a preset duration threshold, it is determined that a regenerative feedback energy spike event has occurred.

[0007] Furthermore, in the sample generation module, a generative adversarial network is used to learn the distribution characteristics of precursor waveforms in the key event dataset, generating synthetic transient risk precursor waveform samples, including: Construct a generative adversarial network, which includes a generator and a discriminator; Using the precursor waveform data in the key event dataset as real samples, the generator and the discriminator are subjected to adversarial training until the discriminator can no longer distinguish the waveform output by the generator from the real samples. The trained generator is used as a synthetic sample generator, and a random noise vector is used as input to generate the synthetic transient risk precursor waveform sample.

[0008] Furthermore, in the model training module, the instantaneous risk identification model is trained and constructed using the enhanced training set, including: Using the samples in the enhanced training set as input and the risk label corresponding to each sample as the output expectation, supervised training is performed on the initial classifier based on the temporal deep network to obtain the instantaneous risk identification model. The risk label indicates whether the sample corresponds to a precursor state of a regenerative feedback energy spike event.

[0009] Furthermore, in the baseline planning module, during online operation, based on a preset mathematical optimization model and the acquired electricity price and load forecast data, long-term economic optimization is performed to generate a sequence of energy storage state of charge reference values ​​within a first preset time scale, including: The preset mathematical optimization model is a mixed integer linear programming model; Based on the acquired electricity price data, load forecast data, and power constraint parameters of the interaction between the energy storage system and the power grid in the microgrid, the mixed integer linear programming method is used to construct the mixed integer linear programming model. The mixed-integer linear programming model is optimized to minimize the total operating cost of the microgrid within the first preset time scale, and generates the energy storage state of charge reference value sequence. The energy storage state of charge reference value sequence is composed of the energy storage state of charge values ​​at each discrete time point obtained by solving the mixed integer linear programming model.

[0010] Furthermore, the risk identification module acquires high-frequency measurement data of the microgrid in real time, including: The voltage, current, drive control signals and instantaneous power data of the microgrid, as well as the regenerative feedback load, are collected in real time at a sampling frequency of not less than 50Hz.

[0011] Furthermore, in the risk identification module, a transient risk index is output with a second preset time scale as the update cycle, including: The instantaneous risk identification model extracts time-series features and performs classification reasoning on the input high-frequency measurement data, and outputs a value normalized to the interval [0, 1] as the transient risk index, where 0 indicates no transient risk and 1 indicates the highest transient risk.

[0012] Furthermore, in the rolling optimization module, based on the energy storage state of charge reference value sequence and the transient risk index, a multi-objective optimization problem with dynamic weights is constructed and solved in a rolling manner, generating and issuing real-time power commands for controlling the energy storage system, including: A state-space prediction model for the energy storage system in the microgrid is established, wherein the state-space prediction model takes the energy storage charging and discharging power as input and the energy storage state of charge as output. Obtain the measured value of the energy storage state of charge at the current moment, and use the state-space prediction model to recursively calculate the energy storage state of charge in the future prediction time domain to obtain the sequence of predicted energy storage state of charge values ​​at each discrete moment in the prediction time domain; and construct the objective function of the multi-objective optimization problem. The objective function is used as the optimization objective of the multi-objective optimization problem. The charging and discharging power limit and the energy storage state of charge limit of the energy storage system are used as constraints to solve the optimal energy storage charging and discharging power sequence in the prediction time domain. The power command of the first control cycle in the optimal energy storage charging and discharging power sequence is used as the real-time power command and sent to the energy storage converter of the energy storage system for execution.

[0013] Furthermore, the objective function includes an economic deviation penalty term, a transient risk penalty term, and a power change penalty term; The economic deviation penalty term is used to penalize the deviation between the predicted energy storage state of charge and the energy storage state of charge reference value sequence, and the transient risk penalty term is positively correlated with the transient risk index and is used to penalize high energy storage state of charge. The weight of the economic deviation penalty term is negatively correlated with the transient risk index, while the weight of the transient risk penalty term is positively correlated with the transient risk index.

[0014] Compared with the prior art, the present invention can achieve at least the following beneficial effects: 1. This invention learns the precursor waveform distribution characteristics of regenerative feedback energy peak events through generative adversarial networks, generating synthetic transient risk precursor waveform samples to expand key event data. This solves the problem of insufficient model training due to the natural sparsity of such events in historical data, providing a reliable data foundation for risk identification in the online phase. During online operation, a two-layer collaborative architecture with different time scales is adopted: long-cycle economic optimization generates a reference value sequence of energy storage state of charge, while the real-time risk identification model outputs a transient risk index at a second-level cycle. The two are dynamically weighted and fused under the model predictive control framework. When the transient risk index is low, the system primarily tracks the economically optimal trajectory; when the transient risk index increases, the system automatically adjusts the optimization weights and proactively reduces the energy storage state of charge to reserve capacity margin, thus preparing for absorption before the regenerative feedback energy peak occurs. This enables the scheduling strategy to smoothly transition between safety and economy based on real-time operating conditions, avoiding the safety hazards of unabsorbed regenerative energy due to high state of charge while also taking into account the economic benefits of peak-valley electricity price arbitrage, effectively improving the overall operational efficiency of the microgrid.

[0015] 2. This invention addresses the safety hazard of energy storage systems losing their regenerative energy absorption capacity due to maintaining a high state of charge by constructing a technical solution of offline data augmentation and online dual-layer collaborative scheduling. It solves the problem of insufficient model training caused by sparse data for critical events by employing a generative adversarial network to learn the distribution characteristics of precursor waveforms of peak events and generate synthetic samples. During online operation, using the energy storage state of charge reference value sequence generated by long-cycle economic optimization as a benchmark and the transient risk index output by the real-time risk identification model as an adjustment signal, the system achieves smooth switching between economic and safety objectives through dynamic weights within the model predictive control framework. This allows the system to track the optimal economic trajectory when transient risk is low and proactively reduce the energy storage state of charge to reserve absorption margin when transient risk increases, fundamentally resolving the structural contradiction between economy and safety, and effectively improving the overall operational efficiency of the microgrid. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a system principle block diagram of the present invention. Detailed Implementation

[0017] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0018] Please refer to Figure 1 This invention provides a microgrid source-load prediction and optimal scheduling system, comprising: The sample generation module is used to acquire historical operating data of the microgrid and divide it into a normal operating condition dataset and a key event dataset. A generative adversarial network is used to learn the distribution characteristics of the precursor waveforms in the key event dataset and generate synthetic transient risk precursor waveform samples.

[0019] It should be noted that the historical operating data of the microgrid can be obtained from the microgrid energy management system or data acquisition and monitoring control system, including time-series data such as voltage, current, power, and control signals related to the operating status of regenerative feedback loads. Regenerative feedback loads refer to electrical equipment that can convert kinetic or potential energy into electrical energy and feed it back to the microgrid under specific operating conditions. Typical regenerative feedback loads include the descent of heavy loads on port cranes and the downhill braking of mining transport vehicles.

[0020] In an optional embodiment, in the sample generation module, the normal operating condition dataset consists of sampled data from continuous operating segments in the historical operating data that do not contain regenerative feedback energy peak events; The key event dataset consists of event window data containing regenerative feedback energy peak events in the historical operating data, and the event window data includes at least time-series data within a first preset time period before the peak event occurs.

[0021] It should be noted that the normal operating condition dataset consists of sampled data from continuous operation periods of the microgrid under normal operating conditions without any regenerative feedback energy peak events. Such data is abundant in the historical database. The critical event dataset consists of historically actual regenerative feedback energy peak events and their waveform data for a period before and after them. Such data is scarce in the historical database.

[0022] It should be noted that the first preset time period in the event window data refers to a time window preceding the occurrence of the peak event. The time-series data within this window contains characteristic information about the changes in the system state before the peak event, i.e., the precursor waveform. The precursor waveform is a key basis for training the instantaneous risk identification model, reflecting the changing patterns and characteristics of parameters such as voltage, current, and power in the microgrid before the occurrence of the regenerative feedback energy peak event.

[0023] In an optional embodiment, the regenerative feedback energy spike event is identified in the following way: When the detected power feedback amplitude exceeds a preset power feedback threshold and the duration of exceeding the power feedback threshold is greater than a preset duration threshold, it is determined that a regenerative feedback energy spike event has occurred.

[0024] It should be noted that the identification of regenerative feedback energy spike events employs a dual criterion combining a power threshold and a time threshold. The power feedback threshold refers to the amplitude threshold of the feedback power in the microgrid; when the feedback power exceeds this threshold, it indicates a possible regenerative feedback event. The duration threshold refers to the length of time the feedback power exceeds the power feedback threshold. The purpose of setting the duration threshold is to eliminate false triggers caused by measurement noise or transient disturbances. When the feedback power amplitude exceeds the power feedback threshold and the duration of this state exceeds the duration threshold, the system determines that a regenerative feedback energy spike event has occurred.

[0025] In an optional embodiment, the sample generation module employs a generative adversarial network to learn the distribution characteristics of precursor waveforms in the key event dataset, generating synthetic transient risk precursor waveform samples, including: Construct a generative adversarial network, which includes a generator and a discriminator; Using the precursor waveform data in the key event dataset as real samples, the generator and the discriminator are subjected to adversarial training until the discriminator can no longer distinguish the waveform output by the generator from the real samples. The trained generator is used as a synthetic sample generator, and a random noise vector is used as input to generate the synthetic transient risk precursor waveform sample.

[0026] It's important to note that Generative Adversarial Networks (GANs) are deep learning frameworks consisting of a generator and a discriminator. The generator takes a random noise vector as input and produces synthetic data with a distribution similar to real data. The discriminator distinguishes whether the input data comes from a real dataset or is synthetic data generated by the generator. During adversarial training, the generator continuously optimizes its parameters to generate more realistic synthetic data to deceive the discriminator, while the discriminator continuously optimizes its parameters to improve its discriminative ability. When the adversarial training reaches equilibrium—that is, when the discriminator's accuracy in distinguishing the generator's output waveform from real samples converges to approximately 50%—it indicates that the generator has learned the inherent distribution characteristics of the real precursor waveform. At this point, the discriminator can no longer effectively distinguish between generated and real samples.

[0027] It should be noted that the trained generator can take noise vectors randomly sampled from Gaussian or uniform distributions as input to generate massive and diverse synthetic transient risk precursor waveform samples. These synthetic samples are similar to real critical event data in statistical distribution characteristics, but they are not simple copies of real data. Therefore, they can effectively expand the sample size of critical event data and provide a sufficient data foundation for subsequent construction of enhanced training sets and training of transient risk identification models.

[0028] The model training module is used to construct an enhanced training set based on the conventional working condition dataset and the synthetic transient risk precursor waveform samples, and to train and construct an instantaneous risk identification model using the enhanced training set.

[0029] It should be noted that the purpose of constructing the augmented training set is to address the problem of insufficient model training caused by the inherent sparsity of critical event data. Since regenerative feedback energy spikes are rare in actual microgrid operation, the sample size of critical event datasets extracted directly from historical data is usually much smaller than that of datasets based on normal operating conditions. This class imbalance can cause classifiers to be biased in predicting input samples as normal operating conditions during training, thus reducing their ability to identify risk events. By mixing synthetic transient risk precursor waveform samples generated by generative adversarial networks with samples from the normal operating condition dataset, a relatively balanced augmented training set can be constructed, effectively improving the above problem.

[0030] In an optional embodiment, the model training module trains and constructs an instantaneous risk identification model using the augmented training set, including: Using the samples in the enhanced training set as input and the risk label corresponding to each sample as the output expectation, supervised training is performed on the initial classifier based on the temporal deep network to obtain the instantaneous risk identification model. The risk label indicates whether the sample corresponds to a precursor state of a regenerative feedback energy spike event.

[0031] It should be noted that the risk label is a category annotation for each sample in the enhanced training set, used to indicate whether the time window corresponding to the sample is in the precursor state of a regenerative feedback energy peak event. The risk label can use a binary annotation method. For example, a sample corresponding to the first preset time period before the peak event is labeled as "1", indicating that the sample belongs to the risk precursor state; a sample in the normal operating condition dataset is labeled as "0", indicating that the sample belongs to the normal operating state. The synthetic transient risk precursor waveform sample is generated by the trained generator. Its statistical characteristics are similar to the real critical event precursor waveform, so its risk label is also labeled as "1".

[0032] It should be noted that the initial classifier based on temporal deep networks can employ deep learning models suitable for processing time series data, such as Long Short-Term Memory (LSTM) networks or Temporal Convolutional Networks (TCNNs). LSM networks, by introducing a gating mechanism, can effectively capture long-term dependencies in time series data, making them suitable for modeling features spanning multiple time steps in precursor waveforms. TCNNs, by expanding the receptive field through convolutional operations, can process the entire time series in parallel, offering advantages in training efficiency.

[0033] It should be noted that supervised training refers to the process of using augmented training set samples with risk labels, inputting the samples into an initial classifier to obtain predicted outputs, calculating the error between the predicted outputs and the risk labels, and updating the network parameters through backpropagation. This process is repeated multiple times until the classifier's prediction accuracy reaches a preset requirement. The resulting instantaneous risk identification model, after training, possesses the ability to extract temporal features and perform classification inference on high-frequency measurement data.

[0034] The baseline planning module, used during online operation, performs long-term economic optimization based on a preset mathematical optimization model and the acquired electricity price and load forecast data, generating a sequence of energy storage state of charge reference values ​​within a first preset time scale.

[0035] It should be noted that the first preset time scale refers to the time span of long-term economic optimization, usually set to the minute level, such as 15 minutes. The generated energy storage state of charge reference value sequence covers a complete future scheduling cycle, such as 24 hours. The next 24 hours are divided into 96 discrete time points with an optimization cycle of 15 minutes. Each discrete time point corresponds to an energy storage state of charge reference value. The reference values ​​at all discrete time points are arranged in chronological order to form the energy storage state of charge reference value sequence.

[0036] It should be noted that the electricity price data refers to the time-of-use electricity price information for the microgrid's purchase and sale of electricity from and to the external grid, typically including peak-hour, off-peak, and flat-hour prices. Load forecasting data refers to the predicted load power demand of the microgrid during future dispatch cycles, which can be obtained through statistical analysis of historical load data or load forecasting methods based on machine learning.

[0037] In an optional embodiment, during online operation, the baseline planning module performs long-term economic optimization based on a preset mathematical optimization model and the acquired electricity price and load forecast data, generating a sequence of energy storage state of charge reference values ​​within a first preset time scale, including: The preset mathematical optimization model is a mixed integer linear programming model; Based on the acquired electricity price data, load forecast data, and power constraint parameters of the interaction between the energy storage system and the power grid in the microgrid, the mixed integer linear programming method is used to construct the mixed integer linear programming model. The mixed-integer linear programming model is optimized to minimize the total operating cost of the microgrid within the first preset time scale, and generates the energy storage state of charge reference value sequence. The energy storage state of charge reference value sequence is composed of the energy storage state of charge values ​​at each discrete time point obtained by solving the mixed integer linear programming model.

[0038] It should be noted that mixed-integer linear programming is a mathematical optimization method where both the objective function and constraints are linear expressions, and the decision variables include both continuous and integer variables. In this embodiment, continuous variables may include the charging and discharging power of the energy storage system, and integer variables may include the charging and discharging status flags of the energy storage system. The total operating cost includes the cost of the microgrid purchasing electricity from the external grid, the loss cost of the energy storage system's charging and discharging cycles, and the possible load shedding penalty cost, minus the revenue from selling electricity to the external grid. The constraints include limits on the charging and discharging power of the energy storage system, upper and lower limits on the energy storage's state of charge, and limits on power exchange between the microgrid and the external grid.

[0039] It should be noted that in long-term economic optimization, the mixed-integer linear programming model, based on the peak-valley electricity price difference reflected in the electricity price data, schedules the energy storage system to charge during off-peak hours to store low-priced electricity, and schedules the energy storage system to discharge during peak hours to supply loads or sell electricity to the grid, thereby minimizing operating costs by utilizing the peak-valley price difference. The energy storage state-of-charge reference value sequence obtained after optimization reflects the target state-of-charge value that the energy storage system should achieve at each discrete time point from an economic optimal perspective throughout the entire scheduling cycle. This reference value sequence provides a long-term economic benchmark for real-time correction and rolling optimization in subsequent steps.

[0040] The risk identification module is used to acquire high-frequency measurement data of the microgrid in real time, input the instantaneous risk identification model, and output the transient risk index with a second preset time scale as the update period.

[0041] It should be noted that the high-frequency measurement data refers to the microgrid electrical quantities and load status quantities collected in real time at a high sampling frequency. Voltage and current are the basic electrical quantities of the microgrid, reflecting its real-time operating status. The drive control signal for regenerative feedback loads refers to the control commands or status feedback signals that control the operating status of the load, such as the hoisting / lowering commands and speed setpoint signals of a port crane. Instantaneous power data refers to the actual active power of the regenerative feedback load at the current moment. These data collectively constitute the multi-dimensional input features used for transient risk identification.

[0042] In an optional embodiment, the risk identification module acquires high-frequency measurement data of the microgrid in real time, including: The voltage, current, drive control signals and instantaneous power data of the microgrid, as well as the regenerative feedback load, are collected in real time at a sampling frequency of not less than 50Hz.

[0043] It should be noted that the sampling frequency is set to no less than 50Hz because the precursor waveform characteristics of regenerative energy feedback spike events typically occur on a sub-second to second timescale. Taking the example of a crane lowering a heavy load, the transition from issuing the braking command to the rapid power feedback usually takes hundreds of milliseconds to several seconds. If the sampling frequency is too low, the key waveform characteristics of this transition process will not be captured, causing the instantaneous risk identification model to fail to identify the impending spike event in a timely and accurate manner. A sampling frequency of no less than 50Hz ensures that at least one data point is collected within each power frequency cycle, providing sufficient time resolution for feature extraction of the precursor waveform.

[0044] In an optional embodiment, the risk identification module outputs a transient risk index with a second preset time scale as the update period, including: The instantaneous risk identification model extracts time-series features and performs classification reasoning on the input high-frequency measurement data, and outputs a value normalized to the interval [0, 1] as the transient risk index, where 0 indicates no transient risk and 1 indicates the highest transient risk.

[0045] It should be noted that the transient risk index is a scalar value normalized to the interval [0, 1], used to quantitatively characterize the probability and estimated impact intensity of an impending regenerative energy feedback peak event. A transient risk index of 0 indicates that there is no transient risk at the current moment, the system is in normal operation, or the probability of a peak event is extremely low; a transient risk index of 1 indicates the highest transient risk, meaning that a regenerative energy feedback peak event is highly likely to occur in a very short time and the estimated impact intensity is very high. The median value of the transient risk index reflects different levels of transient risk; for example, 0.3 indicates a low level of transient risk, and 0.7 indicates a high level of transient risk. This normalized output format facilitates the calculation of dynamic weights and the solution of multi-objective optimization problems in subsequent steps.

[0046] The rolling optimization module is used to construct a multi-objective optimization problem with dynamic weights based on the energy storage state of charge reference value sequence and the transient risk index, and solve it in a rolling manner to generate and issue real-time power commands for controlling the energy storage system.

[0047] In an optional embodiment, the rolling optimization module constructs a multi-objective optimization problem with dynamic weights based on the energy storage state of charge reference value sequence and the transient risk index, and solves it in a rolling manner to generate and issue real-time power commands for controlling the energy storage system, including: A state-space prediction model for the energy storage system in the microgrid is established, wherein the state-space prediction model takes the energy storage charging and discharging power as input and the energy storage state of charge as output. It should be noted that the state-space prediction model is a mathematical model describing the dynamic behavior of an energy storage system. It is used to recursively calculate the predicted state of charge (SBC) values ​​for future discrete moments based on the current SBC and the input charge / discharge power sequence. This state-space prediction model can be established based on the energy conservation principle of the energy storage system. Its core relationship is: the SBC at the next moment equals the current SBC plus the product of the charge / discharge power and the time step, divided by the rated capacity of the energy storage system, taking into account the influence of charge / discharge efficiency.

[0048] Obtain the measured value of the energy storage state of charge at the current moment, and use the state-space prediction model to recursively calculate the energy storage state of charge in the future prediction time domain to obtain the sequence of predicted energy storage state of charge values ​​at each discrete moment in the prediction time domain; and construct the objective function of the multi-objective optimization problem. It should be noted that the prediction time domain refers to the time range used in model predictive control to predict the future system state, typically set to several control cycles. The control cycle refers to the interval between rolling optimization executions of model predictive control, usually set to the second level, such as 1 second. In each control cycle, model predictive control recursively calculates the energy storage state of charge at each discrete moment within the future prediction time domain based on the measured value of the energy storage state of charge at the current moment and the state-space prediction model, obtaining the sequence of predicted energy storage state of charge values.

[0049] The objective function is used as the optimization objective of the multi-objective optimization problem. The charging and discharging power limit and the energy storage state of charge limit of the energy storage system are used as constraints to solve the optimal energy storage charging and discharging power sequence in the prediction time domain. The power command of the first control cycle in the optimal energy storage charging and discharging power sequence is used as the real-time power command and sent to the energy storage converter of the energy storage system for execution.

[0050] In an optional embodiment, the objective function includes an economic deviation penalty term, a transient risk penalty term, and a power change penalty term; The economic deviation penalty term is used to penalize the deviation between the predicted energy storage state of charge and the energy storage state of charge reference value sequence, and the transient risk penalty term is positively correlated with the transient risk index and is used to penalize high energy storage state of charge. The weight of the economic deviation penalty term is negatively correlated with the transient risk index, while the weight of the transient risk penalty term is positively correlated with the transient risk index.

[0051] It should be noted that the economic deviation penalty term in the objective function measures the degree of deviation between the predicted energy storage state of charge (SPC) sequence and the reference SPC sequence in the prediction time domain. This term guides the energy storage system to track the economically optimal trajectory determined by long-cycle economic optimization as much as possible while meeting safety constraints. The transient risk penalty term measures the degree of transient risk faced by the current system state. This term imposes a larger penalty on higher SPC predictions when the transient risk index increases. Its function is to prompt the optimizer to actively reduce the SPC when the transient risk increases, thereby reserving capacity margin for absorbing upcoming regenerative feedback energy spikes. The power variation penalty term penalizes the variation in energy storage charging and discharging power between adjacent control cycles. Its function is to avoid frequent and large fluctuations in the energy storage system's power command, extending the lifespan of the energy storage system.

[0052] It should be noted that the dynamic adjustment relationship between the weights of the economic deviation penalty term and the transient risk penalty term reflects the dynamic balance mechanism between the long-term economic optimization objective and the instantaneous transient safety objective. When the transient risk index is low, the weight of the economic deviation penalty term is large, and the weight of the transient risk penalty term is small, with the model predictive control decision primarily focused on tracking the economically optimal trajectory. When the transient risk index increases, the weight of the economic deviation penalty term decreases, and the weight of the transient risk penalty term increases, with the model predictive control decision primarily focused on ensuring transient safety, actively reducing the energy storage state of charge to cope with the impending regenerative feedback energy peak.

[0053] It should be noted that the rolling solution refers to the process where, in each control cycle, the model predictive control re-solves the multi-objective optimization problem based on the latest energy storage state-of-charge reference value sequence and the latest output transient risk index to obtain the optimal control sequence for the current moment. However, only the power command for the first control cycle in the sequence is issued to the energy storage converter for execution. In the next control cycle, the system re-acquires the latest measurement data and transient risk index, repeating the above optimization solution and execution process. This rolling optimization method can continuously correct control decisions using real-time feedback information, effectively addressing uncertainties in the microgrid's operating state.

[0054] Wherein, the objective function for: ; in, This is the sequence of reference values ​​for the energy storage state of charge. The transient risk index is... To predict the power change sequence formed by the difference between the energy storage charging and discharging power at each discrete moment in the time domain and the energy storage charging and discharging power at the current moment; Weighting for economic deviation penalties As a transient risk penalty weight, Weights for power change penalties; For transient risk penalty function, in the same Value below, The function value varies with The value increases with the increase, and at the same level... Value below, The function value varies with It increases as the value increases.

[0055] It should be noted that the objective function It consists of a weighted sum of three penalty items, namely, the economic deviation penalty item. Transient risk penalty items and power change penalty term .

[0056] It should be noted that the economic deviation penalty term is used to measure the sequence of predicted energy storage state of charge values ​​at each discrete moment in the prediction time domain. With the energy storage state of charge reference value sequence The deviation between them. This term is calculated using the squared Euclidean distance. This is the weighting coefficient for this term. The purpose of the economic deviation penalty term is to guide the optimization solver to follow the economically optimal trajectory determined by long-cycle economic optimization as much as possible during decision-making, thereby ensuring the economic efficiency of the system operation.

[0057] It should be noted that the transient risk penalty term is determined by the transient risk penalty weight. With transient risk penalty function Multiplication constitutes the product. Among them, This refers to the transient risk index output by the instantaneous risk identification model. Transient risk penalty function. It has the following properties: when the transient risk index At a fixed value, the predicted state of charge of energy storage The higher the value, the larger the function value, and the heavier the penalty; when the predicted state of charge of energy storage... When fixed, transient risk index The higher the value, the larger the function value, and the heavier the penalty. The purpose of this penalty term is to impose a greater penalty on the decision to maintain a high state of charge when the system detects an increased transient risk of a regenerative energy feedback peak event. This prompts the optimization solver to actively choose a power command to reduce the state of charge, thereby reserving sufficient capacity margin to absorb the impending energy peak. In a specific implementation, the transient risk penalty function... It can be taken as: ; In the form of, The preset adjustment coefficient, This is a preset safe state of charge threshold. When... When the exponent increases, the exponent term A sharp increase, thus affecting Exceed The portion is subject to an exponentially amplified penalty.

[0058] It should be noted that the power change penalty term is used to measure the power change sequence formed by the difference between the energy storage charging and discharging power at each discrete moment in the prediction time domain and the energy storage charging and discharging power at the current moment. Size, This is the weighting coefficient for this term. The purpose of the power change penalty term is to suppress large jumps in energy storage charging and discharging power commands between adjacent control cycles, avoid accelerated aging of the energy storage system due to frequent and drastic fluctuations in power commands, and extend the service life of the energy storage system.

[0059] It should be noted that weight and The value is determined based on the transient risk index. Dynamic adjustments will be made. Among them, Set as The function is a monotonically increasing function, meaning that the higher the transient risk, the greater the transient risk penalty weight. Correspondingly, the higher the transient risk, the smaller the penalty weight for economic deviation. This dynamic weighting mechanism enables the system to prioritize economy under normal operating conditions, and automatically and smoothly switch to prioritizing safety when transient risk increases, thereby achieving a dynamic balance between economy and safety within a single optimization framework.

[0060] In summary, this invention application solves the structural contradiction between the high-charge state of energy storage and the capacity for renewable energy absorption under a single economic optimization strategy by using a generative adversarial network-enhanced instantaneous risk identification model and a two-layer collaborative scheduling architecture based on dynamic weights, thereby achieving a dynamic balance between economy and transient safety in microgrid operation.

[0061] Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0062] In the embodiments provided by this invention, it should be understood that the disclosed system or method can be implemented in other ways. For example, the embodiments of the invention described above are merely illustrative; for instance, the division of modules is only a logical functional division, and there may be other division methods in actual implementation.

[0063] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0064] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in the form of hardware plus software functional modules.

[0065] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the basic characteristics of the present invention.

[0066] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A microgrid source-load prediction and optimal scheduling system, characterized in that, include: The sample generation module is used to acquire historical operating data of the microgrid, which is divided into a routine operating condition dataset and a key event dataset. Generative adversarial networks are used to learn the distribution characteristics of precursor waveforms in the key event dataset and generate synthetic transient risk precursor waveform samples. The model training module is used to construct an enhanced training set based on the conventional working condition dataset and the synthetic transient risk precursor waveform samples, and to train and construct an instantaneous risk identification model using the enhanced training set. The baseline planning module is used during online operation. Based on a preset mathematical optimization model, it performs long-term economic optimization according to the acquired electricity price and load forecast data, and generates a series of energy storage state of charge reference values ​​within the first preset time scale. The risk identification module is used to acquire high-frequency measurement data of the microgrid in real time, input the instantaneous risk identification model, and output the transient risk index with the second preset time scale as the update period. The rolling optimization module is used to construct a multi-objective optimization problem with dynamic weights based on the energy storage state of charge reference value sequence and the transient risk index, and solve it in a rolling manner to generate and issue real-time power commands for controlling the energy storage system.

2. The microgrid source-load prediction and optimal scheduling system as described in claim 1, characterized in that: In the sample generation module, the normal operating condition dataset consists of sampled data from continuous operating segments in the historical operating data that do not include regenerative feedback energy peak events; The key event dataset consists of event window data containing regenerative feedback energy peak events in the historical operating data, and the event window data includes at least time-series data within a first preset time period before the occurrence of the peak event.

3. The microgrid source-load prediction and optimal scheduling system as described in claim 2, characterized in that: The regenerative feedback energy spike events are identified in the following ways: When the detected power feedback amplitude exceeds a preset power feedback threshold and the duration of exceeding the power feedback threshold is greater than a preset duration threshold, it is determined that a regenerative feedback energy spike event has occurred.

4. The microgrid source-load prediction and optimal scheduling system as described in claim 2, characterized in that: In the sample generation module, a generative adversarial network is used to learn the distribution characteristics of precursor waveforms in the key event dataset, generating synthetic transient risk precursor waveform samples, including: Construct a generative adversarial network, which includes a generator and a discriminator; Using the precursor waveform data in the key event dataset as real samples, the generator and the discriminator are subjected to adversarial training until the discriminator can no longer distinguish the waveform output by the generator from the real samples. The trained generator is used as a synthetic sample generator, and a random noise vector is used as input to generate the synthetic transient risk precursor waveform sample.

5. The microgrid source-load prediction and optimal scheduling system as described in claim 4, characterized in that: In the model training module, the instantaneous risk identification model is trained and constructed using the augmented training set, including: Using the samples in the enhanced training set as input and the risk label corresponding to each sample as the output expectation, supervised training is performed on the initial classifier based on the temporal deep network to obtain the instantaneous risk identification model. The risk label indicates whether the sample corresponds to a precursor state of a regenerative feedback energy spike event.

6. The microgrid source-load prediction and optimal scheduling system as described in claim 1, characterized in that: In the baseline planning module, during online operation, based on a preset mathematical optimization model and the acquired electricity price and load forecast data, long-term economic optimization is performed to generate a sequence of energy storage state of charge reference values ​​within a first preset time scale, including: The preset mathematical optimization model is a mixed integer linear programming model; Based on the acquired electricity price data, load forecast data, and power constraint parameters of the interaction between the energy storage system and the power grid in the microgrid, the mixed integer linear programming method is used to construct the mixed integer linear programming model. The mixed-integer linear programming model is optimized to minimize the total operating cost of the microgrid within the first preset time scale, and generates the energy storage state of charge reference value sequence. The energy storage state of charge reference value sequence is composed of the energy storage state of charge values ​​at each discrete time point obtained by solving the mixed integer linear programming model.

7. The microgrid source-load prediction and optimal scheduling system as described in claim 1, characterized in that: The risk identification module acquires high-frequency measurement data of the microgrid in real time, including: The voltage, current, drive control signals and instantaneous power data of the microgrid, as well as the regenerative feedback load, are collected in real time at a sampling frequency of not less than 50Hz.

8. The microgrid source-load prediction and optimal scheduling system as described in claim 7, characterized in that: In the risk identification module, a transient risk index is output with a second preset time scale as the update cycle, including: The instantaneous risk identification model extracts time-series features and performs classification reasoning on the input high-frequency measurement data, and outputs a value normalized to the interval [0, 1] as the transient risk index, where 0 indicates no transient risk and 1 indicates the highest transient risk.

9. The microgrid source-load prediction and optimal scheduling system as described in claim 1, characterized in that: In the rolling optimization module, based on the energy storage state of charge reference value sequence and the transient risk index, a multi-objective optimization problem with dynamic weights is constructed and solved in a rolling manner, generating and issuing real-time power commands for controlling the energy storage system, including: A state-space prediction model for the energy storage system in the microgrid is established, wherein the state-space prediction model takes the energy storage charging and discharging power as input and the energy storage state of charge as output. Obtain the measured value of the energy storage state of charge at the current moment, and use the state-space prediction model to recursively calculate the energy storage state of charge in the future prediction time domain to obtain the sequence of predicted energy storage state of charge values ​​at each discrete moment in the prediction time domain; and construct the objective function of the multi-objective optimization problem. The objective function is used as the optimization objective of the multi-objective optimization problem. The charging and discharging power limit and the energy storage state of charge limit of the energy storage system are used as constraints to solve the optimal energy storage charging and discharging power sequence in the prediction time domain. The power command of the first control cycle in the optimal energy storage charging and discharging power sequence is used as the real-time power command and sent to the energy storage converter of the energy storage system for execution.

10. The microgrid source-load prediction and optimal scheduling system as described in claim 9, characterized in that: The objective function includes an economic deviation penalty term, a transient risk penalty term, and a power change penalty term; The economic deviation penalty term is used to penalize the deviation between the predicted energy storage state of charge and the energy storage state of charge reference value sequence, and the transient risk penalty term is positively correlated with the transient risk index and is used to penalize high energy storage state of charge. The weight of the economic deviation penalty term is negatively correlated with the transient risk index, while the weight of the transient risk penalty term is positively correlated with the transient risk index.

Images