Micro-grid multi-source collaborative optimization scheduling method based on AI

Abstract

The invention relates to the technical field of micro-grid operation control, in particular to an AI-based micro-grid multi-source collaborative optimization scheduling method, which comprises the following steps: collecting multi-source heterogeneous original operation data of a distributed power supply, energy storage, an adjustable load and a superior power grid in a micro-grid, and extracting key feature vectors after cleaning and standardization processing; and inputting the feature vectors into a pre-trained multi-source cooperative representation neural network model to generate a cooperative state vector set fusing the multi-source complementary characteristics and the time sequence dependency relationship. And on the basis, a scheduling decision generation network model is called to carry out multi-step prospective optimization deduction, and a coordinated scheduling instruction set of multiple time periods in the future is generated. And performing security constraint verification on the instruction set to form a final executable scheduling control instruction sequence, and issuing the final executable scheduling control instruction sequence to the controllable equipment. According to the invention, the deep learning model is utilized to deeply mine the multi-source cooperation characteristic and make a look-ahead decision, and the intelligent level and the overall optimization effect of micro-grid operation are improved.

Description

Technical Field

[0001] This invention relates to the field of microgrid operation and control technology, and in particular to an AI-based multi-source collaborative optimization scheduling method for microgrids. Background Technology

[0002] Microgrids are autonomous systems integrating various heterogeneous resources such as distributed photovoltaics, wind turbines, energy storage, and adjustable loads. Their efficient and stable operation depends on the coordinated scheduling of multiple types of internal equipment. Existing scheduling methods are mainly divided into two categories: centralized optimization and decomposition coordination. Centralized optimization methods typically construct a unified mixed-integer programming model to integrate and solve the operational constraints of various equipment with system security constraints. However, microgrid operational data comes from diverse sources and varies in format and spatiotemporal scale, making it difficult for centralized models to effectively handle this multi-source heterogeneity, resulting in insufficient accuracy in depicting the actual operating state of the system. Decomposition coordination methods, on the other hand, decompose the problem into several sub-problems according to equipment type or region, and coordinate them through alternating iterations. These methods reduce the complexity of the problem to some extent, but the coordination mechanism between sub-problems often relies on fixed boundary conditions or simple price signals, failing to fully characterize the strong randomness of distributed power output, the temporal coupling of energy storage charging and discharging, and the deep dynamic complementary relationship between load adjustable potential. Existing technologies lack effective joint characterization methods for the complex, nonlinear, and time-dependent collaborative relationships among multi-source devices. Dispatch commands are mostly generated based on the current or short-term state, lacking forward-looking and comprehensive projections of the evolution of operating states over subsequent periods. When facing drastic fluctuations in renewable energy, this can easily lead to short-sighted dispatch commands, causing operational risks such as frequent charging and discharging of energy storage devices or exceeding power limits in interactions with the upstream grid. Therefore, how to accurately perceive and encode the overall collaborative operating state of the system from multi-source heterogeneous data, and generate forward-looking and coordinated optimized dispatch commands based on this, is a key problem that urgently needs to be solved in the field of microgrid operation and control. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies by proposing an AI-based multi-source collaborative optimization scheduling method for microgrids.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: an AI-based microgrid multi-source collaborative optimization scheduling method, comprising: Collect power generation data from distributed power sources, status data from energy storage devices, operational data from adjustable loads, and interaction constraint data from the upper-level power grid in the microgrid system to generate a multi-source heterogeneous microgrid raw operational data set; The original microgrid operation data set is cleaned and format-unified to form a standardized microgrid operation dataset. Key feature vectors representing multi-source collaborative relationships are extracted from the standardized microgrid operation dataset. The key feature vectors are input into a pre-trained multi-source collaborative representation neural network model to perform joint state encoding on the distributed power source, the energy storage device, and the adjustable load, generating a set of collaborative state vectors that includes multi-source complementary characteristics and time-series dependencies. Based on the aforementioned set of collaborative state vectors, the scheduling decision generation network model is invoked to perform multi-step forward-looking optimization and deduction, generating a set of coordinated scheduling instructions for multiple future scheduling periods; The coordination and scheduling instruction set is subjected to security constraint verification, and a final executable scheduling control instruction sequence is generated based on the verification results and sent to various controllable devices in the microgrid system for execution.

[0005] As a further aspect of the present invention, the step of performing data cleaning and format unification processing on the original microgrid operation data set to form a standardized microgrid operation dataset includes: The standardized microgrid operation dataset includes timestamps, power generation sequences of each distributed power source, state of charge sequences of energy storage devices, power sequences of adjustable loads, and upper and lower limits sequences of interactive constraints. Abnormal values ​​in the power generation data, status data, operation data, and interaction constraint data are identified and corrected. The identification of abnormal values ​​is based on setting a deviation threshold according to the statistical distribution of historical data, and the correction is completed by interpolation of nearby valid data. The corrected power generation data, status data, operation data, and interaction constraint data are clock-synchronized and resampled according to a unified time base so that all data sequences have the same time resolution; The power generation data, state of charge data, power data, and upper and lower limit constraint data that have completed clock synchronization are normalized and mapped to the same numerical range to form the standardized microgrid operation dataset.

[0006] As a further aspect of the present invention, the extraction of key feature vectors representing multi-source collaborative relationships from the standardized microgrid operation dataset includes: The key feature vectors include source-load power matching characteristics at multiple time scales, energy storage fluctuation mitigation capabilities, and power trend characteristics in interaction with the power grid. Under multiple preset time window scales, the real-time difference between the power generation sequence and the adjustable load power sequence is calculated to form a source-load power deviation sequence with multiple time scales; Based on the state of charge sequence and the rated parameters of the energy storage device, a quantitative index characterizing the energy storage system's ability to absorb or release power per unit time to smooth source-load fluctuations is calculated, thus obtaining the energy storage fluctuation smoothing capability characteristics. By analyzing the relationship between the power generation sequence, the adjustable load power sequence, and the upper and lower limit sequences of the interaction constraints, the power change trend pattern of the microgrid purchasing or selling electricity to the grid during the dispatch cycle is extracted, and the power trend characteristics of the interaction with the grid are obtained. The key feature vector is constructed by combining the source-load power deviation sequence at multiple time scales, the energy storage fluctuation smoothing capability characteristics, and the power trend characteristics interacting with the power grid.

[0007] As a further aspect of the present invention, the step of inputting the key feature vector into a pre-trained multi-source collaborative representation neural network model to perform joint state encoding on the distributed power source, the energy storage device, and the adjustable load, generating a set of collaborative state vectors containing multi-source complementary characteristics and temporal dependencies, includes: The multi-source collaborative representation neural network model includes a feature interaction coding layer and a temporal dependency coding layer; The feature interaction coding layer receives the key feature vector and calculates the dynamic correlation weights between different distributed power sources, between distributed power sources and energy storage, and between source, storage and load through an internal multi-head attention mechanism to generate preliminary fusion features. The temporal-dependent coding layer receives the preliminary fused features, captures the evolution of the preliminary fused features in the time dimension through its internal recurrent neural network structure, and outputs a feature sequence containing temporal context information. The feature sequence containing temporal context information is pooled and compressed to generate a fixed-dimensional state vector for each scheduling moment. The state vectors of all scheduling moments constitute the collaborative state vector set.

[0008] As a further aspect of the present invention, the step of invoking a scheduling decision generation network model based on the cooperative state vector set to perform multi-step forward-looking optimization and deduction, generating a coordinated scheduling instruction set for multiple future scheduling periods, includes: The coordinated scheduling instruction set includes planned output instructions for each distributed power source, charging and discharging power instructions for energy storage devices, and control instructions for adjustable loads. The set of cooperative state vectors is input into the scheduling decision generation network model, which is constructed based on a deep reinforcement learning framework. The scheduling decision generation network model takes the current and historical cooperative state vectors as input to simulate the dynamic evolution process of the microgrid system in a preset multi-step scheduling time domain. In each step of the evolution simulation, the scheduling decision generation network model outputs the temporary scheduling action corresponding to the scheduling time. The temporary scheduling action includes the simulated power of various distributed power sources, the simulated charging and discharging power of energy storage devices, and the simulated adjustment amount of adjustable loads. Collect all temporary scheduling actions during the simulation process, and combine them with the evaluation of the overall system operation status at the end of the simulation to optimize and adjust the sequence of temporary scheduling actions, forming the final coordinated scheduling instruction set.

[0009] As a further aspect of the present invention, the scheduling decision generation network model is constructed based on a deep reinforcement learning framework, including: The state space of the scheduling decision generation network model is defined as the space spanned by the set of cooperative state vectors; The action space of the scheduling decision generation network model is defined as the allowable adjustment range of each controllable device in a single scheduling period. A reward function is defined, which is used to evaluate the temporary scheduling action in real time during the simulation. Its calculation is based on the source-load matching degree, energy storage health degree and limit violation penalty of interaction with the grid under the simulation state. The scheduling decision generation network model is trained offline and fine-tuned online using a large amount of historical and simulation data, enabling it to learn to select the scheduling action sequence that maximizes the cumulative reward in the state space.

[0010] As a further aspect of the present invention, the step of performing security constraint verification on the coordinated scheduling instruction set includes: The output limits and ramp rate limits of each distributed power source, the safe range of the state of charge and the charging and discharging power limits of the energy storage device, the adjustment range and adjustment rate limits of the adjustable load, and the power exchange limits of the connection point between the microgrid and the upper-level grid are obtained from the standardized microgrid operation dataset and used as a set of physical safety constraints. The planned output instructions, charging and discharging power instructions, and control instructions in the coordination and scheduling instruction set are compared with the physical safety constraint set in each time period. Check whether the planned output command meets the upper and lower limits of output and the ramp rate limit; check whether the charging and discharging power command causes the energy storage state of charge to exceed the safe range or the charging and discharging power to exceed the limit; check whether the control command exceeds the adjustment range and rate limit of the adjustable load; and check whether the net exchange power with the grid exceeds the connection point exchange limit. Record all instructions that do not meet the set of physical security constraints and the specific constraints they violate, and generate a constraint violation report.

[0011] As a further aspect of the present invention, the step of generating a final executable sequence of scheduling control instructions based on the verification results includes: If the constraint violation report is empty, it indicates that the coordinated scheduling instruction set fully satisfies all physical security constraints, and it is directly arranged into the scheduling control instruction sequence in chronological order; If the constraint violation report is not empty, the instruction correction process is initiated: for each instruction in the report that violates the constraint, under the premise of satisfying its original optimization objective, it is corrected to satisfy the corresponding physical security constraint using the principle of minimum adjustment. The revised instructions are then reintegrated with the original instructions that do not violate constraints to form a revised set of scheduling instructions that satisfies all safety constraints. The instructions in the modified scheduling instruction set are subjected to a timing and logic consistency check to ensure that there are no time or logical conflicts between the instructions. After passing the check, the instructions are arranged in chronological order to generate the final executable scheduling control instruction sequence.

[0012] As a further aspect of the present invention, the execution of the commands sent to various controllable devices in the microgrid system includes: The sequence of scheduling control instructions is encapsulated into a data packet of a specified format according to the communication protocol; Through the communication network of the microgrid energy management system, the encapsulated data packets are sent to the corresponding distributed power controller, energy storage system manager and load aggregation controller respectively; The distributed power controller, energy storage system manager, and load aggregation controller receive and parse the data packet, extract their respective control instructions, and convert them into drive signals or switching control signals for the underlying power electronic devices to execute the scheduling instructions.

[0013] As a further aspect of the present invention, the step of analyzing the relationship between the power generation sequence, the adjustable load power sequence, and the interactive constraint upper and lower limit sequences, and extracting the power change trend pattern of the microgrid purchasing or selling electricity to the grid within the dispatch cycle, includes: Establish a real-time net power calculation model for the power generation sequence and the adjustable load power sequence, and generate the net load power sequence; The net load power sequence is compared point by point with the upper and lower limit sequences of the interaction constraint to identify the time intervals in which the net power exceeds the constraint boundary. Within the identified time interval, calculate the degree of deviation of the net power relative to the constraint boundary and the duration of the deviation; Based on the combined characteristics of the degree and duration of deviation, the severity level of power limit exceedance events is classified; Extract the frequency and distribution patterns of power limit exceedance events of different severity levels within each scheduling cycle; By combining the overall fluctuation pattern of the net load power sequence with the characteristics of power over-limit events, a multi-dimensional feature vector representing the power change trend pattern is constructed.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: By processing the extracted key feature vectors through a pre-trained multi-source collaborative representation neural network model, joint state encoding of heterogeneous devices such as distributed power sources, energy storage, and adjustable loads is achieved. This model can autonomously learn from historical and real-time data and capture the dynamic complementary characteristics between devices and the temporal dependencies of system operation, mapping multi-source heterogeneous data into a unified collaborative state vector rich in collaborative information. This overcomes the limitations of traditional methods that rely on manual modeling to describe coupling relationships or simple data overlay, resulting in a more accurate and comprehensive perception of the overall microgrid operating status and providing a high-quality input representation foundation for subsequent optimization decisions.

[0015] A scheduling decision generation network model is adopted, based on a set of cooperative state vectors, for multi-step forward-looking optimization. This model can simulate the optimal action sequence of various controllable devices under complex constraints and their mutual influence over multiple consecutive scheduling periods in the future, generating a coordinated scheduling instruction set in one go. This approach breaks through the limitations of traditional rolling optimization or single-step decision-making frameworks, enabling scheduling instructions to not only consider the current optimal state but also to take into account the evolution path of the system state over a period of time. This effectively avoids problems such as energy storage state runaway and power interaction exceeding limits caused by focusing only on short-term benefits, and significantly improves the overall robustness and economy of microgrids in response to renewable energy fluctuations and load changes. Attached Figure Description

[0016] Figure 1 The flowchart shows the AI-based microgrid multi-source collaborative optimization scheduling method described in this invention. Figure 2 A flowchart for extracting key feature vectors; Figure 3 This is a flowchart for a multi-step forward-looking scheduling simulation. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0019] See Figure 1 This invention provides an AI-based multi-source collaborative optimization scheduling method for microgrids. The method is implemented through the following steps: collecting generation data from distributed power sources, state data from energy storage devices, operational data from adjustable loads, and interaction constraint data from the upper-level power grid within the microgrid system to generate a multi-source heterogeneous microgrid raw operational data set; cleaning and format standardizing the raw operational data set to form a standardized microgrid operational dataset; extracting key feature vectors representing multi-source collaborative relationships from the standardized microgrid operational dataset; inputting the key feature vectors into a pre-trained multi-source collaborative representation neural network model to perform joint state encoding on distributed power sources, energy storage devices, and adjustable loads, generating a collaborative state vector set containing multi-source complementary characteristics and temporal dependencies; based on the collaborative state vector set, calling a scheduling decision generation network model to perform multi-step forward-looking optimization deduction to generate a coordinated scheduling instruction set for multiple future scheduling periods; verifying the security constraints of the coordinated scheduling instruction set, and generating a final executable scheduling control instruction sequence based on the verification results, which is then distributed to various controllable devices in the microgrid system for execution.

[0020] In one embodiment of the present invention, the standardized microgrid operation dataset includes timestamps, power generation sequences of each distributed power source, state-of-charge sequences of energy storage devices, adjustable load power sequences, and upper and lower limit sequences of interactive constraints. Abnormal values ​​in the power generation data, state data, operation data, and interactive constraint data are identified and corrected. Abnormal value identification is based on setting deviation thresholds according to the statistical distribution of historical data, and correction is achieved through interpolation using nearby valid data. The corrected power generation data, state data, operation data, and interactive constraint data are clock-synchronized and resampled according to a unified time base, ensuring that all data sequences have the same time resolution. The clock-synchronized power generation data, state-of-charge data, power data, and upper and lower limit constraint data are normalized and mapped to the same numerical range, thus forming the standardized microgrid operation dataset.

[0021] In practical implementation, taking a microgrid system comprising photovoltaic arrays, wind turbines, lithium-ion battery energy storage systems, and adjustable industrial and commercial loads as an example, this paper describes in detail the process of data cleaning and format unification to form a standardized microgrid operation dataset. The original microgrid operation data set comes from real-time monitoring data uploaded by local controllers of each subsystem and dispatch instructions from the upper-level grid received by the energy management system. In its initial state, the data set exhibits multi-source heterogeneous characteristics. For example, photovoltaic power generation data consists of power records every 5 seconds, wind power generation data consists of power and wind speed records every 10 seconds, energy storage status data consists of state of charge and terminal voltage records every minute, adjustable load data consists of total active power records every 15 seconds, and interactive constraint data consists of power purchase and sale limits updated every 15 minutes.

[0022] A standardized microgrid operation dataset needs to include timestamps, power generation sequences for each distributed power source, state-of-charge sequences for energy storage devices, adjustable load power sequences, and upper and lower limit sequences for interaction constraints. In practice, outliers in the original dataset's power generation, state, operation, and interaction constraint data are identified and corrected. Outlier identification is based on analysis of the statistical distribution of historical long-term operational data for each data point. For photovoltaic power generation sequences, a threshold is set for deviations from the historical average of three standard deviations; for energy storage state-of-charge sequences, a threshold is set for deviations exceeding 0% to 100% of the rated capacity. The correction process uses interpolation of nearby valid data. For example, for a zero-value anomaly in photovoltaic power generation at a certain point in time caused by a communication interruption, the power values ​​of two adjacent valid sampling points are linearly interpolated to replace the abnormal zero value.

[0023] In some embodiments, the corrected power generation data, status data, operational data, and interaction constraint data are clock-synchronized and resampled according to a unified time base. The unified time base adopts the master clock signal of the microgrid energy management system, and all data sources are synchronized with this clock. The resampling process unifies all data sequences to the same time resolution of one data point per minute. For data sequences with an original sampling frequency higher than one minute, downsampling is performed, for example, generating a new sequence by calculating the average value of photovoltaic power data every 5 seconds within each minute; for data sequences with an original sampling frequency lower than one minute, up-interpolation sampling is performed, for example, expanding the interaction constraint upper and lower limit data every 15 minutes to a one-minute sequence by forward padding.

[0024] It is understandable that the power generation data, state of charge data, power data, and upper and lower limit constraint data that have completed clock synchronization are normalized. Normalization maps physical quantities with different dimensions and ranges to the same numerical interval. For example, it divides the power generation data of all distributed generation sources by their respective rated installed capacity, divides the state of charge data of energy storage by 100%, divides the power data of adjustable loads by the historical maximum demand of that load, and divides the upper and lower limit power data of interactive constraints by the maximum transmission capacity of the connection point between the microgrid and the upper-level grid. After the above mapping, all data are transformed into a dimensionless numerical interval between 0 and 1, thus forming a standardized microgrid operation dataset.

[0025] Optionally, during data cleaning, for cases where anomalies occur at multiple consecutive time points, an interpolation method based on a time series prediction model can be used. During resampling, for power-type data, downsampling preferentially uses an averaging algorithm to smooth high-frequency fluctuations, while upsampling preferentially uses a nearest-neighbor hold algorithm to avoid introducing spurious fluctuations. In specific implementation, the benchmark values ​​used for normalization, such as rated capacity or historical extreme values, are stored in the database as system configuration parameters. These can be adjusted according to equipment updates or operating strategies, and the adjusted benchmark values ​​will be applied to the next round of data standardization processing.

[0026] See Figure 2In one embodiment of the present invention, the key feature vector includes source-load power matching characteristics at multiple time scales, energy storage fluctuation mitigation capability characteristics, and power trend characteristics interacting with the power grid. At multiple preset time window scales, the real-time difference between the power generation sequence and the adjustable load power sequence is calculated to form a multi-time-scale source-load power deviation sequence. Based on the state-of-charge sequence and the rated parameters of the energy storage device, a quantitative index characterizing the energy storage system's ability to absorb or release power per unit time to smooth source-load fluctuations is calculated, thus obtaining the energy storage fluctuation mitigation capability characteristics. This study analyzes the relationship between power generation sequences, adjustable load power sequences, and the upper and lower limits of interactive constraints to extract the power change trend patterns of microgrids purchasing or selling electricity to the grid during the dispatch cycle, thus obtaining the power trend characteristics of interaction with the grid. Specifically, this process includes: establishing a real-time net power calculation model for the power generation sequence and adjustable load power sequence to generate the net load power sequence; comparing the net load power sequence with the upper and lower limits of interactive constraints point by point to identify the time intervals in which net power exceeds the constraint boundaries; calculating the degree and duration of deviation of net power relative to the constraint boundaries within the identified time intervals; classifying the severity levels of power exceedance events based on the combined characteristics of the degree and duration of deviation; extracting the frequency and distribution patterns of power exceedance events of different severity levels within each dispatch cycle; and constructing a multi-dimensional feature vector characterizing the power change trend pattern by integrating the overall fluctuation pattern of the net load power sequence and the characteristics of power exceedance events. Finally, a key feature vector is constructed by combining the source-load power deviation sequences at multiple time scales, the fluctuation mitigation capabilities of energy storage, and the power trend characteristics of interaction with the grid.

[0027] In practical implementation, a microgrid comprising photovoltaic power generation units, wind power generation units, battery energy storage systems, and adjustable loads in an industrial park is used as an example scenario. The process of extracting key feature vectors representing multi-source collaborative relationships from a standardized microgrid operation dataset is described in detail. These key feature vectors include source-load power matching characteristics across multiple time scales, energy storage fluctuation mitigation capabilities, and power trend characteristics related to grid interaction.

[0028] In practice, photovoltaic power generation sequences, wind power generation sequences, and adjustable load power sequences from industrial parks are obtained from standardized microgrid operation datasets. Real-time differences between the power generation sequences and adjustable load power sequences are calculated at multiple preset time window scales, including 5 minutes, 15 minutes, and 60 minutes. For a 5-minute time window, the average of the total power generation and total load power for the current moment and the preceding 5 data points is calculated, and the difference between the two averages is calculated to form a data point in the 5-minute source-load power deviation sequence. This sliding window calculation is repeated to form a complete 5-minute source-load power deviation sequence. For 15-minute and 60-minute time windows, the same sliding average and difference calculation method is used to form 15-minute and 60-minute source-load power deviation sequences, respectively, which together constitute a multi-time-scale source-load power deviation sequence.

[0029] In some embodiments, the energy storage fluctuation smoothing capability characteristics are calculated based on the state-of-charge (SOC) sequence of energy storage devices and the rated parameters of the energy storage devices in a standardized microgrid operation dataset. The rated parameters include the rated power and rated capacity of the energy storage system. The energy storage fluctuation smoothing capability characteristic is a quantitative indicator characterizing the ability of the energy storage system to absorb or release power per unit time to smooth source-load fluctuations. This indicator is calculated using a function that considers the current SOC, rated power, and time scale. Calculation of Energy Storage Fluctuation Smoothing Capability Characteristics The formula is as follows: in: Indicates at time The quantitative value of the energy storage's ability to smooth out fluctuations. Indicates the rated power of the energy storage device. Represents the standardized time. The state of charge, This represents a curvature adjustment coefficient that is greater than zero. The larger the value, the stronger the power throughput capability of the energy storage at that moment to smooth out fluctuations.

[0030] It is understandable that analyzing the relationship between the power generation sequence, the adjustable load power sequence, and the upper and lower limits of the interaction constraint sequence is used to extract the power trend characteristics interacting with the power grid. A real-time net power calculation model is established for the power generation sequence and the adjustable load power sequence. The net power calculation model is defined as the total power generation minus the total load power, generating the net load power sequence. The net load power sequence is compared point-by-point with the upper and lower limits of the interaction constraint sequence to identify the time intervals in which the net power exceeds the constraint boundaries. Within the identified time intervals in which the net power exceeds the constraint boundaries, the degree of deviation of the net power relative to the constraint boundaries and the duration of the deviation are calculated. Based on the combined characteristics of the degree of deviation and the duration of the deviation, the severity level of the power exceedance event is classified. For example, a deviation of less than 10% of the constraint range and a duration of less than 5 minutes is defined as a minor exceedance, and a deviation of more than 30% of the constraint range or a duration of more than 15 minutes is defined as a severe exceedance.

[0031] Optionally, the frequency and distribution patterns of power exceedance events of different severity levels within each scheduling cycle are extracted. A multi-dimensional feature vector characterizing the power change trend pattern is constructed by integrating the overall fluctuation pattern of the net load power sequence with the characteristics of power exceedance events. Fluctuation pattern features include the mean, standard deviation, skewness, and zero-crossing rate of net load power within the scheduling cycle. Power exceedance event features include the number of minor exceedance events, the number of severe exceedance events, the average exceedance duration, and the maximum exceedance power ratio. These scalar features are combined to form a multi-dimensional feature vector, serving as the power trend feature interacting with the grid. Finally, the source-load power deviation sequences at multiple time scales, the energy storage fluctuation mitigation capability characteristics, and the power trend features interacting with the grid are concatenated along the feature dimensions to construct a key feature vector for subsequent model input.

[0032] In one embodiment of the present invention, the multi-source collaborative representation neural network model includes a feature interaction coding layer and a temporal dependency coding layer. The feature interaction coding layer receives key feature vectors and calculates the dynamic correlation weights between different distributed power sources, between distributed power sources and energy storage, and between source, storage, and load through its internal multi-head attention mechanism, generating preliminary fused features. The temporal dependency coding layer receives the preliminary fused features and captures the evolution of the preliminary fused features over time through its internal recurrent neural network structure, outputting a feature sequence containing temporal context information. The feature sequence containing temporal context information is pooled and compressed to generate a fixed-dimensional state vector for each scheduling moment, and the state vectors of all scheduling moments constitute a collaborative state vector set.

[0033] In practical implementation, a microgrid system with an operating cycle of 24 hours and a scheduling interval of 15 minutes is used as an example scenario to describe in detail the process of inputting key feature vectors into a pre-trained multi-source collaborative representation neural network model and generating a set of collaborative state vectors. The key feature vectors serve as the input to the multi-source collaborative representation neural network model, which has been trained using a large amount of historical operating data before deployment. The model includes a feature interaction coding layer and a temporal dependency coding layer.

[0034] In practical implementation, the feature interactive coding layer receives an input sequence composed of key feature vectors from multiple consecutive scheduling moments. The feature interactive coding layer incorporates a multi-head attention mechanism, which transforms each input key feature vector into a query vector, a key vector, and a value vector. In practice, by calculating the dot product correlation between the query vector and the key vector among different key feature vector components, a weight matrix reflecting the dynamic relationships between different distributed power sources, between distributed power sources and energy storage, and between source, storage, and load is obtained. The calculation process of the correlation weights can be expressed by the following formula: in: Indicates the first In the attention head, the first The eigencomponent and the ... Attention weights between feature components Indicates the first The eigencomponent in the th th eigenvalue 1 The query vector corresponding to each attention head. Indicates the first The eigencomponent in the th th is the ? The key vector corresponding to each attention head This represents the dimension of the key vector. The value vectors are weighted and summed based on attention weights. Then, the outputs of all attention heads are concatenated and linearly transformed to generate preliminary fusion features containing rich interactive information.

[0035] In some embodiments, the temporal dependency coding layer receives the preliminary fused feature sequence output by the feature interaction coding layer. Internally, the temporal dependency coding layer employs a gated recurrent unit (GRU) structure, which processes the preliminary fused feature sequence sequentially over time. For each scheduled moment's preliminary fused features, the GRU structure, combined with its internally stored historical hidden states, calculates the update and reset gate signals for the current moment, thereby determining how much historical information to retain and how to fuse the current input information into the new state. The operation of the GRU structure enables the temporal dependency coding layer to capture the evolution and long-term dependencies of the preliminary fused features over time, outputting a feature sequence containing temporal context information.

[0036] It can be understood that pooling and compression operations are performed on the feature sequence containing temporal context information to generate a fixed-dimensional state vector for each scheduling moment. The pooling operation uses max pooling to extract the most salient feature activation values ​​from the feature vector corresponding to each scheduling moment. The compression operation maps the pooled feature vector to a pre-defined, lower-dimensional vector space through a fully connected layer. For a 24-hour period containing 96 scheduling moments, after the above processing, an ordered set of 96 fixed-dimensional state vectors is finally generated; this ordered set is the co-state vector set.

[0037] Optionally, the number of heads in the multi-head attention mechanism of the feature interaction coding layer is set to 8, and the key vector dimension of each head is set to 64. The number of hidden layer units in the gated recurrent unit structure of the temporal dependency coding layer is set to 256. The fixed dimension of each state vector in the cooperative state vector set is set to 128. These network structure parameters are determined during the model training phase and remain unchanged after model deployment. The pre-trained multi-source cooperative representation neural network model parameters are stored in the computing unit of the microgrid energy management system in the form of files, and are loaded and the above coding process is executed at the beginning of each scheduling cycle.

[0038] See Figure 3In one embodiment of the present invention, the coordinated scheduling instruction set includes planned output instructions for each distributed power source, charging and discharging power instructions for energy storage devices, and control instructions for adjustable loads. A set of cooperative state vectors is input into a scheduling decision generation network model, which is constructed based on a deep reinforcement learning framework. The construction method includes: defining the state space of the scheduling decision generation network model as the space generated by the set of cooperative state vectors; defining the action space of the scheduling decision generation network model as the allowable adjustment range of each controllable device in a single scheduling period; defining a reward function, which is used to evaluate temporary scheduling actions in real time during the simulation process, and its calculation is based on the source-load matching degree, energy storage health status, and over-limit penalties for interaction with the grid under the simulated state; and training the scheduling decision generation network model offline and fine-tuning it online using a large amount of historical and simulation data, enabling it to learn to select the scheduling action sequence that maximizes the cumulative reward in the state space. The scheduling decision generation network model takes the current and historical cooperative state vectors as input to simulate the dynamic evolution of the microgrid system in a preset multi-step scheduling time domain. In each step of the evolution simulation, the scheduling decision generation network model outputs the temporary scheduling actions corresponding to the scheduling time. The temporary scheduling actions include the proposed power of various distributed power sources, the proposed charging and discharging power of energy storage devices, and the proposed adjustment amount of adjustable loads. The model collects the temporary scheduling actions at all scheduling times during the simulation process and, combined with the evaluation of the overall system operating status at the end of the simulation, optimizes and adjusts the sequence of temporary scheduling actions to form the final coordinated scheduling instruction set. The safety constraint verification of the coordinated dispatch instruction set includes: obtaining the upper and lower limits of output and ramp rate limits of each distributed power source, the safe range of state of charge and charging / discharging power limits of energy storage devices, the adjustment range and adjustment rate limits of adjustable loads, and the power exchange limits at the connection point between the microgrid and the upper-level grid from the standardized microgrid operation dataset, as a physical safety constraint set; comparing the planned output instructions, charging / discharging power instructions, and control instructions in the coordinated dispatch instruction set with the physical safety constraint set on a time-by-time basis; checking whether the planned output instructions meet the upper and lower limits of output and ramp rate limits, checking whether the charging / discharging power instructions cause the state of charge of energy storage to exceed the safe range or the charging / discharging power to exceed the limit, checking whether the control instructions exceed the adjustment range and rate limits of adjustable loads, and checking whether the net exchange power with the grid exceeds the exchange limit at the connection point; recording all instructions that do not meet the physical safety constraint set and the specific constraint content they violate, and generating a constraint violation report.

[0039] In practical implementation, taking a microgrid optimization scheduling scenario with a scheduling time domain of 4 hours in the future and a scheduling interval of 15 minutes as an example, this paper describes in detail the process of calling the scheduling decision generation network model to perform multi-step forward-looking optimization simulation, generating a coordinated scheduling instruction set, and performing safety constraint verification. The set of cooperative state vectors is used as the input to the scheduling decision generation network model, which is built based on a deep reinforcement learning framework. Its construction includes the definition of the state space, action space, and reward function. The state space of the scheduling decision generation network model is defined as the space composed of the set of cooperative state vectors, that is, each state is a fixed-dimensional vector in the set of cooperative state vectors. The action space of the scheduling decision generation network model is defined as the allowable adjustment range of each controllable device in a single scheduling period. The action vectors specifically include the planned output adjustment of photovoltaic power generation units, the planned output adjustment of wind power generation units, the charging and discharging power command values ​​of energy storage devices, and the power regulation command values ​​of adjustable loads. Each dimension of the action value has its corresponding physical upper and lower limits. The reward function is used to evaluate temporary scheduling actions in real time during the simulation simulation. The calculation is based on the source-load matching degree, energy storage health, and over-limit penalty for interaction with the grid under simulated conditions, and its expression is: in: Indicates the instant reward value. The score represents the evaluation score for the source-load power matching degree. This indicates the standardized state of charge of the energy storage device. This represents the penalty value for exceeding the power exchange limit at the connection point between the microgrid and the upper-level grid. , , These are the weight coefficients for each item. The scheduling decision-generating network model is trained offline and fine-tuned online using a large amount of historical and simulation data. The training objective is to enable the model to learn to select the scheduling action sequence that maximizes the cumulative reward over multiple future steps in the state space.

[0040] In some embodiments, the scheduling decision generation network model takes the current and historical coordinated state vectors as input to simulate the dynamic evolution of a microgrid system over 16 preset scheduling periods. The simulation evolution is based on a simplified microgrid operation simulation environment that updates the simulated system state according to scheduling actions. In each step of the evolution simulation, the scheduling decision generation network model outputs temporary scheduling actions corresponding to the current simulation state. These temporary scheduling actions include the proposed power output for photovoltaic power generation units, the proposed power output for wind power generation units, the proposed charge / discharge power for energy storage devices, and the proposed adjustment amount for adjustable loads. All temporary scheduling actions from the 16 scheduling periods during the simulation are collected to form a temporary scheduling action sequence. Combined with the evaluation of the overall system operating state at the end of the simulation (evaluation indicators include cumulative reward, final state of charge of energy storage, and total over-limit penalty), the temporary scheduling action sequence is optimized and adjusted. The adjustment method includes small-scale iterative corrections of some action values ​​in the sequence to improve the cumulative reward, ultimately forming a coordinated scheduling instruction set containing 16 sets of instructions.

[0041] It is understandable that verifying the security constraints of the coordinated dispatch instruction set requires obtaining the physical security constraint set. The operating limit parameters of each device are parsed and extracted from the standardized microgrid operation dataset to form the physical security constraint set. Refer to Table 1 for the specific contents of the physical security constraint set.

[0042] Table 1: Set of Physical Security Constraints The planned output commands, charging / discharging power commands, and control commands in the coordinated dispatch command set are compared with the physical safety constraint set on a time-by-time basis. This includes checking whether the planned output commands of photovoltaic and wind power generation units meet their respective output upper and lower limits and ramp-up rate limits; checking whether the charging / discharging power commands of energy storage devices cause the energy storage state of charge to exceed the safe range after simulation calculations or whether the charging / discharging power itself exceeds the limit; checking whether the control commands of adjustable loads exceed their adjustment range and adjustment rate limits; and checking whether the net exchange power between the microgrid and the grid exceeds the connection point exchange limit.

[0043] Optionally, all instructions that do not meet the physical safety constraint set and the specific constraints they violate are recorded, generating a constraint violation report. The constraint violation report is in list format, with each record containing the scheduling period number of the violation, the name of the controlled device, the instruction content, the type of constraint violated, and the constraint limit. For example, a constraint violation report might record "Period 5, Energy Storage Device, Charging Power Instruction 450kW, Violation of Charging / Discharging Power Upper Limit Constraint (400kW)" and "Period 10, Photovoltaic Power Generation Unit, Planned Output Instruction 550kW, Violation of Output Upper Limit Constraint (500kW)". The constraint violation report will serve as input for subsequent instruction correction processes.

[0044] In one embodiment of the present invention, if the constraint violation report is empty, it indicates that the coordinated scheduling instruction set fully satisfies all physical security constraints, and it is directly arranged into a scheduling control instruction sequence in chronological order. If the constraint violation report is not empty, the instruction correction process is initiated: for each instruction that violates the constraint in the report, under the premise of satisfying its original optimization objective, it is corrected to satisfy the corresponding physical security constraint using the principle of minimum adjustment; the corrected instructions are reintegrated with the original instructions that do not violate the constraint to form a corrected scheduling instruction set that satisfies all security constraints; the instructions in the corrected scheduling instruction set are checked for temporal and logical consistency to ensure that there are no temporal or logical conflicts between the instructions, and after passing the check, they are arranged in chronological order to generate the final executable scheduling control instruction sequence. The various controllable devices in the microgrid system execute the following steps: encapsulate the sequence of dispatch control commands into data packets of a specified format according to the communication protocol; send the encapsulated data packets to the corresponding distributed power controllers, energy storage system managers, and load aggregation controllers through the communication network of the microgrid energy management system; the distributed power controllers, energy storage system managers, and load aggregation controllers receive and parse the data packets, extract their respective control commands, and convert them into drive signals or switching control signals for the underlying power electronic devices to realize the execution of the dispatch commands.

[0045] In practical implementation, following the constraint violation report generated in the previous example, a coordinated scheduling instruction set containing 16 scheduling periods is used as the processing object. The process of generating the final executable sequence of scheduling control instructions based on the verification results and issuing them for execution is described in detail. If the constraint violation report is empty, it indicates that the coordinated scheduling instruction set fully satisfies all physical security constraints. In this case, the instructions in the coordinated scheduling instruction set are directly arranged into a scheduling control instruction sequence in chronological order. If the constraint violation report is not empty, the instruction correction process is initiated, which applies to each instruction recorded in the report that violates a physical security constraint.

[0046] In practical implementation, the instruction correction process follows the principle of minimum adjustment, that is, under the premise of satisfying the original optimization objective of the instruction, the instruction is corrected to meet the corresponding physical safety constraints with the smallest adjustment range. For an output or charging power instruction that violates the upper limit constraint, its correction value is set to the upper limit value; for an output or discharging power instruction that violates the lower limit constraint, its correction value is set to the lower limit value. For instructions that violate the ramp rate constraint, the correction needs to consider the final instruction value of the previous scheduling period. If the change in instruction value exceeds the positive ramp limit, the instruction value is corrected to the previous instruction value plus the positive ramp limit; if the change in instruction value is lower than the negative ramp limit, the instruction value is corrected to the previous instruction value minus the negative ramp limit. For cases where the state of charge exceeds the safe range after simulation calculation due to the energy storage charging and discharging power instruction, the correction process needs to iteratively adjust the charging and discharging power instructions of the current period and adjacent periods to ensure that the state of charge trajectory is within the safe range. The objective function is to minimize the total adjustment amount of the original instruction sequence. This optimization problem can be expressed as: in: This represents the set of relevant time periods that need to be adjusted. Indicates time period The revised energy storage charging and discharging power command values, Indicates time period The original energy storage charging and discharging power command value is constrained by the fact that the energy storage state of charge obtained by simulation calculation of the modified command sequence is within a safe range throughout the entire process.

[0047] In some embodiments, the modified instructions are reintegrated with the original instructions that do not violate physical security constraints to form a modified scheduling instruction set that satisfies all security constraints. A timing and logic consistency check is performed on the instructions in the modified scheduling instruction set to ensure that there are no temporal or logical conflicts between instructions. The check includes whether instructions from the same device at different time periods meet the temporal ramp-up or adjustment rate continuity requirements, and whether the superposition of instructions from different devices at the same time period leads to new constraint violations. After passing the timing and logic consistency check, the instructions in the modified scheduling instruction set are arranged in chronological order to generate a final executable scheduling control instruction sequence.

[0048] It is understandable that the sequence of dispatch control commands is sent to various controllable devices in the microgrid system for execution. The sending process first encapsulates the sequence of dispatch control commands into data packets of a specified format according to a communication protocol, which adopts the IEC 61850 standard. Through the communication network of the microgrid energy management system, the encapsulated data packets are sent to the corresponding distributed generation controllers, energy storage system managers, and load aggregation controllers. The distributed generation controllers receive and parse the data packets, extract the planned output control commands corresponding to the photovoltaic or wind power generation units, and convert these commands into power setpoint signals for the inverters. The energy storage system managers receive and parse the data packets, extract the charging and discharging power commands for the energy storage devices, and convert these commands into power control signals for the battery management system and power conversion system. The load aggregation controllers receive and parse the data packets, extract the regulation commands for adjustable loads, and convert these commands into control signals for load switches or power regulation signals for flexible loads.

[0049] Optionally, the data packets encapsulated by the communication protocol in a specified format include a frame header, device address code, instruction function code, instruction data area, checksum, and frame trailer. The communication network of the microgrid energy management system adopts a hybrid topology of industrial Ethernet and fieldbus to ensure the real-time performance and reliability of data transmission. After receiving the data packets and completing instruction parsing, the distributed power controller, energy storage system manager, and load aggregation controller return an acknowledgment signal confirming execution to the microgrid energy management system.

[0050] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An AI-based multi-source collaborative optimization scheduling method for microgrids, characterized in that, The method includes: Collect power generation data from distributed power sources, status data from energy storage devices, operational data from adjustable loads, and interaction constraint data from the upper-level power grid in the microgrid system to generate a multi-source heterogeneous microgrid raw operational data set; The original microgrid operation data set is cleaned and format-unified to form a standardized microgrid operation dataset. Key feature vectors representing multi-source collaborative relationships are extracted from the standardized microgrid operation dataset. The key feature vectors are input into a pre-trained multi-source collaborative representation neural network model to perform joint state encoding on the distributed power source, the energy storage device, and the adjustable load, generating a set of collaborative state vectors that includes multi-source complementary characteristics and time-series dependencies. Based on the aforementioned set of collaborative state vectors, the scheduling decision generation network model is invoked to perform multi-step forward-looking optimization and deduction, generating a set of coordinated scheduling instructions for multiple future scheduling periods; The coordination and scheduling instruction set is subjected to security constraint verification, and a final executable scheduling control instruction sequence is generated based on the verification results and sent to various controllable devices in the microgrid system for execution.

2. The AI-based microgrid multi-source collaborative optimization scheduling method according to claim 1, characterized in that, The original microgrid operation data set is cleaned and formatted to form a standardized microgrid operation dataset, including: The standardized microgrid operation dataset includes timestamps, power generation sequences of each distributed power source, state of charge sequences of energy storage devices, power sequences of adjustable loads, and upper and lower limits sequences of interactive constraints. Abnormal values ​​in the power generation data, status data, operation data, and interaction constraint data are identified and corrected. The identification of abnormal values ​​is based on setting a deviation threshold according to the statistical distribution of historical data, and the correction is completed by interpolation of nearby valid data. The corrected power generation data, status data, operation data, and interaction constraint data are clock-synchronized and resampled according to a unified time base so that all data sequences have the same time resolution; The power generation data, state of charge data, power data, and upper and lower limit constraint data that have completed clock synchronization are normalized and mapped to the same numerical range to form the standardized microgrid operation dataset.

3. The AI-based multi-source collaborative optimization scheduling method for microgrids according to claim 2, characterized in that, Key feature vectors characterizing multi-source collaborative relationships are extracted from the standardized microgrid operation dataset, including: The key feature vectors include source-load power matching characteristics at multiple time scales, energy storage fluctuation mitigation capabilities, and power trend characteristics in interaction with the power grid. Under multiple preset time window scales, the real-time difference between the power generation sequence and the adjustable load power sequence is calculated to form a source-load power deviation sequence with multiple time scales; Based on the state of charge sequence and the rated parameters of the energy storage device, a quantitative index characterizing the energy storage system's ability to absorb or release power per unit time to smooth source-load fluctuations is calculated, thus obtaining the energy storage fluctuation smoothing capability characteristics. By analyzing the relationship between the power generation sequence, the adjustable load power sequence, and the upper and lower limit sequences of the interaction constraints, the power change trend pattern of the microgrid purchasing or selling electricity to the grid during the dispatch cycle is extracted, and the power trend characteristics of the interaction with the grid are obtained. The key feature vector is constructed by combining the source-load power deviation sequence at multiple time scales, the energy storage fluctuation smoothing capability characteristics, and the power trend characteristics interacting with the power grid.

4. The AI-based microgrid multi-source collaborative optimization scheduling method according to claim 3, characterized in that, The key feature vectors are input into a pre-trained multi-source collaborative representation neural network model to perform joint state encoding on the distributed power source, the energy storage device, and the adjustable load, generating a set of collaborative state vectors containing multi-source complementary characteristics and temporal dependencies, including: The multi-source collaborative representation neural network model includes a feature interaction coding layer and a temporal dependency coding layer; The feature interaction coding layer receives the key feature vector and calculates the dynamic correlation weights between different distributed power sources, between distributed power sources and energy storage, and between source, storage and load through an internal multi-head attention mechanism to generate preliminary fusion features. The temporal-dependent coding layer receives the preliminary fused features, captures the evolution of the preliminary fused features in the time dimension through its internal recurrent neural network structure, and outputs a feature sequence containing temporal context information. The feature sequence containing temporal context information is pooled and compressed to generate a fixed-dimensional state vector for each scheduling moment. The state vectors of all scheduling moments constitute the collaborative state vector set.

5. The AI-based microgrid multi-source collaborative optimization scheduling method according to claim 4, characterized in that, Based on the aforementioned set of cooperative state vectors, a scheduling decision generation network model is invoked to perform multi-step forward-looking optimization and deduction, generating a set of coordinated scheduling instructions for multiple future scheduling periods, including: The coordinated scheduling instruction set includes planned output instructions for each distributed power source, charging and discharging power instructions for energy storage devices, and control instructions for adjustable loads. The set of cooperative state vectors is input into the scheduling decision generation network model, which is constructed based on a deep reinforcement learning framework. The scheduling decision generation network model takes the current and historical cooperative state vectors as input to simulate the dynamic evolution process of the microgrid system in a preset multi-step scheduling time domain. In each step of the evolution simulation, the scheduling decision generation network model outputs the temporary scheduling action corresponding to the scheduling time. The temporary scheduling action includes the simulated power of various distributed power sources, the simulated charging and discharging power of energy storage devices, and the simulated adjustment amount of adjustable loads. Collect all temporary scheduling actions during the simulation process, and combine them with the evaluation of the overall system operation status at the end of the simulation to optimize and adjust the sequence of temporary scheduling actions, forming the final coordinated scheduling instruction set.

6. The AI-based multi-source collaborative optimization scheduling method for microgrids according to claim 5, characterized in that, The scheduling decision generation network model is built based on a deep reinforcement learning framework and includes: The state space of the scheduling decision generation network model is defined as the space spanned by the set of cooperative state vectors; The action space of the scheduling decision generation network model is defined as the allowable adjustment range of each controllable device in a single scheduling period. A reward function is defined, which is used to evaluate the temporary scheduling action in real time during the simulation. Its calculation is based on the source-load matching degree, energy storage health degree and limit violation penalty of interaction with the grid under the simulation state. The scheduling decision generation network model is trained offline and fine-tuned online using a large amount of historical and simulation data, enabling it to learn to select the scheduling action sequence that maximizes the cumulative reward in the state space.

7. The AI-based multi-source collaborative optimization scheduling method for microgrids according to claim 6, characterized in that, The security constraint verification of the coordinated scheduling instruction set includes: The output limits and ramp rate limits of each distributed power source, the safe range of the state of charge and the charging and discharging power limits of the energy storage device, the adjustment range and adjustment rate limits of the adjustable load, and the power exchange limits of the connection point between the microgrid and the upper-level grid are obtained from the standardized microgrid operation dataset and used as a set of physical safety constraints. The planned output instructions, charging and discharging power instructions, and control instructions in the coordination and scheduling instruction set are compared with the physical safety constraint set in each time period. Check whether the planned output command meets the upper and lower limits of output and the ramp rate limit; check whether the charging and discharging power command causes the energy storage state of charge to exceed the safe range or the charging and discharging power to exceed the limit; check whether the control command exceeds the adjustment range and rate limit of the adjustable load; and check whether the net exchange power with the grid exceeds the connection point exchange limit. Record all instructions that do not meet the set of physical security constraints and the specific constraints they violate, and generate a constraint violation report.

8. The AI-based microgrid multi-source collaborative optimization scheduling method according to claim 7, characterized in that, The step of generating the final executable sequence of scheduling control instructions based on the verification results includes: If the constraint violation report is empty, it indicates that the coordinated scheduling instruction set fully satisfies all physical security constraints, and it is directly arranged into the scheduling control instruction sequence in chronological order; If the constraint violation report is not empty, the instruction correction process is initiated: for each instruction in the report that violates the constraint, under the premise of satisfying its original optimization objective, it is corrected to satisfy the corresponding physical security constraint using the principle of minimum adjustment. The revised instructions are then reintegrated with the original instructions that do not violate constraints to form a revised set of scheduling instructions that satisfies all safety constraints. The instructions in the modified scheduling instruction set are subjected to a timing and logic consistency check to ensure that there are no time or logical conflicts between the instructions. After passing the check, the instructions are arranged in chronological order to generate the final executable scheduling control instruction sequence.

9. The AI-based multi-source collaborative optimization scheduling method for microgrids according to claim 8, characterized in that, The execution of the commands sent to various controllable devices in the microgrid system includes: The sequence of scheduling control instructions is encapsulated into a data packet of a specified format according to the communication protocol; Through the communication network of the microgrid energy management system, the encapsulated data packets are sent to the corresponding distributed power controller, energy storage system manager and load aggregation controller respectively; The distributed power controller, energy storage system manager, and load aggregation controller receive and parse the data packet, extract their respective control instructions, and convert them into drive signals or switching control signals for the underlying power electronic devices to execute the scheduling instructions.

10. The AI-based multi-source collaborative optimization scheduling method for microgrids according to claim 9, characterized in that, The analysis of the relationship between the power generation sequence, the adjustable load power sequence, and the upper and lower limit sequences of the interaction constraints extracts the power change trend pattern of the microgrid's purchase or sale of electricity to the grid within the dispatch cycle, including: Establish a real-time net power calculation model for the power generation sequence and the adjustable load power sequence, and generate the net load power sequence; The net load power sequence is compared point by point with the upper and lower limit sequences of the interaction constraint to identify the time intervals in which the net power exceeds the constraint boundary. Within the identified time interval, calculate the degree of deviation of the net power relative to the constraint boundary and the duration of the deviation; Based on the combined characteristics of the degree and duration of deviation, the severity level of power limit exceedance events is classified; Extract the frequency and distribution patterns of power limit exceedance events of different severity levels within each scheduling cycle; By combining the overall fluctuation pattern of the net load power sequence with the characteristics of power over-limit events, a multi-dimensional feature vector representing the power change trend pattern is constructed.

Images