Artificial intelligence-based user-side microgrid load regulation method, system and device

Abstract

The application provides a user-side micro-grid load regulation method, system and equipment based on artificial intelligence, and relates to the technical field of micro-grid management.The method comprises the following steps: analyzing multi-dimensional heterogeneous data to obtain load time series data, and determining the demand weight of each user according to power demand data through an analytic hierarchy process; processing the load time series data through empirical mode decomposition to obtain intrinsic mode function components; dynamically predicting the intrinsic mode function components, and obtaining a multi-modal load prediction result by weighting and fusing the prediction results through an attention mechanism; optimizing the multi-modal load prediction result in combination with the operation target of a micro-grid cluster to generate a basic load regulation strategy of the micro-grid cluster; and generating a final regulation strategy corresponding to each user in combination with the demand weight corresponding to the user according to the basic load regulation strategy, and regulating the final regulation strategy. The application significantly improves the regulation accuracy of the user-side micro-grid load.

Description

Technical Field

[0001] This invention relates to the field of microgrid management technology, and more specifically, to a user-side microgrid load regulation method, system, and equipment based on artificial intelligence. Background Technology

[0002] Driven by dual-carbon goals and the accelerated construction of new power systems, the penetration rate of distributed photovoltaic power generation and wind power, among other renewable energy sources, in user-side microgrids continues to increase, deeply coupling the operation of user-side microgrids with the characteristics of new energy output. Traditional load regulation methods typically rely on a single forecasting model to predict the load of user-side microgrids, and then perform centralized regulation on the user side based on the forecast results.

[0003] In related technologies, user-side microgrids are influenced by a combination of factors, including user energy consumption behavior, weather conditions, and electricity pricing policies. Furthermore, they are affected by differences in AC / DC load characteristics, industrial impulsive loads, and time-domain response variations. This results in complex, multi-dimensional, heterogeneous coupling characteristics in user-side microgrid loads, exhibiting high randomness and volatility. Therefore, traditional methods using a single prediction model and algorithm cannot deconstruct the coupling patterns of multiple load types, nor can they adapt to the regulation requirements of heterogeneous AC / DC loads. This leads to the failure of centralized regulation strategies under combined extreme operating conditions, further increasing the difficulty of balancing microgrid power. Summary of the Invention

[0004] The problem addressed by this invention is how to improve the control accuracy of user-side microgrid load.

[0005] To address the aforementioned issues, this invention provides a user-side microgrid load regulation method, system, and device based on artificial intelligence.

[0006] In a first aspect, the user-side microgrid load regulation method based on artificial intelligence of the present invention includes:

[0007] Acquire multi-dimensional heterogeneous data of the microgrid cluster on the user side and the electricity demand data of each user in the microgrid cluster;

[0008] Based on the analysis of the multi-dimensional heterogeneous data, the load time series data of the microgrid cluster is obtained, and the demand weight of each user in the microgrid cluster is determined by the analytic hierarchy process (AHP) based on the electricity demand data of the users.

[0009] The load time series data is processed by empirical mode decomposition to obtain the intrinsic mode function components of the multidimensional heterogeneous data.

[0010] The intrinsic mode function components are divided into high-frequency random components and low-frequency trend components. Then, the high-frequency random components and the low-frequency trend components are dynamically predicted by Kalman filtering algorithm and kernel extreme learning machine algorithm, respectively. The prediction results are weighted and fused by attention mechanism to obtain multimodal load prediction results.

[0011] Based on the multimodal load forecast results and combined with the operational objectives of the microgrid cluster, the basic load control strategy of the microgrid cluster is generated.

[0012] Based on the basic load control strategy and combined with the demand weights corresponding to the users, a final control strategy is generated for each user.

[0013] The user load in the microgrid cluster is regulated according to the final regulation strategy.

[0014] Optionally, acquiring multi-dimensional heterogeneous data of the microgrid cluster on the user side and electricity demand data of each user in the microgrid cluster includes:

[0015] According to the preset sampling interval, the AC / DC hybrid load power data, distributed photovoltaic power output data, fuel cell power output data and environmental meteorological data of the microgrid cluster are collected synchronously. The AC / DC hybrid load power data, distributed photovoltaic power output data, fuel cell power output data and environmental meteorological data are preprocessed to obtain the multi-dimensional heterogeneous data.

[0016] Obtain the interruptible power limit, maximum continuous interruption time, temperature-controlled load range, and demand price elasticity coefficient based on dynamic electricity price for each user's interruptible load;

[0017] The interruptible power limit, the maximum continuous interruption time, the temperature-controlled load range, and the demand price elasticity coefficient are used as the electricity demand data.

[0018] Optionally, the step of analyzing the multi-dimensional heterogeneous data to obtain the load time-series data of the microgrid cluster, and determining the demand weight of each user in the microgrid cluster using the analytic hierarchy process (AHP) based on the user's electricity demand data, includes:

[0019] A correlation analysis was performed on the load data, distributed power output data, and environmental meteorological data in the multi-dimensional heterogeneous data to obtain load dynamic characteristic indicators affected by environmental factors.

[0020] Based on the mapping relationship between the load dynamic characteristic index and environmental factors, a load power sequence indexed by timestamp is constructed.

[0021] The load power sequence is standardized to obtain the load time series data of the microgrid cluster;

[0022] A judgment matrix is ​​established based on the user's electricity demand data using a pre-defined hierarchical analysis model; the hierarchical analysis model includes multiple criterion layers.

[0023] The local weights of each user corresponding to each criterion layer are determined using the judgment matrix.

[0024] The local weights are weighted and combined to obtain the user's demand weight in the microgrid cluster.

[0025] Optionally, the step of processing the load time-series data through empirical mode decomposition to obtain the intrinsic mode function components of the multi-dimensional heterogeneous data includes:

[0026] Identify all local maxima and local minima in the load time series data;

[0027] Spline interpolation is performed on the local maxima and local minima respectively to obtain the upper and lower envelopes of the load time series data;

[0028] The mean envelope is obtained by taking the mean of the upper and lower envelopes at each data point.

[0029] A data sequence is obtained by combining the load time series data with the mean envelope;

[0030] The data sequence is separated by an iterative sieving process to obtain the intrinsic mode function components of the multidimensional heterogeneous data.

[0031] Optionally, dividing the intrinsic mode function components into high-frequency random components and low-frequency trend components includes:

[0032] Obtain the zero-crossing rate and energy percentage of each intrinsic mode function component;

[0033] The intrinsic mode function components whose zero-crossing rate is greater than or equal to a first preset threshold and whose energy percentage is less than a second preset threshold are divided into high-frequency random components;

[0034] The intrinsic mode function components whose zero-crossing rate is less than a third preset threshold and whose energy percentage is greater than or equal to a fourth preset threshold are classified as low-frequency trend components.

[0035] Optionally, the step of dynamically predicting the high-frequency random component and the low-frequency trend component using the Kalman filter algorithm and the kernel extreme learning machine algorithm, respectively, and then weighting and fusing the prediction results through an attention mechanism to obtain the multimodal load prediction result includes:

[0036] For the high-frequency random components, a single-step prediction is performed using the extended Kalman filter algorithm to obtain the high-frequency component prediction sequence;

[0037] For the low-frequency trend component, a multi-step prediction is performed using the kernel extreme learning machine algorithm to obtain the low-frequency component prediction sequence;

[0038] The weights corresponding to the high-frequency component prediction sequence and the low-frequency component prediction sequence are dynamically calculated through an attention mechanism.

[0039] Based on the weights corresponding to the high-frequency component prediction sequence and the low-frequency component prediction sequence, the high-frequency component prediction sequence and the low-frequency component prediction sequence are weighted and summed to obtain the multimodal load prediction result.

[0040] Optionally, the step of optimizing the microgrid cluster based on the multimodal load forecast results and the operational objectives of the microgrid cluster to generate the basic load control strategy for the microgrid cluster includes:

[0041] The multimodal load prediction results are input into the preset optimization model corresponding to the operation target. At the same time, the operation target of the microgrid cluster in the preset future rolling time domain is input into the preset optimization model.

[0042] The model prediction control framework of the preset optimization model is used to solve the problem at the beginning of each optimization cycle in the preset future rolling time domain, so as to obtain the load regulation pre-instruction sequence for multiple time periods in the preset future rolling time domain.

[0043] Based on the first load control preparatory instruction in the load control preparatory instruction sequence, the load control preparatory instruction sequence is continuously updated by repeatedly executing optimization solutions to obtain the basic load control strategy.

[0044] Optionally, generating a final control strategy for each user based on the basic load control strategy and the corresponding demand weight of the user includes:

[0045] The total cluster load adjustment amount in the basic load control strategy is taken as the total control task to be allocated, and the control task ratio of each user is determined according to the demand weight of each user.

[0046] Based on the control task ratio, the total control task is decomposed and allocated to each user, and a preliminary load control instruction is generated for each user.

[0047] By combining the specific electrical equipment constraints of each user with the electricity demand data, the feasibility of the initial load control command is verified and fine-tuned to obtain the final control strategy for each user.

[0048] Secondly, the user-side microgrid load control system based on artificial intelligence of the present invention includes:

[0049] The data acquisition unit is used to acquire multi-dimensional heterogeneous data of the microgrid cluster on the user side and the electricity demand data of each user in the microgrid cluster;

[0050] The user analysis unit is used to analyze the multi-dimensional heterogeneous data to obtain the load time series data of the microgrid cluster, and to determine the demand weight of each user in the microgrid cluster based on the user's electricity demand data using the analytic hierarchy process.

[0051] The data processing unit is used to process the load time series data through empirical mode decomposition to obtain the intrinsic mode function components of the multidimensional heterogeneous data.

[0052] The dynamic prediction unit is used to divide the intrinsic mode function components into high-frequency random components and low-frequency trend components, and then use the Kalman filter algorithm and the kernel extreme learning machine algorithm to dynamically predict the high-frequency random components and the low-frequency trend components respectively. The prediction results are weighted and fused through an attention mechanism to obtain the multimodal load prediction results.

[0053] The basic strategy generation unit is used to optimize the basic load control strategy of the microgrid cluster based on the multimodal load prediction results and the operation objectives of the microgrid cluster.

[0054] The user optimization unit is used to generate a final control strategy for each user based on the basic load control strategy and the corresponding demand weight of the user.

[0055] The control unit is used to control the user load in the microgrid cluster according to the final control strategy.

[0056] Thirdly, the electronic device of the present invention includes: a processor and a memory, the memory being used to store a computer program;

[0057] When the computer program is loaded by the processor, it causes the processor to execute the user-side microgrid load regulation method based on artificial intelligence as described above.

[0058] The present invention relates to an artificial intelligence-based user-side microgrid load control method, system, and equipment. First, it comprehensively acquires multi-dimensional heterogeneous data of the microgrid cluster and the electricity demand data of each user, providing comprehensive and realistic basic data support. Then, it uses the analytic hierarchy process (AHP) to accurately quantify the demand weight of each user, thereby analyzing individual user differences. Subsequently, it employs empirical mode decomposition to decompose the complex load time-series data into high-frequency random components and low-frequency trend components. Kalman filtering and kernel extreme learning machine algorithms are applied to dynamically predict the different characteristics of these two types of components. Since high-frequency random components are characterized by strong volatility and low regularity, while low-frequency trend components are characterized by strong regularity and low dispersion, the prediction of components with different characteristics is solved by using Kalman filtering and kernel extreme learning machine algorithms respectively. The challenge lies in effectively adapting to the high randomness and trend characteristics of load data. Then, an attention mechanism is used to weight and fuse the prediction results of the two models, significantly improving the accuracy of multimodal load prediction. Next, based on the multimodal load prediction results and combined with the microgrid cluster operation objectives, a basic load control strategy is optimized to ensure the global rationality of the control strategy. Then, the basic control strategy is differentiated by combining previously determined user demand weights to generate a final control strategy tailored to the actual situation of each user, achieving precise control transmission from the cluster as a whole to individual users. Finally, through a layered, progressive, and precisely adapted process, the coupling patterns of multiple load types are effectively deconstructed, adapting to the control needs of AC / DC heterogeneous loads, avoiding the failure of centralized control under combined extreme operating conditions, and significantly improving the control accuracy of the microgrid load on the user side. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating a user-side microgrid load control method based on artificial intelligence in one embodiment of the present invention.

[0060] Figure 2 This is a schematic diagram of the time window for rolling optimization in one embodiment of the present invention;

[0061] Figure 3 This is a schematic diagram of the structure of a user-side microgrid load control system based on artificial intelligence in another embodiment of the present invention. Detailed Implementation

[0062] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the accompanying drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0063] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0064] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to"; the term "based on" means "at least partially based on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; and the term "optionally" means "optional embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first," "second," etc., mentioned in this invention are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0065] It should be noted that the terms "a" and "a plurality of" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0066] The names of the messages or information exchanged between the multiple devices in the embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0067] Combination Figure 1 As shown in the figure, an embodiment of the present invention provides a user-side microgrid load regulation method based on artificial intelligence, characterized in that it includes:

[0068] Acquire multi-dimensional heterogeneous data of the microgrid cluster on the user side and the electricity demand data of each user in the microgrid cluster.

[0069] Specifically, acquiring multi-dimensional heterogeneous data of the user-side microgrid cluster and the electricity demand data of each user is the fundamental prerequisite for the entire control method. User-side microgrid loads are comprehensively affected by various factors such as user energy consumption behavior, weather conditions, and electricity pricing policies. Furthermore, AC and DC load characteristics differ significantly, and industrial impulsive loads and time-domain response differences further exacerbate the complexity of the load. Relying on only one type of data cannot fully depict the dynamic changes in the load. Each user's electricity demand exhibits individual differences, such as electricity priority and load rigidity, directly affecting the rationality of the control strategy and user acceptance. Therefore, comprehensively collecting multi-dimensional heterogeneous data and user electricity demand data can fully capture the multi-dimensional heterogeneous coupling characteristics of the load and individual user differences, providing a comprehensive and accurate data source for subsequent load forecasting, weight determination, and strategy formulation, avoiding control measures that are out of touch with reality due to incomplete data.

[0070] Based on the analysis of the multi-dimensional heterogeneous data, the load time series data of the microgrid cluster is obtained. Then, using the analytic hierarchy process (AHP), the demand weight of each user in the microgrid cluster is determined based on the user's electricity demand data.

[0071] Specifically, the load time-series data of the microgrid cluster is obtained through multi-dimensional heterogeneous data analysis. Essentially, this involves integrating and structuring dispersed, multi-source data. Multi-dimensional heterogeneous data comes from different sources and types, making it difficult to directly reflect the inherent patterns of load changes over time. By analyzing and integrating this data to form load time-series data, the fluctuation characteristics of load at different times can be intuitively reflected. Simultaneously, the analytic hierarchy process (AHP) is used to determine the demand weight of each user because user electricity demand includes many qualitative indicators, such as electricity priority and load importance, which are difficult to quantify directly. The AHP can transform these qualitative indicators into quantitative weight values, clarifying the importance proportion of different users in the microgrid cluster. This solves the problem of traditional centralized control ignoring individual user differences, providing a quantitative basis for subsequently decomposing cluster-level control strategies into user-level personalized strategies, ensuring that control measures are both comprehensive and tailored to actual user needs.

[0072] The load time series data is processed by empirical mode decomposition to obtain the intrinsic mode function components of the multidimensional heterogeneous data.

[0073] Specifically, empirical mode decomposition (EMD) is used to process load time-series data to obtain intrinsic mode function (EMF) components. The load time-series data of a microgrid cluster is essentially a complex signal containing multiple frequency components. It exhibits both low-frequency trend changes caused by users' stable daily electricity consumption and high-frequency random fluctuations caused by sudden weather changes and unexpected electricity demand. A single algorithm struggles to simultaneously adapt to these two distinct characteristics. Directly predicting the raw load time-series data leads to decreased prediction accuracy due to interference between different frequency components. EMD, based on the signal's inherent time-scale characteristics, adaptively decomposes complex load time-series data into EMF components of different frequencies, effectively separating low-frequency trend components from high-frequency random components. This lays the foundation for selecting appropriate prediction algorithms for different component characteristics, improving the accuracy of load forecasting.

[0074] The intrinsic mode function components are divided into high-frequency random components and low-frequency trend components. Then, the high-frequency random components and the low-frequency trend components are dynamically predicted by Kalman filtering algorithm and kernel extreme learning machine algorithm, respectively. The prediction results are weighted and fused by attention mechanism to obtain the multimodal load prediction results.

[0075] Specifically, after dividing the intrinsic mode function components into high-frequency random components and low-frequency trend components, dynamic prediction is performed using the Kalman filter algorithm and the kernel extreme learning machine algorithm, respectively. A weighted fusion using an attention mechanism further improves the accuracy of load prediction. High-frequency random components are characterized by strong volatility and low regularity. The Kalman filter algorithm, with its excellent nonlinear tracking and state estimation capabilities, can accurately capture the dynamic fluctuation patterns of these components. Low-frequency trend components, on the other hand, exhibit strong regularity and stability. The kernel extreme learning machine algorithm, with its efficient regression prediction capabilities, can accurately fit the trends of these components. The targeted application of these two algorithms solves the prediction challenges of components with different characteristics, avoiding the inadequacy of a single algorithm for complex load components. Furthermore, the attention mechanism dynamically allocates fusion weights based on the prediction accuracy and importance of different components. Compared to fixed-weight fusion, this approach highlights the contribution of components with better prediction performance to the final result, further improving the accuracy of multimodal load prediction.

[0076] Based on the multimodal load forecast results and combined with the operational objectives of the microgrid cluster, the basic load control strategy of the microgrid cluster is generated through optimization.

[0077] Specifically, based on the multimodal load forecast results and the operational objectives of the microgrid cluster, optimization is performed to generate a cluster-level basic load control strategy. The multimodal load forecast results accurately reflect the future load change trends, ensuring that the control strategy is based on reliable load prediction and avoiding strategy failure due to forecast deviations. The operational objectives of the microgrid cluster typically include maximizing renewable energy consumption, minimizing operating costs, and optimizing power supply reliability. Optimization based on these objectives allows the basic load control strategy to consider the overall interests of the cluster and avoid conflicts between local control behaviors and global goals. The basic load control strategy generated in the above steps provides a scientific and reasonable global framework for subsequent user-level strategy decomposition, clarifying the control direction and overall requirements at the cluster level, and ensuring that all user control behaviors revolve around the overall cluster objectives.

[0078] Based on the basic load control strategy and the corresponding demand weights of the users, a final control strategy is generated for each user.

[0079] Specifically, based on the basic load control strategy and the user's corresponding demand weight, a final control strategy is generated for each user. The basic load control strategy is formulated globally for the entire cluster; if directly applied to all users, it would lead to poor control effects due to differences in user needs, and might even cause user resistance. By decomposing and adjusting the basic control strategy based on user demand weights, the cluster-level global control objectives can be transformed into personalized solutions adapted to each user's actual situation. For users with high demand weights and strong load rigidity, a more lenient control strategy can be formulated to prioritize their electricity needs; for users with low demand weights and strong load flexibility, a more targeted control strategy can be formulated to fully explore their load adjustment potential. This step achieves precise control transmission from the cluster-wide level to the individual user, ensuring the achievement of the overall cluster control objectives while meeting the differentiated electricity needs of users, thus improving the acceptability and effectiveness of control.

[0080] The user load in the microgrid cluster is regulated according to the final regulation strategy.

[0081] Specifically, the final user-level control strategy fully considers the multi-dimensional heterogeneous coupling characteristics of microgrid cluster loads and the uncertainty of renewable energy output, while also taking into account the overall operational goals of the cluster and the individual differences of users. This effectively adapts to the control requirements of heterogeneous AC and DC loads. By implementing this control strategy, the loads of different users can be precisely adjusted, achieving cluster-level power balance. This avoids the failure problem of traditional centralized control under combined extreme operating conditions, significantly reducing the difficulty of microgrid power balancing, while simultaneously improving renewable energy absorption rates, reducing operating costs, and achieving precise and efficient control of user-side microgrid loads.

[0082] The present invention relates to an artificial intelligence-based user-side microgrid load control method, system, and equipment. First, it comprehensively acquires multi-dimensional heterogeneous data of the microgrid cluster and the electricity demand data of each user, providing comprehensive and realistic basic data support. Then, it uses the analytic hierarchy process (AHP) to accurately quantify the demand weight of each user, thereby analyzing individual user differences. Subsequently, it employs empirical mode decomposition to decompose the complex load time-series data into high-frequency random components and low-frequency trend components. Kalman filtering and kernel extreme learning machine algorithms are applied to dynamically predict the different characteristics of these two types of components. Since high-frequency random components are characterized by strong volatility and low regularity, while low-frequency trend components are characterized by strong regularity and low dispersion, the prediction of components with different characteristics is solved by using Kalman filtering and kernel extreme learning machine algorithms respectively. The challenge lies in effectively adapting to the high randomness and trend characteristics of load data. Then, an attention mechanism is used to weight and fuse the prediction results of the two models, significantly improving the accuracy of multimodal load prediction. Next, based on the multimodal load prediction results and combined with the microgrid cluster operation objectives, a basic load control strategy is optimized to ensure the global rationality of the control strategy. Then, the basic control strategy is differentiated by combining previously determined user demand weights to generate a final control strategy tailored to the actual situation of each user, achieving precise control transmission from the cluster as a whole to individual users. Finally, through a layered, progressive, and precisely adapted process, the coupling patterns of multiple load types are effectively deconstructed, adapting to the control needs of AC / DC heterogeneous loads, avoiding the failure of centralized control under combined extreme operating conditions, and significantly improving the control accuracy of the microgrid load on the user side.

[0083] Optionally, acquiring multi-dimensional heterogeneous data of the microgrid cluster on the user side and electricity demand data of each user in the microgrid cluster includes:

[0084] According to the preset sampling interval, the AC / DC hybrid load power data, distributed photovoltaic power output data, fuel cell power output data and environmental meteorological data of the microgrid cluster are collected synchronously. The AC / DC hybrid load power data, distributed photovoltaic power output data, fuel cell power output data and environmental meteorological data are preprocessed to obtain the multi-dimensional heterogeneous data.

[0085] Obtain the interruptible power limit, maximum continuous interruption time, temperature-controlled load range, and demand price elasticity coefficient based on dynamic electricity price for each user's interruptible load;

[0086] The interruptible power limit, the maximum continuous interruption time, the temperature-controlled load range, and the demand price elasticity coefficient are used as the electricity demand data.

[0087] Specifically, for multi-dimensional heterogeneous data, AC / DC hybrid load power data, distributed photovoltaic power output data, fuel cell power output data, and environmental meteorological data are collected synchronously at preset sampling intervals. These three types of data correspond to load characteristics, new energy power output characteristics, and influencing factors, respectively, comprehensively covering the core data dimensions of microgrid cluster operation. Then, data noise and outliers are removed through preprocessing to ensure data quality. For electricity demand data, the interruptible power limit, maximum continuous interruption time (defining the load regulation capacity boundary), temperature-controlled load range (clarifying the rigid load operation range), and demand price elasticity coefficient based on dynamic electricity price (reflecting the user's response sensitivity to electricity price) of the user's interruptible load are obtained in a targeted manner. These quantitative indicators are used as electricity demand data to achieve accurate characterization of individual user electricity consumption characteristics. For example, by setting a preset sampling interval of 15 minutes, smart meters, distributed power monitoring terminals, and meteorological sensors deployed in the microgrid cluster are used to simultaneously collect AC load power, DC load power (such as electric vehicle charging load and LED lighting load), distributed photovoltaic power output, fuel cell power output, and environmental meteorological data such as light intensity, wind speed, and temperature. Outliers are removed from the collected data using the 3σ principle, and the data is normalized to obtain standardized multi-dimensional heterogeneous data. Through user-side interaction platforms and intelligent device monitoring, the interruptible load power limit for industrial users is 50kW, the maximum continuous interruption time is 2 hours, the temperature control load range for residential users is 22℃-26℃, and the demand price elasticity coefficient for commercial users based on dynamic electricity pricing is -0.8. These specific parameters are used as the electricity demand data for each user to complete the entire data collection process.

[0088] In this embodiment of the invention, by clearly defining the specific content, method, and quantitative indicators of data collection, the comprehensiveness and synchronicity of multi-dimensional heterogeneous data are ensured. This allows for the complete capture of the load coupling characteristics, renewable energy output patterns, and environmental influencing factors of the AC / DC hybrid microgrid, providing a high-quality data foundation for subsequent empirical mode decomposition and multi-algorithm prediction. Furthermore, it achieves the quantification and precision of electricity demand data, avoiding the ambiguity of traditional data collection. This makes the determination of demand weights based on the analytic hierarchy process more scientific and objective. At the same time, it provides a clear quantitative basis for the subsequent decomposition of basic control strategies into user-level personalized strategies, effectively improving the operability and accuracy of the entire control method.

[0089] Optionally, the step of analyzing the multi-dimensional heterogeneous data to obtain the load time-series data of the microgrid cluster, and determining the demand weight of each user in the microgrid cluster using the analytic hierarchy process (AHP) based on the user's electricity demand data, includes:

[0090] A correlation analysis was performed on the load data, distributed power output data, and environmental meteorological data in the multi-dimensional heterogeneous data to obtain load dynamic characteristic indicators affected by environmental factors.

[0091] Based on the mapping relationship between the load dynamic characteristic index and environmental factors, a load power sequence indexed by timestamp is constructed.

[0092] The load power sequence is standardized to obtain the load time series data of the microgrid cluster;

[0093] A judgment matrix is ​​established based on the user's electricity demand data using a pre-defined hierarchical analysis model; the hierarchical analysis model includes multiple criterion layers.

[0094] The local weights of each user corresponding to each criterion layer are determined using the judgment matrix.

[0095] The local weights are weighted and combined to obtain the user's demand weight in the microgrid cluster.

[0096] Specifically, a correlation analysis is first conducted on load data, distributed power generation output data, and environmental meteorological data from multi-dimensional heterogeneous data to uncover the intrinsic relationship between environmental factors and load changes, and to extract load dynamic characteristic indicators. Then, based on the mapping relationship between load dynamic characteristic indicators and environmental factors, a timestamp-indexed load power sequence is constructed. After standardization, structured load time-series data is formed, solving the problem that the original multi-source data is messy and difficult to use directly for prediction. At the same time, with the help of a pre-set hierarchical analysis model containing multiple criterion layers, a judgment matrix is ​​established in combination with user electricity demand data. The local weights of each criterion layer corresponding to each user are first calculated, and then the global demand weight is obtained through weighted synthesis. The weight determination method of the abstract hierarchical analysis method in the sovereign is transformed into an operable quantitative process, ensuring that the weight allocation is both in line with the individual electricity consumption characteristics of users and in line with the overall control logic of the microgrid cluster. In a preferred embodiment of the present invention, multi-dimensional heterogeneous data at 15-minute intervals from a microgrid cluster are selected, including AC / DC load power data, distributed photovoltaic output data, fuel cell output data, and environmental meteorological data such as irradiance, temperature, and wind speed. Pearson correlation analysis reveals that the correlation coefficients between photovoltaic output and irradiance and temperature are 0.85 and 0.62, respectively, and the correlation coefficient between DC load power and electric vehicle charging periods is 0.78. Based on this, dynamic characteristic indicators such as load response sensitivity and load fluctuation coefficient are extracted. Based on the mapping relationship between these indicators and environmental factors, a 24-hour load power sequence indexed by timestamps is constructed. Min-max standardization is used to map the data to the [0,1] interval, obtaining standardized load time-series data. Specifically, load response sensitivity, load fluctuation coefficient, and environmental factors are constructed... To generate a 24-hour load power sequence by mapping the elements, it is necessary to first clarify the correlation strength between each environmental factor (such as sunlight, temperature, and wind speed) and the load response sensitivity (the amplitude of load changes with the environment) and the load fluctuation coefficient (the degree of dispersion of load fluctuation) through correlation analysis. For example, when the sunlight intensity increases, the output of distributed photovoltaics will increase and reduce the load response sensitivity, and when the temperature rises, the increase in cooling load will increase the load fluctuation coefficient. Based on this correlation, using the timestamp as an index, the environmental factor data at each time point is substituted into a pre-trained mapping model (such as multiple linear regression or BP neural network) to calculate the load response sensitivity and fluctuation coefficient at the corresponding time. Combined with the load benchmark value under the same historical environmental conditions, the dynamic adjustment amount corresponding to the sensitivity and fluctuation coefficient is superimposed to finally obtain the load power value at each timestamp, which is then concatenated to form a 24-hour load power sequence.The criterion layers of the pre-defined hierarchical analysis model are interruptible load capacity, load rigidity, and electricity price response sensitivity. For the electricity demand data of 3 industrial users, 5 commercial users, and 20 residential users, pairwise comparison judgment matrices are established. The local weights of each user under the interruptible load capacity criterion are calculated as 0.3-0.7, under the load rigidity criterion as 0.2-0.8, and under the electricity price response sensitivity criterion as 0.1-0.6 using the eigenvalue method. Then, the weights of each criterion layer are weighted and summed to obtain the global demand weight of each user, which is between 0.15 and 0.72.

[0097] In this embodiment of the invention, by using correlation analysis and structured processing, scattered multi-dimensional heterogeneous data is transformed into standardized load time-series data. This preserves the dynamic pattern of load changes over time while eliminating interference caused by differences in data dimensions, providing a high-quality and highly adaptable data foundation for subsequent empirical mode decomposition and multi-algorithm prediction. Simultaneously, by quantifying user demand weights through a multi-criteria hierarchical analysis model, the one-sidedness of single-dimensional weight allocation is avoided. The weights reflect differences in core electricity consumption characteristics such as user interruptible load capacity and load rigidity, and through weighted synthesis, a balance is achieved between individual characteristics and overall cluster control needs. This provides a scientific and precise quantitative basis for subsequently decomposing basic control strategies into personalized user strategies, effectively improving the targeting and rationality of control strategies.

[0098] Optionally, the step of processing the load time-series data through empirical mode decomposition to obtain the intrinsic mode function components of the multi-dimensional heterogeneous data includes:

[0099] Identify all local maxima and local minima in the load time series data;

[0100] Spline interpolation is performed on the local maxima and local minima respectively to obtain the upper and lower envelopes of the load time series data;

[0101] The mean envelope is obtained by taking the mean of the upper and lower envelopes at each data point.

[0102] A data sequence is obtained by combining the load time series data with the mean envelope;

[0103] The data sequence is separated by an iterative sieving process to obtain the intrinsic mode function components of the multidimensional heterogeneous data.

[0104] Specifically, the process begins by identifying all local maxima and minima in the load time series data, which forms the basis for constructing the load data fluctuation boundary. Then, spline interpolation is used to fit the upper and lower envelopes to determine the fluctuation range of the load data. Subsequently, the mean envelope is obtained by calculating the average of the two envelopes at each data point, thereby removing the trend component from the load data. Then, the difference between the load time series data and the mean envelope is used to obtain the preliminary data sequence. Finally, through an iterative screening process, components that satisfy the intrinsic mode function conditions are continuously separated to ensure that the decomposed components can accurately reflect the different frequency characteristics of the load data, providing reliable support for subsequent component classification and targeted prediction. In a preferred embodiment of the present invention, 24-hour load time-series data (sampling interval 15 minutes, a total of 96 data points) after standardization of the microgrid cluster is selected. First, 28 local maxima and 26 local minima in the data are identified using the sliding window method. Cubic spline interpolation is used to fit the maxima and minima respectively to obtain smooth upper and lower envelopes. The mean of the two envelopes at each data point is calculated and connected to form a mean envelope. The mean envelope is subtracted from the original load time-series data to obtain the initial data sequence. The above envelope construction and difference calculation process is repeated for the data sequence. When the mean of the upper and lower envelopes of the data sequence (the arithmetic mean of the values ​​of the upper and lower envelopes at each data point) is close to zero and the number of extreme points and the number of zero-crossing points do not differ by more than 1, the data sequence is determined as the first intrinsic mode function component. Iteration and screening continue to obtain 6 intrinsic mode function components of different frequencies, covering the high-frequency fluctuations and low-frequency trend characteristics of the load data.

[0105] In this embodiment of the invention, an adaptive decomposition of load time series data is achieved through a refined process of progressively identifying extreme points, constructing envelopes, and iteratively screening. This process eliminates the need for pre-defined basis functions and accurately separates the fluctuation and trend components of different frequencies from the load data. The resulting intrinsic mode function components exhibit good orthogonality and physical meaning. The decomposition process fully preserves the original dynamic characteristics of the load data, avoiding the distortion caused by complex load fluctuations in traditional decomposition methods. This makes subsequent classification of high-frequency random components and low-frequency trend components more accurate, laying a high-quality foundation for the targeted application of Kalman filtering and kernel extreme learning machine algorithms, thereby improving the overall accuracy and reliability of load forecasting.

[0106] Optionally, dividing the intrinsic mode function components into high-frequency random components and low-frequency trend components includes:

[0107] Obtain the zero-crossing rate and energy percentage of each intrinsic mode function component;

[0108] The intrinsic mode function components whose zero-crossing rate is greater than or equal to a first preset threshold and whose energy percentage is less than a second preset threshold are divided into high-frequency random components;

[0109] The intrinsic mode function components whose zero-crossing rate is less than a third preset threshold and whose energy percentage is greater than or equal to a fourth preset threshold are classified as low-frequency trend components.

[0110] Specifically, firstly, the zero-crossing rate and energy percentage of each intrinsic mode function (EMF) component are obtained. Specifically, for the zero-crossing rate, the EMF component data sequence comes from the output of Empirical Mode Decomposition (EMD), that is, after performing EMD decomposition on the microgrid cluster load time-series data (after standardization), the individual EMF components obtained are denoted as the sequence x = [x1, x2, ..., x...]. N ], where N is the data length (determined by the sampling interval and total duration of the load time series data, for example, N=96 when the sampling interval is 24 hours and 15 minutes).

[0111] Pair adjacent data points (x) one by one i x i+1 (i from 1 to N) 1) Make a judgment: If (x) is satisfied i ×x i+1 If the value is <0 (i.e., adjacent data points have opposite signs and directly cross the zero value), or if the current data point is zero and the next data point is non-zero, it is recorded as 1 zero crossing, and the cumulative number of zero crossings is Z.

[0112] The formula for calculating the zero-crossing rate ZCR is: ZCR = ;

[0113] Where the denominator is N 1 is because N data points contain N One adjacent data pair ensures the rationality of the ratio calculation.

[0114] Regarding energy proportion, the energy proportion of a single intrinsic mode function component refers to the ratio of the total energy of that component to the sum of the total energy of all intrinsic mode function components, reflecting the component's energy contribution weight to the original load time series data.

[0115] For the kth intrinsic mode function component (denoted as IMF) k ), whose data sequence is x k =[x k1 x k2 , ..., x kN Total energy E k The sum of squares method is used for calculation (which conforms to the classical definition of signal energy), and the formula is as follows:

[0116] ;

[0117] in, This represents the value of the i-th data point of the k-th intrinsic mode function component.

[0118] Assume that the EMD decomposition yields M IMF components (M is determined by the number of decomposition iterations, until the remaining components are monotonically trending terms), and the total energy E of all components is... total for:

[0119] ;

[0120] Energy percentage R k The calculation formula is: R k = .

[0121] The zero-crossing rate reflects the intensity of fluctuations in a component (the more frequent the fluctuations, the higher the zero-crossing rate), while the energy proportion reflects the component's contribution to the overall load data (trend components typically have a higher energy proportion). By setting first to fourth preset thresholds, components with high zero-crossing rates and low energy proportions are defined as high-frequency random components, while components with low zero-crossing rates and high energy proportions are defined as low-frequency trend components. This ensures that the classification results accurately match the essential characteristics of the two types of components, providing a reliable basis for subsequent targeted selection of Kalman filtering and kernel extreme learning machine algorithms for prediction. Specifically, the first preset threshold is greater than the third preset threshold, and the second preset threshold is less than the fourth preset threshold. The first preset threshold is the lower limit for judging the zero-crossing rate of high-frequency random components (a zero-crossing rate ≥ the first preset threshold is required to meet the characteristics of high-frequency components), and the third preset threshold is the upper limit for judging the zero-crossing rate of low-frequency trend components (a zero-crossing rate < the third preset threshold is required to meet the characteristics of low-frequency components). The core characteristics of high-frequency random components are dense fluctuations and high zero-crossing rates, while the core characteristics of low-frequency trend components are smooth fluctuations and low zero-crossing rates. Therefore, to clearly distinguish between the two types of components, the first preset threshold needs to be greater than the third preset threshold. Similarly, the second preset threshold is the upper limit for judging the energy proportion of high-frequency random components (only when the energy proportion is less than the second preset threshold does it meet the characteristics of high-frequency components), and the fourth preset threshold is the lower limit for judging the energy proportion of low-frequency trend components. Considering the characteristics of low energy proportion of high-frequency components and high energy proportion of low-frequency components, the second preset threshold needs to be less than the fourth preset threshold.

[0122] In a preferred embodiment of the present invention, for a microgrid cluster, six intrinsic mode function components (IMF1-IMF6) obtained by empirical mode decomposition, each component contains 96 data points (sampling interval of 24 hours and 15 minutes). When calculating the zero-crossing rate, the number of times IMF1-IMF6 crossed the zero value are counted as 42, 36, 31, 12, 8, and 5 times, respectively. Dividing these by the data length of 96, the zero-crossing rates are 0.4375, 0.375, 0.3229, 0.125, 0.0833, and 0.0521, respectively. When calculating the energy proportion, the energy of each component is calculated as follows: The values ​​are 128, 165, 203, 452, 587, and 732, with a total energy of 2267. The resulting energy percentages are 5.65%, 7.28%, 8.95%, 19.94%, 25.90%, and 32.29%, respectively. The first preset threshold is set to 0.3, the second to 10%, the third to 0.3, and the fourth to 15%. Based on these settings, IMF1-IMF3, with a zero-crossing rate ≥ 0.3 and an energy percentage < 10%, are classified as high-frequency random components, while IMF4-IMF6, with a zero-crossing rate < 0.3 and an energy percentage ≥ 15%, are classified as low-frequency trend components.

[0123] In this embodiment of the invention, by clarifying the specific calculation methods for zero-crossing rate and energy proportion, the quantitative acquisition of component classification indicators is realized, avoiding the uncertainty of subjective judgment. By combining the collaborative judgment of the two types of indicators, the high-frequency fluctuation components and low-frequency trend components in the load data can be accurately separated, making the classification results more objective and scientific. The characteristics of the classified components are more compatible with the subsequently selected Kalman filter algorithm and kernel extreme learning machine algorithm, which can give full play to the core advantages of the two types of algorithms in tracking random fluctuations and fitting trend changes, thereby improving the accuracy of multimodal load prediction results.

[0124] Optionally, the step of dynamically predicting the high-frequency random component and the low-frequency trend component using the Kalman filter algorithm and the kernel extreme learning machine algorithm, respectively, and then weighting and fusing the prediction results through an attention mechanism to obtain the multimodal load prediction result includes:

[0125] For the high-frequency random components, a single-step prediction is performed using the extended Kalman filter algorithm to obtain the high-frequency component prediction sequence;

[0126] For the low-frequency trend component, a multi-step prediction is performed using the kernel extreme learning machine algorithm to obtain the low-frequency component prediction sequence;

[0127] The weights corresponding to the high-frequency component prediction sequence and the low-frequency component prediction sequence are dynamically calculated through an attention mechanism.

[0128] Based on the weights corresponding to the high-frequency component prediction sequence and the low-frequency component prediction sequence, the high-frequency component prediction sequence and the low-frequency component prediction sequence are weighted and summed to obtain the multimodal load prediction result.

[0129] Specifically, to address the characteristics of high-frequency random components—strong volatility and low regularity—an extended Kalman filter algorithm is selected for single-step prediction, leveraging its excellent nonlinear signal tracking and state estimation capabilities to accurately capture the instantaneous fluctuations of the components. For low-frequency trend components—characterized by strong regularity and low dispersion—a kernel extreme learning machine algorithm is employed for multi-step prediction, utilizing its efficient regression fitting capabilities to accurately predict the long-term trends of the components. Simultaneously, an attention mechanism is introduced, dynamically calculating weights based on the real-time prediction accuracy and component importance of the two types of prediction sequences, rather than using fixed weight fusion. This ensures that the multimodal load prediction results not only accurately reflect the essential characteristics of each component but also dynamically highlight the contribution of high-quality prediction results, thereby improving overall prediction accuracy. In a preferred embodiment of the present invention, the microgrid cluster is partitioned into three high-frequency random components (IMF1-IMF3) and three low-frequency trend components (IMF4-IMF6). For IMF1-IMF3, an extended Kalman filter algorithm is used to construct a state-space model, with the process noise covariance Q=0.01 and the observation noise covariance R=0.05. Single-step prediction with a step size of 15 minutes is performed to obtain the high-frequency component prediction sequence P_high=[P1_high, P2_high, ..., P96_high]. For IMF4-IMF6, a kernel extreme learning machine algorithm is used, with the RBF kernel function (kernel parameter σ=2) selected. The number of hidden layer nodes is set to 50. After training the model based on 12 months of historical data, multi-step prediction is performed. The prediction process yields a low-frequency component prediction sequence P_low=[P1_low, P2_low, ..., P96_low]. Weights are calculated using an attention mechanism: for each timestamp t, weights ω_high(t) and ω_low(t) (ω_high(t)+ω_low(t)=1) are dynamically generated based on the historical prediction errors of the two prediction sequences (smaller errors result in higher weights). Specifically, ω_high=0.32 and ω_low=0.68 at t=30, and ω_high=0.45 and ω_low=0.55 at t=60. Finally, the multimodal load prediction result is obtained using P_pred(t)=ω_high(t)×P_high(t)+ω_low(t)×P_low(t).

[0130] In this embodiment of the invention, by specifically applying the extended Kalman filter algorithm and the kernel extreme learning machine algorithm, the core advantages of both algorithms in tracking random fluctuations and fitting trend changes are fully utilized, solving the problem that a single algorithm cannot adapt to predictions with different characteristic components. Simultaneously, the introduction of dynamic weights in the attention mechanism avoids the limitation of fixed-weight fusion in adapting to dynamic changes in prediction accuracy. It can highlight the contribution of the component with better prediction performance to the final result in real time, enabling the multimodal load prediction results to accurately capture both high-frequency load fluctuations and accurately predict low-frequency load trends, significantly improving the accuracy and reliability of the prediction results.

[0131] Optionally, the step of optimizing the microgrid cluster based on the multimodal load forecast results and the operational objectives of the microgrid cluster to generate the basic load control strategy for the microgrid cluster includes:

[0132] The multimodal load prediction results are input into the preset optimization model corresponding to the operation target. At the same time, the operation target of the microgrid cluster in the preset future rolling time domain is input into the preset optimization model.

[0133] The model prediction control framework of the preset optimization model is used to solve the problem at the beginning of each optimization cycle in the preset future rolling time domain, so as to obtain the load regulation pre-instruction sequence for multiple time periods in the preset future rolling time domain.

[0134] Based on the first load control preparatory instruction in the load control preparatory instruction sequence, the load control preparatory instruction sequence is continuously updated by repeatedly executing optimization solutions to obtain the basic load control strategy.

[0135] Specifically, firstly, high-precision multimodal load forecasting results, distributed power generation output forecasts, and electricity price information are used as feedforward inputs and combined with an optimization model aiming to minimize total operating costs. Then, within the model predictive control framework at the start of each rolling optimization cycle, the constrained optimization problem is solved, generating a sequence of pre-emptive load control instructions for multiple future time periods. However, only the first instruction for the current time period is executed, achieving precise optimization in the finite time domain. Finally, by introducing a feedback correction mechanism, subsequent instructions are dynamically fine-tuned based on prediction errors occurring during real-time operation, forming a closed-loop rolling optimization mechanism of "prediction-optimization-execution-feedback." This dynamically transforms static forecast data into a basic load control strategy that adapts to real-time system changes and possesses both economic efficiency and robustness. Specifically, combining... Figure 2 As shown, at time k, the control instructions are optimized in the time domain for the next M time periods. Only the instructions for the first time period are executed. At time k+1, the time window is moved forward by one time period. Based on the latest data, the instructions for the next M time periods are re-optimized, forming a cyclical and rolling optimization mode.

[0136] In a preferred embodiment of the present invention, at the beginning of a certain optimization cycle, the system, based on the load and renewable energy output data up to the current time, uses a trained EMD-KELM-EKF combined model to predict the multimodal load values ​​for the next 4 hours (i.e., 9:00 to 13:00) at 15-minute intervals. These predicted values, photovoltaic output prediction values, and future electricity prices are input into a preset mixed-integer linear programming optimization model. The model is solved with the objective of minimizing the sum of the total electricity purchase cost and the load interruption compensation cost for the next 4 hours, resulting in preliminary control commands for 16 time intervals: 9:00-9:15, 9:15-9:30, ..., 12:45-13:00. The system then executes commands only for the first time interval, 9:00-9:15, such as reducing a certain type of flexible load by 50kW. At 9:15, the start of the next cycle, the system collects actual data from 9:00 to 9:15, rolls the prediction window to 9:15-13:15, resolves the optimization model, and executes new instructions from 9:15 to 9:30, repeating this cycle. Simultaneously, if a significant deviation is found between the actual load and the predicted value from 9:00 to 9:15, the feedback correction module adjusts subsequent pre-defined instructions proportionally to ensure overall control effectiveness. It is worth noting that the mixed-integer linear programming optimization model consists of three parts: decision variables, objective function, and constraints. The mixed-integer linear programming optimization model is constructed with the objective of minimizing the sum of the total electricity purchase cost and the load interruption compensation cost for the next 4 hours. The decision variables include continuous variables such as the main grid interaction power and the user load reduction power, as well as 0-1 integer variables representing the load interruption execution status. The objective function is obtained by accumulating the product of the main grid electricity purchase price and the interaction power for each time period, and the product of the load interruption compensation cost and the reduction power for each user unit. The constraints cover the core power balance constraints (main grid interaction power + renewable energy consumption power + total load reduction power = multimodal load forecast value), user load reduction constraints (not exceeding the interruptible load limit and only reducing when an interruption is executed), renewable energy consumption constraints (not exceeding the predicted output), main grid interaction power constraints (complying with the distribution network access limit), and integer variable logical constraints. By integrating multimodal load forecast values, renewable energy output forecast values, future electricity prices, and user electricity demand data, the model is not only suitable for the actual scenario of microgrid cluster regulation, but also achieves the synergy of discrete decision-making and continuous regulation through the combination of integer and continuous variables, ensuring the economy and feasibility of the optimization results.

[0137] In this embodiment of the invention, the rolling optimization mechanism of the model predictive control framework combines the high-precision advantage of multimodal load forecasting with real-time dynamic updates of operating conditions, avoiding the problem of the traditional static optimization strategy being out of sync with the actual operating state. The design of pre-setting the future rolling time domain not only ensures the forward-looking nature of the optimization decision, but also responds quickly to the dynamic changes in renewable energy output and load through short-cycle rolling updates, effectively addressing their uncertainties. By executing only the first instruction of each instruction sequence and updating it on a rolling basis, the basic load control strategy is always based on the latest data and operating status, which improves the flexibility and adaptability of the strategy while meeting the multi-dimensional operating objectives of the microgrid cluster.

[0138] Optionally, generating a final control strategy for each user based on the basic load control strategy and the corresponding demand weight of the user includes:

[0139] The total cluster load adjustment amount in the basic load control strategy is taken as the total control task to be allocated, and the control task ratio of each user is determined according to the demand weight of each user.

[0140] Based on the control task ratio, the total control task is decomposed and allocated to each user, and a preliminary load control instruction is generated for each user.

[0141] By combining the specific electrical equipment constraints of each user with the electricity demand data, the feasibility of the initial load control command is verified and fine-tuned to obtain the final control strategy for each user.

[0142] Specifically, the total load adjustment of the cluster in the basic load control strategy is first defined as the total control task to be allocated. Using the user demand weights previously determined through the analytic hierarchy process (AHP) as the quantitative basis, the control task proportion for each user is clarified to ensure that the task allocation matches the user's importance and load characteristics. Then, the total control task is decomposed to each user according to this proportion, generating targeted preliminary load control instructions to achieve precise transmission of the cluster's overall goals to individual users. Finally, combining each user's specific equipment constraints (such as rated power and operating time limits) and power demand data (such as interruptible load parameters and temperature control ranges), the preliminary instructions are tested for feasibility. Instructions exceeding equipment operating capacity or violating the user's core power demand are eliminated and fine-tuned to ensure that the final control strategy meets both the overall cluster operating goals and adapts to the individual user's equipment conditions and power demand. In a preferred embodiment of the invention, the basic load control strategy of a microgrid cluster determines a certain 15 The total load adjustment (load reduction required) for the cluster during the minute-by-minute period is 100kW, which is designated as the total control task to be allocated. Given the demand weights of three users—User A (industrial user) 0.5, User B (commercial user) 0.3, and User C (residential user) 0.2—the control task ratio for each user is determined to be 5:3:2. After decomposition, User A is required to undertake a load reduction task of 50kW, User B 30kW, and User C 20kW, generating preliminary load control instructions. Further verification is then performed based on the constraints of each user's equipment and electricity demand data: User A's... The interruptible power limit for the interruptible load is 60kW, which meets the 50kW control requirement and requires no fine-tuning. User B's temperature control load range is 22℃-26℃. The initial instruction involving the adjustment of the air conditioning load would cause the temperature to exceed the range, so its control task is adjusted to 25kW, and the remaining 5kW is allocated to User A (which does not reach its interruptible limit). User C's interruptible load has a maximum continuous interruption time of 1 hour. The initial instruction meets the requirements, and the final control strategy is as follows: User A: reduce load by 55kW, User B: reduce load by 25kW, and User C: reduce load by 20kW.

[0143] In this embodiment of the invention, a proportional allocation mechanism based on demand weights ensures a fair and reasonable distribution of the overall control task among users, avoiding the problem of traditional allocation methods ignoring individual user differences. By combining user equipment constraints with feasibility verification and fine-tuning of electricity demand data, instructions that do not conform to actual operating conditions are effectively eliminated, improving the executability of the control strategy. The final user-level control strategy not only achieves the cluster's total load adjustment target but also fully adapts to each user's equipment capabilities and core electricity needs. While ensuring the power balance and operational stability of the microgrid cluster, it reduces the impact of control on users' normal electricity consumption, increases users' acceptance of the control strategy, and further enhances the accuracy and practicality of the entire control method.

[0144] Combination Figure 3 As shown, another embodiment of the present invention also provides an artificial intelligence-based user-side microgrid load control system, comprising:

[0145] The data acquisition unit is used to acquire multi-dimensional heterogeneous data of the microgrid cluster on the user side and the electricity demand data of each user in the microgrid cluster;

[0146] The user analysis unit is used to analyze the multi-dimensional heterogeneous data to obtain the load time series data of the microgrid cluster, and to determine the demand weight of each user in the microgrid cluster based on the user's electricity demand data using the analytic hierarchy process.

[0147] The data processing unit is used to process the load time series data through empirical mode decomposition to obtain the intrinsic mode function components of the multidimensional heterogeneous data.

[0148] The dynamic prediction unit is used to divide the intrinsic mode function components into high-frequency random components and low-frequency trend components, and then use the Kalman filter algorithm and the kernel extreme learning machine algorithm to dynamically predict the high-frequency random components and the low-frequency trend components respectively. The prediction results are weighted and fused through an attention mechanism to obtain the multimodal load prediction results.

[0149] The basic strategy generation unit is used to optimize the basic load control strategy of the microgrid cluster based on the multimodal load prediction results and the operation objectives of the microgrid cluster.

[0150] The user optimization unit is used to generate a final control strategy for each user based on the basic load control strategy and the corresponding demand weight of the user.

[0151] The control unit is used to control the user load in the microgrid cluster according to the final control strategy.

[0152] The user-side microgrid load control system based on artificial intelligence of the present invention has the same advantages over the prior art as the aforementioned user-side microgrid load control method based on artificial intelligence, and will not be repeated here.

[0153] Another embodiment of the present invention provides an electronic device, including: a processor and a memory, wherein the memory is used to store a computer program;

[0154] When the computer program is loaded by the processor, it causes the processor to execute the user-side microgrid load regulation method based on artificial intelligence as described above.

[0155] The electronic device of the present invention has the same advantages over the prior art as the aforementioned user-side microgrid load control method based on artificial intelligence, and will not be repeated here.

[0156] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A user-side microgrid load regulation method based on artificial intelligence, characterized in that, include: Acquire multi-dimensional heterogeneous data of the microgrid cluster on the user side and the electricity demand data of each user in the microgrid cluster; Based on the analysis of the multi-dimensional heterogeneous data, the load time series data of the microgrid cluster is obtained, and the demand weight of each user in the microgrid cluster is determined by the analytic hierarchy process based on the electricity demand data of the users. The load time series data is processed by empirical mode decomposition to obtain the intrinsic mode function components of the multidimensional heterogeneous data. The intrinsic mode function components are divided into high-frequency random components and low-frequency trend components. Then, the high-frequency random components and the low-frequency trend components are dynamically predicted by Kalman filtering algorithm and kernel extreme learning machine algorithm, respectively. The prediction results are weighted and fused by attention mechanism to obtain multimodal load prediction results. Based on the multimodal load forecast results and combined with the operational objectives of the microgrid cluster, the basic load control strategy of the microgrid cluster is generated. Based on the basic load control strategy and combined with the demand weights corresponding to the users, a final control strategy is generated for each user. The user load in the microgrid cluster is regulated according to the final regulation strategy.

2. The user-side microgrid load regulation method based on artificial intelligence according to claim 1, characterized in that, The acquisition of multi-dimensional heterogeneous data of the microgrid cluster on the user side and the electricity demand data of each user in the microgrid cluster includes: According to the preset sampling interval, the AC / DC hybrid load power data, distributed photovoltaic power output data, fuel cell power output data and environmental meteorological data of the microgrid cluster are collected synchronously. The AC / DC hybrid load power data, distributed photovoltaic power output data, fuel cell power output data and environmental meteorological data are preprocessed to obtain the multi-dimensional heterogeneous data. Obtain the interruptible power limit, maximum continuous interruption time, temperature-controlled load range, and demand price elasticity coefficient based on dynamic electricity price for each user's interruptible load; The interruptible power limit, the maximum continuous interruption time, the temperature-controlled load range, and the demand price elasticity coefficient are used as the electricity demand data.

3. The user-side microgrid load regulation method based on artificial intelligence according to claim 1, characterized in that, The process involves analyzing the multi-dimensional heterogeneous data to obtain the load time-series data of the microgrid cluster, and then using the analytic hierarchy process (AHP) to determine the demand weight of each user within the microgrid cluster based on the user's electricity demand data. This includes: A correlation analysis was performed on the load data, distributed power output data, and environmental meteorological data in the multi-dimensional heterogeneous data to obtain load dynamic characteristic indicators affected by environmental factors. Based on the mapping relationship between the load dynamic characteristic index and environmental factors, a load power sequence indexed by timestamp is constructed. The load power sequence is standardized to obtain the load time series data of the microgrid cluster; A judgment matrix is ​​established based on the user's electricity demand data using a pre-defined hierarchical analysis model; the hierarchical analysis model includes multiple criterion layers. The local weights of each user corresponding to each criterion layer are determined using the judgment matrix. The local weights are weighted and combined to obtain the user's demand weight in the microgrid cluster.

4. The user-side microgrid load regulation method based on artificial intelligence according to claim 1, characterized in that, The process of processing the load time-series data through empirical mode decomposition to obtain the intrinsic mode function components of the multi-dimensional heterogeneous data includes: Identify all local maxima and local minima in the load time series data; Spline interpolation is performed on the local maxima and local minima respectively to obtain the upper and lower envelopes of the load time series data; The mean envelope is obtained by taking the mean of the upper and lower envelopes at each data point. A data sequence is obtained by combining the load time series data with the mean envelope; The data sequence is separated by an iterative sieving process to obtain the intrinsic mode function components of the multidimensional heterogeneous data.

5. The user-side microgrid load regulation method based on artificial intelligence according to claim 1, characterized in that, The step of dividing the intrinsic mode function components into high-frequency random components and low-frequency trend components includes: Obtain the zero-crossing rate and energy percentage of each intrinsic mode function component; The intrinsic mode function components whose zero-crossing rate is greater than or equal to a first preset threshold and whose energy percentage is less than a second preset threshold are divided into high-frequency random components; The intrinsic mode function components whose zero-crossing rate is less than a third preset threshold and whose energy percentage is greater than or equal to a fourth preset threshold are classified as low-frequency trend components.

6. The user-side microgrid load regulation method based on artificial intelligence according to claim 1, characterized in that, The high-frequency random component and the low-frequency trend component are dynamically predicted using the Kalman filter algorithm and the kernel extreme learning machine algorithm, respectively. The prediction results are then weighted and fused using an attention mechanism to obtain multimodal load prediction results, including: For the high-frequency random components, a single-step prediction is performed using the extended Kalman filter algorithm to obtain the high-frequency component prediction sequence; For the low-frequency trend component, a multi-step prediction is performed using the kernel extreme learning machine algorithm to obtain the low-frequency component prediction sequence; The weights corresponding to the high-frequency component prediction sequence and the low-frequency component prediction sequence are dynamically calculated through an attention mechanism. Based on the weights corresponding to the high-frequency component prediction sequence and the low-frequency component prediction sequence, the high-frequency component prediction sequence and the low-frequency component prediction sequence are weighted and summed to obtain the multimodal load prediction result.

7. The user-side microgrid load regulation method based on artificial intelligence according to claim 1, characterized in that, The step of optimizing the basic load control strategy of the microgrid cluster based on the multimodal load forecast results and the operational objectives of the microgrid cluster includes: The multimodal load prediction results are input into the preset optimization model corresponding to the operation target. At the same time, the operation target of the microgrid cluster in the preset future rolling time domain is input into the preset optimization model. The model prediction control framework of the preset optimization model is used to solve the problem at the beginning of each optimization cycle in the preset future rolling time domain, so as to obtain the load regulation pre-instruction sequence for multiple time periods in the preset future rolling time domain. Based on the first load control preparatory instruction in the load control preparatory instruction sequence, the load control preparatory instruction sequence is continuously updated by repeatedly executing optimization solutions to obtain the basic load control strategy.

8. The user-side microgrid load regulation method based on artificial intelligence according to claim 1, characterized in that, The step of generating a final control strategy for each user based on the basic load control strategy and the corresponding demand weight of the user includes: The total cluster load adjustment amount in the basic load control strategy is taken as the total control task to be allocated, and the control task ratio of each user is determined according to the demand weight of each user. Based on the control task ratio, the total control task is decomposed and allocated to each user, and a preliminary load control instruction is generated for each user. By combining the specific electrical equipment constraints of each user with the electricity demand data, the feasibility of the initial load control command is verified and fine-tuned to obtain the final control strategy for each user.

9. A user-side microgrid load control system based on artificial intelligence, characterized in that, include: The data acquisition unit is used to acquire multi-dimensional heterogeneous data of the microgrid cluster on the user side and the electricity demand data of each user in the microgrid cluster; The user analysis unit is used to analyze the multi-dimensional heterogeneous data to obtain the load time series data of the microgrid cluster, and to determine the demand weight of each user in the microgrid cluster based on the user's electricity demand data using the analytic hierarchy process. The data processing unit is used to process the load time series data through empirical mode decomposition to obtain the intrinsic mode function components of the multidimensional heterogeneous data. The dynamic prediction unit is used to divide the intrinsic mode function components into high-frequency random components and low-frequency trend components, and then use the Kalman filter algorithm and the kernel extreme learning machine algorithm to dynamically predict the high-frequency random components and the low-frequency trend components respectively. The prediction results are weighted and fused through an attention mechanism to obtain the multimodal load prediction results. The basic strategy generation unit is used to optimize the basic load control strategy of the microgrid cluster based on the multimodal load prediction results and the operation objectives of the microgrid cluster. The user optimization unit is used to generate a final control strategy for each user based on the basic load control strategy and the corresponding demand weight of the user. The control unit is used to control the user load in the microgrid cluster according to the final control strategy.

10. An electronic device, characterized in that, include: Processor and memory, the memory being used to store computer programs; When the computer program is loaded by the processor, it causes the processor to execute the user-side microgrid load regulation method based on artificial intelligence as described in any one of claims 1-8.

Images