Micro-grid self-regulating energy management system and method combined with AI and intelligent terminal

Abstract

The application discloses a micro-grid self-adjusting energy management system and method combined with AI and an intelligent terminal, and relates to the technical field of power system automation. The system comprises: a real-time information acquisition module for acquiring real-time operation state information of a target micro-grid and corresponding charging station information; an expected load prediction module for generating expected load prediction information of the target micro-grid within a future preset management window based on the real-time operation state information and the charging station information, in combination with a pre-constructed expected load prediction model; an adjustment margin optimization module for maximizing the adjustment capacity margin of the target micro-grid within a future preset period as a target, establishing an optimization task and solving it, and outputting a power limit value corresponding to each charging interface; and a charging load control module for distributing and managing the charging load of the target micro-grid according to the power limit value. The application effectively improves the load prediction accuracy and enhances the operation stability of the micro-grid self-adjustment.

Description

Technical Field

[0001] This invention relates to the field of power system automation technology, specifically to a microgrid self-regulating energy management system, method, and intelligent terminal that incorporates AI. Background Technology

[0002] With the widespread application of distributed photovoltaic, energy storage, and electric vehicle charging piles in microgrids, charging stations, as a new type of load with significant power fluctuations, are crucial to the stability and efficiency of microgrids. Existing technologies typically employ statistical methods based on historical data for load forecasting and use preset rules or static optimization tasks to schedule charging power in order to achieve a basic supply-demand balance.

[0003] However, the static optimization tasks employed by traditional technologies lack flexibility and are difficult to accurately adapt to the real-time fluctuations and uncertainties on the generation and load sides, resulting in low load forecast accuracy and failing to meet the needs of dynamic microgrid management. At the same time, existing optimization tasks mostly aim to minimize short-term economic costs, neglecting to ensure the margin of microgrid regulation capacity. This makes microgrids prone to power imbalance, equipment overload, and other safety hazards when charging load suddenly surges, making it difficult to achieve stable and reliable self-regulating operation. Summary of the Invention

[0004] This invention provides a microgrid self-regulating energy management system, method, and intelligent terminal that incorporates AI, aiming to solve the technical problems of insufficient load forecasting accuracy and lack of regulation stability in the prior art.

[0005] In view of the above problems, the present invention provides a microgrid self-regulating energy management system, method and intelligent terminal that incorporates AI.

[0006] In a first aspect, the present invention provides a microgrid self-regulating energy management system incorporating AI, comprising:

[0007] The real-time information acquisition module is used to acquire the real-time operating status information of the target microgrid and the corresponding charging station information;

[0008] The expected load forecasting module is used to generate expected load forecasting information for the target microgrid within a future preset management window based on the real-time operating status information and the charging station information, combined with a pre-built expected load forecasting model.

[0009] The regulation margin optimization module is used to establish and solve an optimization task with the goal of maximizing the regulation margin of the target microgrid in a future preset time period, and output the power limit value corresponding to each charging interface.

[0010] The charging load management module is used to allocate and manage the charging load of the target microgrid according to the power limit value.

[0011] Secondly, this invention provides a microgrid self-regulating energy management method incorporating AI, including:

[0012] Obtain real-time operating status information of the target microgrid and corresponding charging station information;

[0013] Based on the real-time operating status information and the charging station information, combined with the pre-constructed expected load prediction model, expected load prediction information of the target microgrid within a future preset management window is generated.

[0014] With the goal of maximizing the regulation margin of the target microgrid in a future preset period, an optimization task is established and solved, and the power limit value corresponding to each charging interface is output.

[0015] The charging load of the target microgrid is allocated and managed according to the power limit value.

[0016] Thirdly, the present invention provides an AI-integrated microgrid self-regulating energy management intelligent terminal, and the AI-integrated microgrid self-regulating energy management system and method provided by the present invention are applied to the intelligent terminal.

[0017] One or more technical solutions provided in this invention have at least the following technical effects or advantages:

[0018] This invention provides an AI-integrated microgrid self-regulating energy management system, method, and intelligent terminal. By accurately acquiring full real-time data on microgrid operation and charging stations, it provides a reliable data foundation for subsequent analysis. Combined with AI models, it achieves high-precision prediction of future loads, overcoming the limitations of traditional statistical methods. With the goal of maximizing the microgrid's regulation capacity margin, it intelligently solves power limit problems, constructing a buffer barrier to cope with load fluctuations. Furthermore, through charging load management, it achieves precise allocation and dynamic management of charging power. This system effectively improves the accuracy of microgrid load prediction, enhances the stability and reliability of microgrid self-regulating operation, fundamentally solves problems such as power imbalance and equipment overload caused by sudden changes in charging load, and realizes the coordinated and efficient operation of the microgrid and charging stations. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0020] Figure 1 This is a schematic diagram of the structure of a microgrid self-regulating energy management system incorporating AI, provided in an embodiment of the present invention.

[0021] Figure 2 A flowchart illustrating the microgrid self-regulating energy management method incorporating AI, provided in an embodiment of the present invention;

[0022] The components represented by each number in the attached diagram are explained below:

[0023] Real-time information acquisition module 11, expected load prediction module 12, adjustment margin optimization module 13, and charging load management module 14. Detailed Implementation

[0024] This invention provides a microgrid self-regulating energy management system, method, and intelligent terminal that incorporates AI, addressing the technical problems of insufficient load forecasting accuracy and lack of regulation stability in existing technologies.

[0025] Example 1, as Figure 1 As shown, this invention provides a microgrid self-regulating energy management system incorporating AI, comprising:

[0026] The real-time information acquisition module 11 is used to acquire the real-time operating status information of the target microgrid and the corresponding charging station information.

[0027] In this embodiment of the invention, real-time operating status information of the target microgrid and corresponding charging station information are acquired. A microgrid is a small-scale power generation and distribution system composed of distributed power sources, energy storage devices, energy conversion devices, loads, monitoring and protection devices, etc., designed to enable flexible and efficient application of distributed power sources, solve the grid connection problem of distributed power sources, and possess islanded operation capability. If microgrid operating status data is missing, the predicted load value may deviate significantly from the actual value, leading to unreasonable optimized power limits. If charging station information is not collected meticulously, charging piles may be misjudged as vacant or missed for scheduled charging, ultimately causing charging queues and microgrid overload. Therefore, real-time collection of microgrid and charging station status data is necessary to provide reliable data support for subsequent operations and is a prerequisite for ensuring the accurate operation of the entire energy management system.

[0028] In one embodiment, the real-time information acquisition module 11 is further configured to: include at least real-time power output data, tie line limit data, and inherent load data in the real-time operating status information; and include at least real-time access status information and access object plan information for each charging interface.

[0029] First, acquire the real-time operating status information of the target microgrid. This real-time operating status information includes at least real-time power output data, tie-line limit data, and inherent load data. Real-time power output data refers to the actual output power of distributed power sources within the microgrid at the current moment, reflecting the microgrid's own power supply capacity. Tie-line limit data refers to the maximum power limit allowed to be transmitted through the tie-line between the microgrid and the main grid, which is a key threshold constraining power interaction between the microgrid and the main grid. Inherent load data refers to the real-time power consumption of non-charging loads within the microgrid, including power data of stable power loads such as factory production equipment in industrial parks, office building lighting, and central air conditioning.

[0030] Specifically, the data collection targets and equipment are first determined. For real-time power output data, smart meters are deployed at the photovoltaic array combiner boxes and energy storage converters. For tie-line limit data, preset tie-line power thresholds are obtained from the microgrid energy management platform. For inherent load data, power monitoring terminals are deployed in each distribution room of the park. Secondly, the data collection frequency is set to 5 minutes per instance based on the microgrid's operating characteristics, ensuring the data's real-time performance meets the time accuracy requirements for subsequent load forecasting. Finally, power line carrier communication technology is used to transmit data from each collection device to the real-time information acquisition module 11. The collected data undergoes preprocessing such as outlier removal and unit standardization to generate a standardized real-time operating status information dataset.

[0031] For example, the target microgrid is an integrated photovoltaic-storage-charging microgrid in an industrial park. This microgrid includes a 3MW photovoltaic array, a 2MWh energy storage system, and a 10kV tie line. The inherent load consists of the electricity load of two production plants and one office building within the park. Real-time output data acquisition: The photovoltaic array output is 1.8MW at the photovoltaic combiner box, and the energy storage system is in a discharging state with an output of 0.5MW. Tie line limit data acquisition: The upper limit of the tie line power is obtained from the microgrid EMS, which is ±1.2MW. Inherent load data acquisition: The current electricity load of the production plants is 1.5MW, and the electricity load of the office building is 0.3MW, with a total inherent load of 1.8MW, collected through the distribution room monitoring terminal.

[0032] Secondly, charging station information is acquired. This information includes at least the real-time access status information and access target plan information for each charging interface. The real-time access status information refers to the current operating status of each charging interface in the charging station, including four states: idle, charging, fault, and reserved lock. This is the direct basis for determining whether a charging interface is available. The access target plan information refers to the preset charging plan of the electrical equipment connected to the charging interface, including key parameters such as planned charging time, planned charging power, and planned charging amount. First, status sensors are deployed at each charging interface in the charging station to monitor the interface's connection status, power supply status, and fault alarm information in real time. Second, the planned charging information of the access targets is collected through the charging station's user reservation system. Finally, the real-time access status information of each charging interface is associated and bound with the corresponding access target plan information to generate a charging station information dataset uniquely identified by the charging interface number.

[0033] For example, the target microgrid is equipped with one 10-interface fast charging station to serve 20 logistics electric vehicles and 5 electric forklifts within the park. Real-time charging interface access status information is collected: interfaces 1-5 are in charging mode, interfaces 6-7 are in reserved lockout mode, interfaces 8-10 are idle, and there are no faulty interfaces. Access object planning information is collected: the reservation system shows that interface 1 is for logistics vehicle A, with a planned charging time of 10:00-11:00, a planned charging power of 60kW, and a planned charging amount of 60kWh; interface 6 is for electric forklift B, with a planned charging time of 11:30-12:00, a planned charging power of 20kW, and a planned charging amount of 10kWh.

[0034] In this embodiment of the invention, a standardized data acquisition and preprocessing process covers core data dimensions such as microgrid power supply capacity, grid interaction constraints, inherent power load, charging interface status, and charging plans, ensuring high accuracy and timeliness of data input to subsequent modules. Simultaneously, classifying, storing, and associating these two types of data forms a standardized dataset that can directly interface with feature input interfaces, reducing the complexity of subsequent data processing and effectively avoiding deviations in subsequent technical processes caused by missing or delayed data. This provides a data foundation for the stable operation of the entire microgrid self-regulating energy management system.

[0035] The expected load prediction module 12 is used to generate expected load prediction information for the target microgrid within a future preset management window based on the real-time operating status information and the charging station information, combined with a pre-built expected load prediction model.

[0036] In this embodiment of the invention, based on the real-time operating status information and the charging station information, and combined with a pre-constructed expected load prediction model, expected load prediction information for the target microgrid within a future preset management window is generated. Relying solely on real-time collected operating status and charging station information cannot predict future load change trends, making it difficult to support the power allocation decisions of the adjustment margin optimization module. If the load prediction results are biased, it will lead to a mismatch between the charging load management strategy and the actual power supply capacity of the microgrid, causing problems of power overcapacity or power undercapacity. Therefore, it is necessary to construct a hierarchical expected load prediction model, and through feature extraction, pattern mapping, probability correction, and other processes, combined with real-time additional features, achieve accurate load prediction, providing reliable data support for subsequent adjustment margin optimization.

[0037] In one embodiment, the expected load forecasting module 12 is further configured to:

[0038] Based on the real-time operating status information and the charging station information, extract the current time information and the number of available interfaces of the target microgrid;

[0039] The current time information and the number of available interfaces are input into the load feature pattern mapping layer of the expected load prediction model to obtain a load feature pattern list, wherein the load feature pattern list is associated with a pattern probability list.

[0040] Based on the load characteristic pattern list, the pattern probability list, and the charging station information, and combined with the artificial intelligence load prediction layer of the expected load prediction model, expected load prediction information of the target microgrid is generated, wherein the expected load prediction information is constrained by the preset management window as the time dimension.

[0041] First, based on the real-time operating status information and the charging station information, the current time information and the number of available interfaces in the target microgrid are extracted. The current time information refers to the real-time time data during load forecasting operations, including date type, time period, and other dimensions; it is a time characteristic parameter affecting microgrid load changes. The number of available interfaces refers to the number of interfaces in the charging station that are idle and ready to connect to new charging objects; it is a core parameter reflecting the potential growth space of charging load. The data set transmitted from the real-time information acquisition module 11 is labeled with the current timestamp, and time attributes are determined according to preset rules, such as distinguishing between weekday production periods and weekend rest periods. The real-time access status information of charging interfaces in the charging station information is retrieved, the number of interfaces in an idle state is counted, and interfaces in charging, reserved, or faulty states are removed. The current time information and the number of available interfaces are associated and stored as basic feature data for input load feature pattern mapping layer.

[0042] For example, the current time is 10:00 AM on a weekday, which is a non-peak production period. Of the 10 interfaces at the charging station, interfaces 1-5 are charging, interfaces 6-7 are reserved, and interfaces 8-10 are idle. Extract the current time information: weekday, 10:00 AM, non-peak production period; count the number of idle interfaces: filter out the idle interfaces 8, 9, and 10, and the number of idle interfaces is 3.

[0043] Next, the current time information and the number of available interfaces are input into the load feature pattern mapping layer of the expected load prediction model to obtain a load feature pattern list, wherein the load feature pattern list is associated with a pattern probability list.

[0044] The steps for constructing the expected load forecasting model include:

[0045] Acquire sample data of the target microgrid, and reorganize the sample data to obtain training sample data, including first training sample data and second training sample data;

[0046] Based on the first training sample data, a load feature pattern mapping layer is constructed and trained, wherein the load feature pattern mapping layer is constructed based on either a rule model or a data regression model.

[0047] Based on the second training sample data, construct and train an artificial intelligence load prediction layer based on a machine learning regression model;

[0048] The load feature pattern mapping layer and the artificial intelligence load prediction layer are connected in series to obtain the expected load prediction model, and the expected load prediction model is retrained as a whole.

[0049] First, sample data of the target microgrid is acquired, and the sample data is reorganized to obtain training sample data, including first training sample data and second training sample data.

[0050] This involves acquiring sample data of the target microgrid, reorganizing the sample data to obtain training sample data, including first training sample data and second training sample data, including:

[0051] Extract sample time information and the number of available sample interfaces from the sample data;

[0052] Based on a predefined statistical analysis period, using sample time information as the first index and the number of available sample interfaces as the second index, the historical load data in the sample data is aggregated and statistically analyzed to determine the typical load characteristic patterns and probability distributions corresponding to the sample time information and the number of available sample interfaces, thus forming the first training sample data.

[0053] Additional features are extracted from the sample data, and the typical load feature patterns and probability distributions are associated and combined with the additional features to form the second training sample data.

[0054] First, sample time information and the number of available interfaces are extracted from the sample data. Sample data refers to a comprehensive dataset accumulated during the historical operation of the target microgrid, including real-time operating status information, charging station information, and historical load data. Sample time information refers to the timestamp data in the historical operating data, containing historical date type and time period attributes. The number of available interfaces refers to the number of idle interfaces at charging stations corresponding to each time point in the historical operating data. The full historical operating sample data of the target microgrid is retrieved, encompassing historical real-time operating status information, historical charging station information, and corresponding historical load data. Historical charging interface status information with timestamps is extracted from the sample data, and the number of idle interfaces at each time point is counted to form a sequence of available interface numbers. The date type and time period attributes are parsed from the timestamps of the sample data to generate a sample time information sequence. The sample time information sequence and the sample available interface number sequence are then correlated one-to-one according to the timestamps to form a basic feature sequence.

[0055] For example, in a model application experiment, the historical operation data of a photovoltaic-storage-charging integrated microgrid in an industrial park throughout 2024 was used. This microgrid includes a 3MW photovoltaic array, a 2MWh energy storage system, and a charging station with 10 charging interfaces. The data sampling frequency was once per hour, totaling 8760 valid sample data, covering the complete operation scenario for 365 days of the year. The sample data includes the following key feature dimensions:

[0056] Time characteristics: timestamp, hour, day of the week, whether it is a weekday, time period classification (off-peak / off-peak / peak);

[0057] Charging station characteristics: number of available charging ports (0-10), port occupancy status;

[0058] Power generation characteristics: Photovoltaic output (0-3MW), energy storage output (-2~2MW);

[0059] Load characteristics: inherent load (1-3MW), charging load (0-0.6MW), total load;

[0060] Environmental characteristics: Temperature (5-35°C);

[0061] Market characteristics: Time-of-use electricity pricing (0.3 yuan during off-peak hours, 0.6 yuan during normal hours, and 1.0 yuan during peak hours).

[0062] Some examples of the obtained valid sample data are shown in Table 1 below:

[0063]

[0064] Table 1. Exemplary valid sample data

[0065] Correspondingly, the statistical characteristics of the 8760 valid sample data in the above example are shown in Table 2 below:

[0066]

[0067] Table 2. Exemplary statistical characteristics of the data - based on 8760 valid sample data.

[0068] For example, retrieve the full historical operation sample data of the target microgrid for the past year, extract the timestamps from 00:00 to 24:00 each day from the sample data, parse out the weekday and weekend date types, and divide the peak electricity consumption periods into 7:00-11:00 AM and 5:00-9:00 PM, and the rest into off-peak or low-peak periods, to form the sample time information; extract the charging interface status corresponding to each timestamp from the historical charging station information, and count that the number of idle interfaces at 10:00 AM on weekdays is mostly 2-4, and the number of idle interfaces at 10:00 AM on weekends is mostly 5-8, to form the sample number of available interfaces; associate the above two types of information by timestamp to form a basic feature combination, for example: weekday - 10:00 AM - off-peak period - 3 available interfaces.

[0069] Secondly, based on a predefined statistical analysis period, using sample time information as the first index and the number of available interfaces as the second index, aggregated statistical analysis is performed on the historical load data in the sample data to determine the typical load characteristic patterns and probability distributions corresponding to the sample time information and the number of available interfaces, thus forming the first training sample data. The first training sample data refers to a dataset that uses sample time information and the number of available interfaces as dual indices to associate typical load characteristic patterns and probability distributions, used to train the load characteristic pattern mapping layer. The predefined statistical analysis period refers to the historical data statistical period set for mining load characteristic patterns; the period length must match the time accuracy requirements of subsequent load prediction. Dual-index aggregated statistics refers to an analysis method that uses sample time information as the first index and the number of available interfaces as the second index to group and statistically analyze historical load data. Typical load characteristic patterns and probability distributions refer to the common changing patterns of historical load data under the same dual-index combination and the frequency proportion of each pattern appearing under that combination.

[0070] Specifically, based on the operating characteristics of the target microgrid, a predefined statistical analysis period is set; the associated basic characteristic sequence is matched with the corresponding historical load data; the matched data is grouped according to the dual indexes of sample time information and the number of available interfaces, with historical load data having the same time attributes and number of available interfaces in the same group; the historical load data in each group is aggregated and statistically analyzed to analyze the load change trend, fluctuation range and other characteristics, and the typical load characteristic patterns corresponding to the group are extracted; the frequency of occurrence of each type of typical load characteristic pattern in each group is counted, and the probability distribution of each pattern is calculated; the dual index information, typical load characteristic patterns and corresponding probability distributions of each group are integrated to form the first training sample data.

[0071] For example, the statistical analysis period is set to 1 hour. After matching the basic feature sequence with historical load data, dual-index aggregation statistics are performed. Using "Weekday - 10:00 AM - Off-peak hours" as the first index and "3 available interfaces" as the second index, the historical load data under this group is statistically analyzed. It is found that in 70% of the cases, the load shows a trend of stable inherent load and slow growth of new charging load; 20% shows a trend of slight fluctuation in inherent load; and 10% shows a trend of decreasing inherent load. This dual-index combination, three typical load feature patterns and their corresponding probability distributions are bound together to form a first training sample data. The above statistics and integration are completed for all dual-index combinations, and finally, a first training sample dataset covering all combinations is generated.

[0072] Next, additional features are extracted from the sample data, and the typical load feature patterns and probability distributions are associated and combined with the additional features to form the second training sample data. The second training sample data refers to the dataset that associates and combines typical load feature patterns and probability distributions with additional features, used to train the AI ​​load prediction layer. Additional features refer to auxiliary prediction parameters such as environmental features, market features, and microgrid preceding features in historical operating data. Feature association and combination refers to the process of binding typical load feature patterns and probability distributions under the same dual-index combination with additional features to form a multi-dimensional feature combination. Historical environmental features, historical market features, and historical microgrid preceding features are extracted from the sample data to form a set of additional features. The set of additional features is matched one-to-one with the dual-index combinations in the first training sample data according to timestamps, ensuring that the typical load feature patterns and probability distributions corresponding to each dual-index combination can be associated with additional features at the same time. The matched dual-index information, typical load feature patterns, probability distributions, and additional features are integrated and encapsulated to form the second training sample data.

[0073] For example, additional features are extracted from the historical sample data of the target microgrid and associated and combined. Historical environmental features such as light intensity and temperature at 10:00 AM on weekdays, historical market features such as the time-of-use electricity price flat period attribute corresponding to that time period, and historical microgrid preceding features such as photovoltaic output and inherent load power at 9:00 AM are extracted from the sample data to constitute additional features. The additional features are then bound to the typical load characteristic patterns and probability distributions corresponding to "weekday - 10:00 AM - off-peak period - 3 spare interfaces". All entries in the first training sample data are associated with the above additional features to finally generate the second training sample dataset.

[0074] Secondly, based on the first training sample data, a load feature pattern mapping layer is constructed and trained. This layer is built upon either a rule-based model or a data regression model. The load feature pattern mapping layer is the underlying module of the expected load prediction model, used to establish the correspondence between the sample time information-sample available interface quantity dual index and the typical load feature pattern-probability distribution. The rule-based model is a model based on expert experience and pre-defined mapping rules, directly defining the load feature patterns and probabilities corresponding to different dual index combinations. The data regression model is a model based on statistical learning algorithms, which learns the correlation patterns of sample data and autonomously generates the mapping relationship between the dual indexes and load feature patterns.

[0075] First, determine the model architecture, choosing either a rule-based model or a data regression model as the foundation for the load feature pattern mapping layer. Define the model's input and output: the input is the dual-indexed features from the first training sample data, namely sample time information and the number of available interfaces; the output is a list of typical load feature patterns and a list of pattern probabilities that match the input dual-indexed features. Second, conduct model training, dividing the first training sample data into a training set and a validation set according to a preset ratio. Use the training set data to optimize the model parameters, allowing the model to learn the mapping rules from dual-indexed features to load patterns and probabilities. Determine training completion: when the matching accuracy of the model's output load feature patterns and probabilities with the validation set sample annotations is ≥90% of a preset threshold, the mapping layer training is considered complete.

[0076] For example, a data regression model is selected as the mapping layer architecture. The model input is the sample time information and the number of sample free interfaces. The regression model parameters are optimized with training set data to learn the load pattern corresponding to the dual index. During the training process, the model outputs a list of pattern A (stable inherent load + slow growth of new charging load), pattern B (small fluctuation of inherent load), and pattern C (decreasing inherent load) and a list of probabilities of 60%, 30%, and 10%. When the validation set test shows that the matching accuracy between the model output pattern and probability and the sample label reaches 92%, that is, when the preset threshold of 90% is met, the training stops, and the training of the load feature pattern mapping layer is completed.

[0077] Further, based on the second training sample data, an AI load prediction layer based on a machine learning regression model is constructed and trained. The AI ​​load prediction layer is the upper-level structure of the desired load prediction model, built upon the machine learning regression model, and is used to achieve probability correction of load feature patterns and final load prediction. The machine learning regression model refers to an algorithmic model with non-linear fitting capabilities, such as gradient boosting trees and neural networks, which can learn the correlation between multi-dimensional features and probability correction rules. A machine learning regression model is selected as the basic architecture of the AI ​​load prediction layer; the model input consists of multi-dimensional features from the second training sample data, namely, a list of load feature patterns, a list of pattern probabilities, and additional sample features; the model output is an adaptive corrected probability list adjusted from the original input probability list. Model training is conducted by dividing the second training sample data into a training set and a validation set. The training set data is used to optimize task hyperparameters such as tree depth and learning rate, allowing the model to learn the correction rules of additional features on pattern probabilities. The training completion criterion is determined when the load prediction error corresponding to the corrected probability output by the model based on the validation set data is ≤ a preset threshold, such as a mean absolute error (MAE) ≤ 5%, at which point the prediction layer training is considered complete.

[0078] For example, a gradient boosting tree regression model is selected as the prediction layer architecture. The model input is a list of load feature patterns, a list of pattern probabilities, and additional features of the samples. During training, the hyperparameters of the gradient boosting tree are adjusted using training set data to learn the rules for adjusting the load pattern probabilities based on the additional features. During training, the model outputs an adaptive correction probability list. When the validation set test shows that the mean absolute error (MAE) of load prediction calculated based on the correction probability is 4.2%, which is lower than the preset threshold of 5%, training is stopped, and the training of the artificial intelligence load prediction layer is completed.

[0079] Subsequently, the load feature pattern mapping layer and the AI ​​load prediction layer are concatenated to obtain the expected load prediction model, and the expected load prediction model is then retrained as a whole. Model concatenation refers to connecting the trained load feature pattern mapping layer and the AI ​​load prediction layer in the order of mapping layer output and prediction layer input, forming an end-to-end model structure. Overall retraining refers to jointly optimizing the concatenated model using the full set of sample data to eliminate the error accumulation caused by hierarchical training and improve the overall performance of the model.

[0080] First, the model is concatenated by connecting the output port of the load feature pattern mapping layer to the input port of the AI ​​load prediction layer. The load feature pattern list – pattern probability list output by the mapping layer is defined as the basic input features for the prediction layer, while reserving interfaces for additional features. The model input consists of basic features + additional features, namely time information, the number of available interfaces, and real-time additional features; the model output is the expected load prediction information for the target microgrid within a preset management window. Second, overall retraining is performed by inputting all sample data into the concatenated model and simultaneously optimizing the parameters of both layers using the backpropagation algorithm to reduce the impact of inter-layer error propagation. The training is considered complete when the overall load prediction accuracy of the concatenated model improves by ≥3% compared to the accuracy during layered training, and the prediction error remains consistently below a preset threshold.

[0081] For example, the output of the mapping layer is first directly connected to the basic feature input port of the prediction layer, and the additional feature input port of the prediction layer is connected to features such as light intensity and electricity price to complete the model chaining. The overall model input is, for example, a weekday at 10 am, 3 spare interfaces, sufficient light intensity, and flat electricity price. During training, the full sample data of the industrial park microgrid over the past year is input. The parameters of the mapping layer and the prediction layer are simultaneously optimized through the backpropagation algorithm to eliminate the accumulation of errors in the layered training. During the training process, the model outputs the expected load prediction information of the microgrid load slowly increasing from 1.8MW to 2.2MW in the next hour. When the average absolute error of the overall load prediction of the model is monitored to decrease from 4.2% during layered training to 3.1%, the accuracy is improved by 26%, which meets the preset improvement ratio requirement of 3%, and the error is stable, the training is stopped, and the expected load prediction model is completed.

[0082] For example, continuing the previous example of a photovoltaic-storage-charging integrated microgrid in an industrial park, a feature mapping layer is constructed using a data regression model. This establishes a mapping relationship between time-interface quantity dual indices and load characteristic patterns, with the convergence condition being a verification accuracy of over 90%. An artificial intelligence load prediction layer is constructed using a gradient boosting regressor machine learning regression model.

[0083] Specifically, the training parameter configuration used is shown in Table 3 below:

[0084]

[0085] Table 3. Exemplary training parameters for an integrated photovoltaic, energy storage, and charging microgrid in an industrial park.

[0086] For example, a training process record of the AI ​​load prediction layer is shown in Table 4 below:

[0087]

[0088] Table 4. Exemplary training records for the AI ​​load prediction layer

[0089] Simultaneously, a corresponding training process analysis is performed, including:

[0090] a. The training MAE gradually decreased from an initial 8.14% to 2.58%, a decrease of 68.3%;

[0091] b. Verification showed that MAE decreased from the initial 9.27% ​​to 2.96%, a decrease of 68.1%;

[0092] c. The training R² steadily increased from 0.095 to 0.933, indicating a significant improvement in model fit;

[0093] d. The validation R² improved from 0.087 to 0.926, and the performance of the training and validation sets was similar, with no obvious overfitting;

[0094] e. The performance stabilizes after the 80th round, indicating that the model has converged sufficiently.

[0095] Furthermore, the trained load feature pattern mapping layer is concatenated with the AI ​​load prediction layer to form an end-to-end expected load prediction model, which is then retrained and optimized as a whole. For example, the optimization method uses the backpropagation algorithm to simultaneously optimize the parameters of both layers, with 30 iterations, thereby eliminating the error accumulation caused by layered training. The optimization effect of the concatenated model is as follows:

[0096] MAE: 2.96% → 2.52% (a decrease of 14.9%);

[0097] R²: 0.926 → 0.942 (an improvement of 1.7%).

[0098] For example, the performance of the overall retrained and optimized expected load prediction model is evaluated, and the results are shown in Table 5 below:

[0099]

[0100] Table 5 Performance evaluation results of an exemplary expected load forecasting model

[0101] Based on this, the current time information and the number of available interfaces are input into the load feature pattern mapping layer of the expected load prediction model to obtain a load feature pattern list, wherein the load feature pattern list is associated with a pattern probability list. The current time information and the number of available interfaces are used as input parameters and input into the pre-trained load feature pattern mapping layer; the mapping layer calls the internally stored "time-interface quantity-load pattern" mapping relationship library to match the typical load feature pattern that best matches the current feature; the matched load feature pattern list is output, and the occurrence probability of each pattern is simultaneously output to form a pattern probability list.

[0102] For example, the input "Workday - 10 AM - Off-peak production period, number of available interfaces 3" is input into the load characteristic pattern mapping layer. The mapping layer matches three typical load characteristic patterns: Pattern A (stable inherent load + slow growth of new charging load), Pattern B (small fluctuations in inherent load + rapid growth of new charging load), and Pattern C (decreasing inherent load + stable new charging load); the corresponding pattern probability list is output: Pattern A probability 60%, Pattern B probability 30%, and Pattern C probability 10%.

[0103] Furthermore, based on the load characteristic pattern list, the pattern probability list, and the charging station information, and combined with the artificial intelligence load prediction layer of the expected load prediction model, expected load prediction information of the target microgrid is generated, wherein the expected load prediction information is constrained by the preset management window as the time dimension.

[0104] Specifically, based on the load characteristic pattern list, the pattern probability list, and the charging station information, and combined with the artificial intelligence load prediction layer of the expected load prediction model, expected load prediction information for the target microgrid is generated, including:

[0105] Based on the charging station information, real-time additional features are extracted, wherein the real-time additional features include at least environmental feature dimension, market feature dimension and microgrid preceding feature dimension;

[0106] The load feature pattern list, the pattern probability list, and the real-time additional features are fused based on feature concatenation to obtain real-time fused features;

[0107] The real-time fusion features are input to the artificial intelligence load prediction layer to adaptively correct the pattern probability list and obtain the adaptively corrected probability list.

[0108] Based on the adaptive correction probability list and the load characteristic pattern list, the expected load prediction information of the target microgrid is calculated and obtained.

[0109] First, based on the charging station information, real-time additional features are extracted. These real-time additional features include at least environmental feature dimensions, market feature dimensions, and microgrid antecedent feature dimensions. Real-time additional features refer to auxiliary predictive features extracted from the charging station information, excluding time and interface quantity, and include environmental feature dimensions, market feature dimensions, and microgrid antecedent feature dimensions. Environmental feature dimensions refer to environmental parameters affecting distributed generation output and electricity load, such as light intensity, temperature, and wind speed. Market feature dimensions refer to market parameters affecting microgrid electricity purchase and sale decisions, such as time-of-use pricing and grid peak-valley time division. Microgrid antecedent feature dimensions refer to the microgrid's operating parameters from the previous data collection period, such as distributed generation output and inherent load power at the previous moment.

[0110] For example, parameters for environmental characteristics, market characteristics, and microgrid antecedent characteristics are extracted from charging station information and real-time operating status information, respectively. The extracted real-time additional features are standardized to eliminate dimensional differences between different features. The processed real-time additional features are then integrated and stored as input data for feature fusion. For example, environmental characteristics include: sufficient current sunlight intensity and a temperature of 25°C; market characteristics include: currently in a period of flat time-of-use pricing; and microgrid antecedent characteristics include: stable photovoltaic output and no significant fluctuations in inherent load during the previous data collection cycle. Standardization of these features completes the extraction of real-time additional features.

[0111] Secondly, the load feature pattern list, the pattern probability list, and the real-time additional features are fused based on feature concatenation to obtain real-time fused features. Feature concatenation fusion refers to the process of concatenating the load feature pattern list, the pattern probability list, and the real-time additional features according to a preset dimension to form a fused feature vector containing multi-dimensional information. Real-time fused features refer to the high-dimensional feature vector formed by concatenating the load feature pattern list, the pattern probability list, and the real-time additional features, containing three types of information: pattern, probability, and auxiliary features. The load feature pattern list, the pattern probability list, and the real-time additional features are retrieved, and the three types of data are concatenated according to the dimensional order of load feature pattern - pattern probability - real-time additional features to form a real-time fused feature vector; the format of the concatenated feature vector is converted to meet the input requirements of the artificial intelligence load prediction layer.

[0112] Furthermore, the real-time fused features are input to the AI ​​load prediction layer to adaptively correct the pattern probability list, obtaining an adaptively corrected probability list. The adaptively corrected probability list refers to the new probability list obtained by the AI ​​load prediction layer after adjusting the original pattern probability list based on real-time additional features, which can improve the accuracy of the prediction results. Adaptive correction refers to the process by which the AI ​​load prediction layer autonomously adjusts the probability values ​​of each load feature pattern based on changes in real-time additional features, making the probability distribution more consistent with the current operating scenario. The real-time fused features are input to the pre-trained AI load prediction layer. Based on the learned feature association patterns, the prediction layer analyzes the influence of real-time additional features on the probability of each load pattern, adaptively adjusts the original pattern probability list according to the degree of influence, and outputs the corrected adaptively corrected probability list.

[0113] For example, real-time fusion features are input into the AI ​​load prediction layer; the model analyzes the characteristics of stable current photovoltaic output, no significant fluctuations in charging demand, and stable preceding output, and determines that the probability of the occurrence of mode A should be increased; an adaptive correction probability list is output: mode A probability 75%, mode B probability 20%, mode C probability 5%.

[0114] Finally, based on the adaptive correction probability list and the load characteristic pattern list, the expected load forecast information of the target microgrid is calculated and obtained. The expected load forecast information refers to the set of load forecast values ​​for the microgrid within a future time period, calculated using a preset management window as the time dimension constraint and combining the adaptive correction probability list and the load characteristic pattern list. The adaptive correction probability list and the corresponding load characteristic pattern list are retrieved to clarify the load change patterns and forecast values ​​under each pattern; a weighted summation algorithm is used, with the correction probability of each pattern as the weight, to calculate the load forecast values ​​at different time nodes; using the preset management window as the time dimension, the forecast values ​​at each time node are integrated to generate complete expected load forecast information.

[0115] For example, the adaptive correction probability list and the corresponding load forecast values ​​for each mode are retrieved: Mode A corresponds to a slow increase in load from the current 1.8MW to 2.2MW within the future management window; Mode B corresponds to a rapid increase in load to 2.5MW; and Mode C corresponds to a decrease in load to 1.6MW. Using the correction probability as the weight, the load values ​​for each mode at different time points within the management window are weighted and calculated. With 15-minute intervals, the specific load forecast values ​​for each node are obtained. For example, at the first 15-minute node, the load for Mode A is 1.85MW, the load for Mode B is 1.95MW, and the load for Mode C is... The first node has a load of 1.75MW. The weighted calculation yields a predicted value of 1.85×0.75+1.95×0.2+1.75×0.05=1.865MW. The second 15-minute node has loads of 1.95MW for Mode A, 2.15MW for Mode B, and 1.70MW for Mode C. The weighted calculation yields a predicted value of 1.95×0.75+2.15×0.2+1.70×0.05=1.9775MW. Subsequent nodes are calculated similarly to obtain their corresponding values. Finally, the predicted load values ​​for all nodes in the entire window are output, generating complete expected load prediction information.

[0116] In this embodiment of the invention, a hierarchical expected load prediction model is constructed, achieving a closed-loop process from basic feature extraction to accurate load prediction, effectively improving the accuracy and reliability of load prediction. Through sample data recombination and dual-layer model cascade training, load characteristic patterns under different time and interface combinations can be accurately matched; combined with adaptive probability correction of real-time additional features, the prediction results are more closely aligned with the actual operating scenarios of the microgrid. The final output expected load prediction information provides accurate load data support for the regulation margin optimization module, ensuring the scientific and rational nature of subsequent charging load management strategies, and improving the operational stability and economy of the entire microgrid self-regulating energy management system.

[0117] The regulation margin optimization module 13 is used to establish and solve an optimization task with the goal of maximizing the regulation margin of the target microgrid in a future preset period, and output the power limit value corresponding to each charging interface.

[0118] In this embodiment of the invention, an optimization task is established and solved to maximize the regulation capacity margin of the target microgrid within a preset future time period, outputting the power limit value corresponding to each charging interface. In the self-regulating energy management system of the target microgrid, if the charging load allocation based on expected load forecast information lacks scientific optimization constraints, it can easily lead to an excessively low regulation capacity margin, making it unable to cope with load fluctuations or sudden changes in distributed power output. This may also cause problems such as excessively high economic costs and equipment overload operation. Therefore, it is necessary to construct an optimization task with the goal of maximizing the regulation capacity margin, combining cost and safety constraints, and generating accurate charging interface power limit values ​​through rolling time-domain solution to ensure the stability and economy of microgrid operation.

[0119] In one embodiment, the adjustment margin optimization module 13 is further configured to:

[0120] The optimization task is established with the objective function of maximizing the regulation capacity margin of the target microgrid within the preset future time period. The regulation capacity margin is obtained by correcting the difference between the available power and the expected load forecast information by a cost factor.

[0121] The expected load prediction information and the real-time operating status information are used as input parameters for the optimization task;

[0122] Under the premise of satisfying the preset constraints, the optimization task is solved in the rolling time domain, and the power limit value sequence of each charging interface in the future preset time period is dynamically output.

[0123] The constraints of the optimization task include:

[0124] Cost constraints, defined based on electricity purchase costs and equipment operating loss costs, are used to limit the total economic cost corresponding to the power limit value;

[0125] Safety operation constraints are defined based on the power balance limit of the target microgrid, the power capacity limit of the equipment, and the minimum charging requirements of users.

[0126] First, the optimization task is established with the objective function of maximizing the regulation capacity margin of the target microgrid within the predetermined future time period. The regulation capacity margin is obtained by adjusting the difference between available power and expected load forecast information using a cost factor. The regulation capacity margin refers to the parameter obtained by adjusting the difference between the microgrid's available power and expected load forecast information using a cost factor, reflecting the microgrid's redundancy capacity to cope with load or output fluctuations. The objective function is a mathematical expression that optimizes by maximizing the regulation capacity margin, serving as the core guiding principle of the optimization task. Cost constraints are constraints set based on electricity purchase costs and equipment operating losses, used to limit the upper limit of the economic cost of the optimization scheme. Safe operation constraints are constraints set based on the microgrid's power balance limit, equipment power capacity limit, and the minimum charging demand of users, representing the bottom-line requirements for ensuring the safe and stable operation of the microgrid. In other words, the optimization task is a set of mathematical models composed of the objective function and constraints.

[0127] Specifically, with the goal of maximizing the adjustability margin within the future preset management window, the objective function formula is defined as follows: Where M is the total adjustment capacity margin and k is the cost factor. The maximum power that the microgrid can provide in the t-th sub-time period is Let be the actual total load power in the t-th sub-period; where ,in For the real-time output of the distributed power source in the t-th sub-time period, This is the upper limit of the tie line power. , Let be the inherent load power of the t-th sub-time period. Let C be the total charging power in the t-th sub-period. Cost constraint: Total operating cost C = electricity purchase cost C1 + equipment operating loss cost C2, which must satisfy total operating cost C ≤ preset cost threshold C0. Wherein, electricity purchase cost C1 is calculated based on the forward power purchased from the tie line and the time-of-use electricity price, and equipment operating loss cost C2 is calculated based on the charging equipment's operating power and loss coefficient. Safe operation constraint: Power balance constraint: ,in The power of the interconnection line in the t-th sub-time period; equipment capacity constraints: charging interface output power ≤ interface rated power, distributed power output ≤ equipment rated output, absolute value of interconnection line power | |≤Connecting Line Limit Minimum charging requirement constraint for users: Output power of a single charging interface ≥ preset minimum charging power threshold.

[0128] For example, the next hour is set as the preset management window, divided into four 15-minute sub-periods. The cost factor k is 0.8, the preset cost threshold is 1.1 times the historical average cost of the period, and the minimum charging power threshold for a single charging interface is 10kW. The objective function is to maximize the total adjustment capacity margin of the four sub-periods. The cost constraint is defined as the sum of the electricity purchase cost and the charging equipment loss cost of the period not exceeding the preset threshold. The safe operation constraints include the photovoltaic array output not exceeding the rated value of 3MW, the interconnection power of the tie line not exceeding ±1.2MW, and the charging power of a single interface not less than 10kW. The optimization task is thus completed.

[0129] Secondly, the expected load prediction information and the real-time operating status information are used as input parameters for the optimization task. Input parameters refer to the basic data driving the optimization task solution, including expected load prediction information from the expected load prediction module 12 and real-time operating status information from the real-time information acquisition module 11. Model output parameters refer to the result data generated after solving the optimization task, i.e., the power limit value sequence for each charging interface in each sub-period within the future preset management window. Imported input parameters include the total expected load prediction value output by the expected load prediction module 12, the real-time output of distributed power sources, tie-line limits, inherent load power, and the number of charging interfaces from the real-time operating status information. The model output parameter format is defined, using the charging interface number-sub-period number-power limit value as an index to generate a structured power limit value sequence.

[0130] For example, the total expected load forecast values ​​of 1.86MW, 1.97MW, 2.05MW, and 2.12MW for the four sub-periods output by the expected load forecast module 12, as well as the real-time collected data such as photovoltaic output of 1.8MW, energy storage discharge of 0.5MW, tie line limit of ±1.2MW, and inherent load of 1.8MW, are imported into the optimization task. The output parameters are defined as the power limit values ​​of 10 charging interfaces in four 15-minute sub-periods, and the structured data generated is, for example, Interface 1 - First Period - 50kW.

[0131] Finally, under the premise of satisfying the preset constraints, the optimization task is solved in a rolling time domain, and the power limit value sequence of each charging interface within the preset future time period is dynamically output. The sub-time period solution window refers to the time unit of each calculation in the rolling time domain solution, with a length consistent with the granularity of the preset management window. The power limit value sequence refers to the set of power limit values ​​for each charging interface in all sub-time periods within the preset future management window, reflecting the dynamic change pattern of power limits. The preset future management window is divided into multiple sub-time periods, with the first sub-time period set as the initial solution window; the input parameters of the current solution window are substituted into the optimization task, and the optimal solution of the objective function is solved under the premise of satisfying cost constraints and safe operation constraints, obtaining the power limit value of each charging interface in that sub-time period; the solution window is rolled to the next sub-time period, the real-time operating status information is updated, and the solution process is repeated; the solution results of all sub-time periods are integrated to generate a complete power limit value sequence for the charging interfaces.

[0132] For example, the 1-hour preset management window is divided into 4 sub-periods of 15 minutes. First, the first 15 minutes is used as the solution window. The input parameters are substituted to obtain the power limit values ​​of the 10 interfaces in this period. For example, the power limit of charging interfaces 1-5 is 50kW, the power limit of reservation interfaces 6-7 is 40kW, and the power limit of idle interfaces 8-10 is 60kW. Then, the solution window is scrolled to the second 15 minutes, the real-time collected photovoltaic output is updated to the new data of 1.7MW, and the power limit values ​​of each interface in this period are re-solved. For example, the power limit value of interface 1-5 is adjusted to 48kW. The solution of the 4 sub-periods is completed in sequence, and the power limit value sequence of the 10 interfaces in the 4 periods is obtained and output to the charging load management module 14.

[0133] In this embodiment of the invention, by constructing an optimization task aimed at maximizing the regulation capacity margin, and combining cost and safety constraints, a refined optimization allocation of charging load is achieved. The rolling time-domain solution method can adapt to the dynamic changes in the real-time operating status of the microgrid. The output power limit value sequence ensures that the microgrid has sufficient regulation redundancy to cope with load or output fluctuations, while avoiding excessive electricity purchase costs and equipment overload operation, effectively improving the stability, economy, and reliability of microgrid operation.

[0134] The charging load management module 14 is used to allocate and manage the charging load of the target microgrid according to the power limit value.

[0135] In this embodiment of the invention, the charging load of the target microgrid is allocated and managed according to the power limit value. The structured sequence of charging interface number-sub-period-power limit value output by the adjustment margin optimization module 13 is received, and the latest charging station information from the real-time information acquisition module 11 is retrieved simultaneously. A matching rule between interface status and power limit value is established. For interfaces in charging, the current output power is adjusted according to the corresponding sub-period power limit value. For reserved interfaces, the power limit value is pushed to the reservation system in advance as the initial power benchmark after the reserved object connects. For idle interfaces, the power limit value is set to the maximum allowable power threshold when the interface connects. The actual charging power of each interface and the charging progress of the connected object are monitored in real time. The power limit value is compared to determine whether there are any abnormal situations such as exceeding the limit or failing to meet the minimum requirement. If an abnormality occurs, a dynamic adjustment mechanism is triggered to fine-tune the interface power without exceeding the total power limit. Data such as charging duration and actual power consumption of each interface are recorded to form a load management log, providing data support for subsequent parameter optimization.

[0136] For example, after receiving the power limit value sequence of 10 charging interfaces in 4 15-minute sub-time periods, the specific execution is as follows: For the logistics vehicle interfaces 1-5 that are currently charging, the power limit value of 50kW for the first time period and 48kW for the second time period are sequentially sent to the interface controller to adjust the output power in real time; For the electric forklift interfaces 8-10 that are reserved, the power limit value of 40kW for the corresponding time period is entered into the reservation system in advance to ensure that the forklift starts charging directly at this power when it connects at 11:30; For the idle interfaces 8-10, the power limit value of 60kW is set as the access threshold. When a temporary logistics vehicle connects, it will automatically supply power at a power not exceeding 60kW. At the same time, if the actual power of a certain interface exceeds the limit, it will be finely adjusted to within the limit value immediately.

[0137] In this embodiment of the invention, the charging load of the target microgrid is allocated and managed according to the power limit value, realizing refined and dynamic control of the charging load. This ensures that all charging behaviors strictly match the optimized power limit requirements, guaranteeing the stability of the microgrid's regulation margin and avoiding power supply fluctuations caused by disorderly growth of the charging load, while also meeting the charging needs of different access objects. At the same time, the complete management and control log provides real data support for subsequent load prediction model optimization and regulation margin parameter calibration, promoting the closed-loop optimization of the entire microgrid self-regulating energy management system and further improving the stability and economy of operation.

[0138] Through the specific implementation methods described above, the embodiments of the present invention achieve the following technical effects:

[0139] This invention provides a microgrid self-regulating energy management system, method, and intelligent terminal that integrates AI. By collecting and standardizing real-time data on microgrid operation status and charging station data across all dimensions, it lays a high-quality data foundation for the entire energy management process. Relying on a hierarchical AI model, it deeply mines load change patterns and accurately predicts future load trends, solving the problem of the disconnect between traditional forecasting and actual operating scenarios. With the goal of maximizing regulation capacity margin, it constructs optimization tasks based on both cost and safety constraints, generating a scientific charging power limiting scheme through rolling time-domain solutions, effectively avoiding the risks of over- or under-supply of power. Based on the power limiting scheme and the real-time status of the charging interface, it achieves refined load allocation and dynamic control, ensuring precise matching between charging behavior and the microgrid's operating status. Through multi-stage collaborative linkage, it enhances the microgrid's adaptive capability to cope with distributed power output fluctuations and charging load changes, achieving a dual improvement in energy utilization efficiency and operational economy. Simultaneously, through full-process data recording and feedback, it drives continuous strategy iteration, providing reliable technical support for the safe, stable, and efficient operation of the microgrid.

[0140] Example 2, as Figure 2 As shown, this invention provides a microgrid self-regulating energy management method incorporating AI, including:

[0141] S100: Obtain the real-time operating status information of the target microgrid and the corresponding charging station information.

[0142] Step S100 in the method provided in this embodiment of the invention includes:

[0143] The real-time operating status information includes at least real-time output data, tie-line limit data, and inherent load data; the charging station information includes at least real-time access status information and access target planning information for each charging interface.

[0144] S200: Based on the real-time operating status information and the charging station information, and combined with the pre-built expected load prediction model, generate expected load prediction information for the target microgrid within a future preset management window.

[0145] Step S200 in the method provided in this embodiment of the invention includes:

[0146] Based on the real-time operating status information and the charging station information, extract the current time information and the number of available interfaces of the target microgrid;

[0147] The current time information and the number of available interfaces are input into the load feature pattern mapping layer of the expected load prediction model to obtain a load feature pattern list, wherein the load feature pattern list is associated with a pattern probability list.

[0148] Based on the load characteristic pattern list, the pattern probability list, and the charging station information, and combined with the artificial intelligence load prediction layer of the expected load prediction model, expected load prediction information of the target microgrid is generated, wherein the expected load prediction information is constrained by the preset management window as the time dimension.

[0149] Specifically, based on the load characteristic pattern list, the pattern probability list, and the charging station information, and combined with the artificial intelligence load prediction layer of the expected load prediction model, expected load prediction information for the target microgrid is generated, including:

[0150] Based on the charging station information, real-time additional features are extracted, wherein the real-time additional features include at least environmental feature dimension, market feature dimension and microgrid preceding feature dimension;

[0151] The load feature pattern list, the pattern probability list, and the real-time additional features are fused based on feature concatenation to obtain real-time fused features;

[0152] The real-time fusion features are input to the artificial intelligence load prediction layer to adaptively correct the pattern probability list and obtain the adaptively corrected probability list.

[0153] Based on the adaptive correction probability list and the load characteristic pattern list, the expected load prediction information of the target microgrid is calculated and obtained.

[0154] The steps for constructing the expected load forecasting model include:

[0155] Acquire sample data of the target microgrid, and reorganize the sample data to obtain training sample data, including first training sample data and second training sample data;

[0156] Based on the first training sample data, a load feature pattern mapping layer is constructed and trained, wherein the load feature pattern mapping layer is constructed based on either a rule model or a data regression model.

[0157] Based on the second training sample data, construct and train an artificial intelligence load prediction layer based on a machine learning regression model;

[0158] The load feature pattern mapping layer and the artificial intelligence load prediction layer are connected in series to obtain the expected load prediction model, and the expected load prediction model is retrained as a whole.

[0159] This involves acquiring sample data of the target microgrid, reorganizing the sample data to obtain training sample data, including first training sample data and second training sample data, including:

[0160] Extract sample time information and the number of available sample interfaces from the sample data;

[0161] Based on a predefined statistical analysis period, using sample time information as the first index and the number of available sample interfaces as the second index, the historical load data in the sample data is aggregated and statistically analyzed to determine the typical load characteristic patterns and probability distributions corresponding to the sample time information and the number of available sample interfaces, thus forming the first training sample data.

[0162] Additional features are extracted from the sample data, and the typical load feature patterns and probability distributions are associated and combined with the additional features to form the second training sample data.

[0163] S300: With the goal of maximizing the regulation margin of the target microgrid in a future preset period, establish an optimization task and solve it, and output the power limit value corresponding to each charging interface.

[0164] Step S300 in the method provided in this embodiment of the invention includes:

[0165] The optimization task is established with the objective function of maximizing the regulation capacity margin of the target microgrid within the preset future time period. The regulation capacity margin is obtained by correcting the difference between the available power and the expected load forecast information by a cost factor.

[0166] The expected load prediction information and the real-time operating status information are used as input parameters for the optimization task;

[0167] Under the premise of satisfying the preset constraints, the optimization task is solved in the rolling time domain, and the power limit value sequence of each charging interface in the future preset time period is dynamically output.

[0168] The constraints of the optimization task include:

[0169] Cost constraints, defined based on electricity purchase costs and equipment operating loss costs, are used to limit the total economic cost corresponding to the power limit value;

[0170] Safety operation constraints are defined based on the power balance limit of the target microgrid, the power capacity limit of the equipment, and the minimum charging requirements of users.

[0171] S400: Based on the power limit value, allocate and manage the charging load of the target microgrid.

[0172] Example 3: The present invention provides an AI-integrated microgrid self-regulating energy management smart terminal, and the AI-integrated microgrid self-regulating energy management system and method provided by the present invention are applied to the smart terminal.

[0173] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A microgrid self-regulating energy management system incorporating AI, characterized in that: include: The real-time information acquisition module is used to acquire the real-time operating status information of the target microgrid and the corresponding charging station information; The expected load forecasting module is used to generate expected load forecasting information for the target microgrid within a future preset management window based on the real-time operating status information and the charging station information, combined with a pre-built expected load forecasting model. The regulation margin optimization module is used to establish and solve an optimization task with the goal of maximizing the regulation margin of the target microgrid in a future preset time period, and output the power limit value corresponding to each charging interface. The charging load management module is used to allocate and manage the charging load of the target microgrid according to the power limit value; The expected load forecasting module is also used for: Based on the real-time operating status information and the charging station information, extract the current time information and the number of available interfaces of the target microgrid; The current time information and the number of available interfaces are input into the load feature pattern mapping layer of the expected load prediction model to obtain a load feature pattern list, wherein the load feature pattern list is associated with a pattern probability list. Based on the load characteristic pattern list, the pattern probability list, and the charging station information, and combined with the artificial intelligence load prediction layer of the expected load prediction model, expected load prediction information of the target microgrid is generated, wherein the expected load prediction information is constrained by the preset management window as the time dimension.

2. The microgrid self-regulating energy management system incorporating AI as described in claim 1, characterized in that, The real-time information acquisition module is also used for: the real-time operating status information including at least real-time output data, tie line limit data, and inherent load data; and the charging station information including at least real-time access status information and access object plan information for each charging interface.

3. The microgrid self-regulating energy management system incorporating AI as described in claim 1, characterized in that, The expected load forecasting module is also used for: Based on the charging station information, real-time additional features are extracted, wherein the real-time additional features include at least environmental feature dimension, market feature dimension and microgrid preceding feature dimension; The load feature pattern list, the pattern probability list, and the real-time additional features are fused based on feature concatenation to obtain real-time fused features; The real-time fusion features are input to the artificial intelligence load prediction layer to adaptively correct the pattern probability list and obtain the adaptively corrected probability list. Based on the adaptive correction probability list and the load characteristic pattern list, the expected load prediction information of the target microgrid is calculated and obtained.

4. The microgrid self-regulating energy management system incorporating AI as described in claim 1, characterized in that, The expected load forecasting module is also used for: Acquire sample data of the target microgrid, and reorganize the sample data to obtain training sample data, including first training sample data and second training sample data; Based on the first training sample data, a load feature pattern mapping layer is constructed and trained, wherein the load feature pattern mapping layer is constructed based on either a rule model or a data regression model. Based on the second training sample data, construct and train an artificial intelligence load prediction layer based on a machine learning regression model; The load feature pattern mapping layer and the artificial intelligence load prediction layer are connected in series to obtain the expected load prediction model, and the expected load prediction model is retrained as a whole.

5. The microgrid self-regulating energy management system incorporating AI as described in claim 4, characterized in that, The expected load forecasting module is also used for: Extract sample time information and the number of available sample interfaces from the sample data; Based on a predefined statistical analysis period, using sample time information as the first index and the number of available interfaces as the second index, the historical load data in the sample data is aggregated and statistically analyzed to determine the typical load characteristic patterns and probability distributions corresponding to the sample time information and the number of available interfaces, thus forming the first training sample data. Additional features are extracted from the sample data, and the typical load feature patterns and probability distributions are associated and combined with the additional features to form the second training sample data.

6. The microgrid self-regulating energy management system incorporating AI as described in claim 1, characterized in that, The adjustment margin optimization module is also used for: The optimization task is established with the objective function of maximizing the regulation capacity margin of the target microgrid within the preset future time period. The regulation capacity margin is obtained by correcting the difference between the available power and the expected load forecast information by a cost factor. The expected load prediction information and the real-time operating status information are used as input parameters for the optimization task; Under the premise of satisfying the preset constraints, the optimization task is solved in the rolling time domain, and the power limit value sequence of each charging interface in the future preset time period is dynamically output.

7. The microgrid self-regulating energy management system incorporating AI as described in claim 1, characterized in that, The adjustment margin optimization module is further configured to: include the following constraints for the optimization task: Cost constraints, defined based on electricity purchase costs and equipment operating loss costs, are used to limit the total economic cost corresponding to the power limit value; Safety operation constraints are defined based on the power balance limit of the target microgrid, the power capacity limit of the equipment, and the minimum charging requirements of users.

8. A microgrid self-regulating energy management method combining AI, characterized in that: The microgrid self-regulating energy management system, which incorporates AI as described in any one of claims 1-7, comprises: Obtain real-time operating status information of the target microgrid and corresponding charging station information; Based on the real-time operating status information and the charging station information, combined with the pre-constructed expected load prediction model, expected load prediction information of the target microgrid within a future preset management window is generated. With the goal of maximizing the regulation margin of the target microgrid in a future preset period, an optimization task is established and solved, and the power limit value corresponding to each charging interface is output. The charging load of the target microgrid is allocated and managed according to the power limit value.

9. A smart terminal for microgrid self-regulating energy management incorporating AI, characterized in that: The microgrid self-regulating energy management system incorporating AI as described in any one of claims 1-7 is applied to the smart terminal.

Images