Virtual power plant energy supply and demand matching system based on AI

Abstract

The invention relates to an AI-based virtual power plant energy supply and demand matching system, and belongs to the technical field of virtual power plant energy scheduling. According to the system, multi-dimensional energy flow data is processed through a global energy flow data module, and a global energy flow data graph is generated; the energy load prediction module introduces an energy load correlation degree dynamic modeling mechanism to realize coupling prediction and result calibration of wind and light output and load demands; the multi-objective collaborative optimization module constructs a multi-objective optimization system fusing full life cycle values, and generates an energy flow scheduling instruction set adaptive to constraint conditions; and the edge intelligent control feedback module completes instruction transmission, equipment driving and data layered feedback based on a lightweight AI preprocessing mechanism. According to the system, through module collaboration and AI technology energizing, the accuracy and collaboration of energy supply and demand matching of the virtual power plant are improved, the scheduling requirements of distributed energy and multi-element loads are adapted, and safe and efficient operation of the virtual power plant is guaranteed.

Description

Technical Field

[0001] This invention belongs to the field of virtual power plant energy dispatching technology, specifically relating to an AI-based virtual power plant energy supply and demand matching system. Background Technology

[0002] As the global energy transition accelerates, distributed energy sources such as photovoltaics and wind power have achieved large-scale grid connection due to their clean and renewable advantages. However, their inherent intermittency, volatility, and randomness pose serious challenges to the safe and stable operation of the power grid and the efficient consumption of energy. Virtual power plants, as a new type of energy management platform that aggregates distributed power sources, flexible loads, energy storage units, and other controllable resources through advanced communication and dispatch control technologies, can effectively smooth out fluctuations in distributed energy output, provide grid ancillary services such as peak shaving and valley filling, and reserve response. They have become a core hub connecting distributed energy sources and the main power grid, playing an indispensable role in improving energy utilization efficiency and promoting market-oriented reforms in the electricity market.

[0003] As virtual power plants aggregate more resource types, expand their coverage, and become increasingly complex in their operation scenarios, the technical limitations of existing virtual power plant energy supply and demand matching systems are becoming increasingly apparent. They struggle to adapt to diversified and highly dynamic operational demands, primarily exhibiting the following technical bottlenecks: Firstly, the sources of multi-source energy flow data are dispersed and diverse, encompassing real-time output data from wind and solar power plants, user-side load monitoring data, energy storage unit operating status data, electricity market price data, and meteorological early warning data. The communication protocols and data formats of different data sources exhibit significant heterogeneity, and the data correlation analysis is insufficient. This leads to data processing being susceptible to interference from redundant and abnormal data, as well as issues such as misaligned data timing and inconsistent dimensions. It is difficult to form a unified, highly reliable data support covering all scenarios, thus affecting the accuracy of subsequent forecasting and scheduling decisions. Secondly, the energy load forecasting stage often uses single-dimensional forecasting models, independently forecasting wind and solar power output or load demand separately, without fully exploring the dynamic correlation between the two, and lacking adaptability to factors such as meteorological fluctuations and the adjustability characteristics of flexible loads. The shortcomings of the existing system lead to low accuracy in forecasting results, especially under extreme weather conditions or sudden load changes, where the forecasting deviation further widens, failing to provide a reliable basis for supply and demand matching. Existing optimized scheduling systems primarily focus on maximizing short-term operational benefits, neglecting to consider the lifecycle losses of energy storage units (including charge / discharge cycle losses, temperature losses, and aging losses). Furthermore, they lack sufficient consideration for the synergy of decision variables such as energy storage cluster charging / discharging strategies, controllable load control schemes, and grid-load interaction modes. This can easily lead to problems such as overcharging / discharging of energy storage units resulting in shortened lifespans and excessive deviations in grid dispatch command tracking, making it difficult to balance short-term gains with long-term operating costs. At the edge, as the terminal link for command execution and data acquisition, existing systems lack efficient, lightweight AI preprocessing mechanisms. Faced with massive amounts of operational data and command information from various devices, problems such as delayed command verification, redundant data transmission, and untimely response to abnormal information can easily occur, resulting in a disconnect between command execution and data feedback, making it impossible to achieve a closed-loop dynamic control throughout the entire "prediction-optimization-execution-feedback" process.

[0004] The aforementioned technical bottlenecks overlap, resulting in low energy supply and demand matching efficiency and poor operational stability of virtual power plants. This not only restricts the large-scale consumption of distributed energy and the release of the market operation value of virtual power plants, but may also have a potential impact on the safe and stable operation of the main power grid. An efficient and accurate energy supply and demand matching system is needed to solve these problems. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, this invention provides an AI-based virtual power plant energy supply and demand matching system. The objective of this invention can be achieved through the following technical solutions: An AI-based virtual power plant energy supply and demand matching system includes: a global energy flow data module, an energy load prediction module, a multi-objective collaborative optimization module, and an edge intelligent control feedback module; The full-domain energy flow data module acquires multi-dimensional energy flow data of the coverage area and related regions of the access virtual power plant, performs redundancy removal, anomaly correction and standardization processing, and generates a full-domain energy flow data map through data time sequence alignment and feature mining. The energy load prediction module introduces a dynamic modeling mechanism for energy load correlation. Based on the full-domain energy flow data map, it captures and predicts the spatiotemporal characteristics of distributed photovoltaic and wind power output. Simultaneously, it couples and predicts the temporal variation patterns and response characteristics of user-side rigid and flexible loads, calibrates the prediction results, and outputs wind and solar power output prediction curves and load demand prediction schemes. The multi-objective collaborative optimization module constructs a multi-objective optimization system that integrates the value of the entire life cycle. With the comprehensive operation benefits of the virtual power plant, the tracking accuracy of the grid dispatching instructions, and the life loss of the energy storage unit as the core objectives, the energy storage cluster charging and discharging strategy, the controllable load flexible regulation scheme, and the power interaction mode with the main grid and microgrid are used as key decision variables. Combined with the prediction results and the constraint parameters of the full-domain energy flow data map, an energy flow dispatching instruction set is generated. The edge intelligent control feedback module is based on a lightweight AI preprocessing mechanism at the edge. It receives the energy flow scheduling instruction set, transmits it to the edge management node, and drives each device to perform corresponding scheduling operations. It obtains the device instruction execution status, operating condition data and safety monitoring information, and then provides feedback in layers after analysis and processing at the edge.

[0006] Specifically, the process of acquiring multi-dimensional energy flow data of the coverage area and associated regions of the access virtual power plant includes: delineating the boundaries between the core coverage area and associated radiation area of ​​the virtual power plant by combining the power grid topology, wind and solar power station distribution and load density; determining the data acquisition nodes and corresponding equipment types that need to be accessed within the area; generating a list of data acquisition nodes and an equipment access ledger; establishing a data transmission link; marking the multi-dimensional energy flow data by source according to the acquisition area; and attaching acquisition node numbers and acquisition timestamps.

[0007] Specifically, the process of generating the full-domain energy flow data map includes: performing protocol conversion and format parsing for communication protocols and data formats of different data sources, and building a unified data access gateway; then entering the data processing stage, synchronously associating data source information during redundancy removal, filtering and removing data with unreliable sources and repetition exceeding the threshold according to preset duplication judgment rules, and using interpolation completion method to correct single-point abnormal data and trend fitting method to correct continuous abnormal data in the anomaly correction stage, and standardization and normalization processing to convert data with different dimensions and different numerical ranges into preset standard formats; associating equipment operating attributes and data characteristics, extracting data time-series correlation characteristics, mining cross-data source feature correlation relationships, and generating the full-domain energy flow data map.

[0008] Specifically, the process of the energy-load correlation dynamic modeling mechanism includes: initializing the basic parameters of the energy-load correlation dynamic model based on historical wind and solar power output data and historical load data in the global energy flow data map, and setting the constraint range for adjusting the model parameters; in the feature capture stage, weighted summing and coupling the extracted spatiotemporal features of distributed photovoltaic and wind power output with the energy-load correlation parameters, and filtering features according to the correlation threshold; during the prediction process, upon reaching the preset update cycle, combining real-time wind and solar power output data and load data, dynamically adjusting the parameters of the energy-load correlation dynamic model, and correcting the correlation weights and thresholds.

[0009] Specifically, the coupling prediction process includes: based on the global energy flow data map, establishing independent and joint feature libraries for the two types of loads respectively through the data collection dimensions of user-side rigid and flexible loads; mining the linkage response law between the two types of loads and wind and solar power output through the dynamic modeling mechanism of energy load correlation; calling the data from the independent and joint feature libraries, and dividing the flexible load into multiple prediction intervals according to the adjustable range based on the load's own time-series variation law and the correlation characteristics of wind and solar power output, and marking the probability distribution of each interval; for the peak and valley periods, duration, and load intensity of the rigid load prediction, linking the prediction results of the two types of loads with the prediction results of wind and solar power output to complete the coupling prediction process.

[0010] Specifically, the calibration prediction result process includes: establishing a dual verification system. The first verification is based on a historical data deviation database, retrieving prediction data and actual data under the same operating conditions within a preset period, calculating the average deviation rate as a verification benchmark, and comparing the preliminary prediction result with the benchmark deviation rate. The second verification is based on real-time data sampling values, collecting real-time data of wind and solar power output and load at preset sampling intervals, and comparing the instantaneous deviation between the real-time data and the preliminary prediction result.

[0011] Specifically, the process of constructing a multi-objective optimization system that integrates the value of the entire life cycle includes: first, weighting and summing the losses of each dimension of the energy storage unit throughout its life cycle and incorporating them into the multi-objective optimization system, and setting an upper limit for loss control; then, combining the entire process of virtual power plant operation, establishing a comprehensive benefit accounting model, and constructing the multi-objective optimization system through a hierarchical modeling approach.

[0012] Specifically, the multi-objective optimization system has an upper objective layer, which sets three core objectives: comprehensive operating efficiency of virtual power plants, accuracy of grid dispatch command tracking, and lifespan loss of energy storage units, and determines the priority constraints and weight adjustment range of the core objectives; a middle objective layer, which constructs the evaluation indicators and accounting standards corresponding to each core objective; and a lower objective layer, which breaks down the specific parameters and calculation logic corresponding to each objective.

[0013] Specifically, the key decision variables include: establishing correlation functions with the three core objectives for the energy storage cluster charging and discharging strategy, the controllable load flexible regulation scheme, and the power interaction mode with the main grid and microgrid, respectively; combining the equipment parameters in the full-domain energy flow data map, setting the adjustment step size and limit threshold of the key decision variables, establishing a variable coordination mechanism, and setting variable linkage adjustment rules.

[0014] Specifically, the process of generating the energy flow scheduling instruction set includes: conducting a risk assessment on the prediction results, extracting constraint parameters from the global energy flow data map, classifying them into three categories according to constraint type: safety constraints, equipment constraints, and scheduling constraints, and setting constraint priorities; determining the optimal variable combination through iterative calculation, and generating the energy flow scheduling instruction set.

[0015] Specifically, the edge-side lightweight AI preprocessing mechanism includes: loading a preset instruction compliance verification rule library, performing compliance verification on each item of the energy flow scheduling instruction set, and grouping the verified instructions into dual groups according to device affiliation and control priority to generate a grouped instruction list.

[0016] Specifically, the hierarchical feedback process includes: classifying the timeliness of the acquired instruction execution status data, operating condition data, and safety monitoring information; generating emergency feedback data packets or regular feedback data packets through the edge-end lightweight AI preprocessing mechanism; and using different feedback channels based on data type and timeliness requirements to synchronously update the full-domain energy flow data map.

[0017] The beneficial effects of this invention are as follows: By adapting to multi-source interfaces, performing hierarchical data processing, and mining features in the global energy flow data module, the heterogeneity and redundancy of multi-dimensional energy flow data are effectively addressed. The generated global energy flow data map provides unified data support, laying the foundation for subsequent prediction and optimization modules and avoiding scheduling deviations caused by data interference.

[0018] Based on the dynamic modeling mechanism of energy-load correlation, the deep coupling prediction of the spatiotemporal characteristics of wind and solar power output and load demand is realized. The prediction results are calibrated through a dual verification system to adapt to the fluctuation of wind and solar power output and load response characteristics, thereby improving the prediction accuracy, providing a reliable basis for optimized scheduling, and reducing the risk of supply and demand imbalance.

[0019] A multi-objective optimization system integrating the full life-cycle value of energy storage units is constructed, taking into account the comprehensive operational benefits of virtual power plants, the accuracy of grid dispatch command tracking, and energy storage life-cycle losses. Through a collaborative decision variable mechanism and priority constraint setting, a balance is achieved between short-term dispatch efficiency and long-term operational benefits, reducing energy storage losses and improving the adaptability of dispatch commands.

[0020] The lightweight AI preprocessing mechanism of the edge intelligent control feedback module can quickly complete instruction compliance verification and group transmission. Combined with the hierarchical data feedback strategy, it can realize the closed-loop control of the entire link of instruction execution, status monitoring, data feedback and strategy iteration, reduce the response lag on the edge side, and ensure the stability and flexibility of system operation.

[0021] The modules work together and have dynamic adjustment capabilities, which can adapt to virtual power plants of different sizes, diverse distributed energy sources and complex load scenarios. It is compatible with grid dispatch and electricity market demands, and provides technical support for the large-scale promotion of virtual power plants and the efficient consumption of distributed energy sources. It has broad application prospects. Attached Figure Description

[0022] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0023] Figure 1 This is a system architecture diagram of an AI-based virtual power plant energy supply and demand matching system according to the present invention; Figure 2 This is a schematic diagram of the working principle of the energy load prediction module in this invention; Figure 3 This is a hierarchical diagram of the multi-objective collaborative optimization system in this invention; Figure 4 This is a diagram showing the data flow and timing relationship in this invention. Detailed Implementation

[0024] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0025] Please see Figure 1-4An AI-based virtual power plant energy supply and demand matching system includes: a global energy flow data module, an energy load prediction module, a multi-objective collaborative optimization module, and an edge intelligent control feedback module; The full-domain energy flow data module acquires multi-dimensional energy flow data of the coverage area and related regions of the access virtual power plant, performs redundancy removal, anomaly correction and standardization processing, and generates a full-domain energy flow data map through data time sequence alignment and feature mining. The energy load prediction module introduces a dynamic modeling mechanism for energy load correlation. Based on the full-domain energy flow data map, it captures and predicts the spatiotemporal characteristics of distributed photovoltaic and wind power output. Simultaneously, it couples and predicts the temporal variation patterns and response characteristics of user-side rigid and flexible loads, calibrates the prediction results, and outputs wind and solar power output prediction curves and load demand prediction schemes. The multi-objective collaborative optimization module constructs a multi-objective optimization system that integrates the value of the entire life cycle. With the comprehensive operation benefits of the virtual power plant, the tracking accuracy of the grid dispatching instructions, and the life loss of the energy storage unit as the core objectives, the energy storage cluster charging and discharging strategy, the controllable load flexible regulation scheme, and the power interaction mode with the main grid and microgrid are used as key decision variables. Combined with the prediction results and the constraint parameters of the full-domain energy flow data map, an energy flow dispatching instruction set is generated. The edge intelligent control feedback module is based on a lightweight AI preprocessing mechanism at the edge. It receives the energy flow scheduling instruction set, transmits it to the edge management node, and drives each device to perform corresponding scheduling operations. It obtains the device instruction execution status, operating condition data and safety monitoring information, and then provides feedback in layers after analysis and processing at the edge.

[0026] Specifically, the process of acquiring multi-dimensional energy flow data of the coverage area and related regions of the virtual power plant includes: delineating the boundaries between the core coverage area and related radiation areas of the virtual power plant based on the power grid topology, wind and solar power station distribution, and load density; identifying the data acquisition nodes (including wind and solar power generation nodes, load monitoring nodes, energy storage operation and maintenance nodes, and electricity price collection nodes) and corresponding equipment types that need to be connected in each area; and compiling a list of data acquisition nodes and an equipment access ledger; establishing a data transmission link through wired and wireless dual-mode communication, with the wired link using fiber optic transmission and the wireless link using industrial-grade wireless communication technology, to access each area in real time. The multi-dimensional energy flow data of the nodes synchronously monitors the communication quality and connection status of the data access link at the millisecond level. When a link interruption or data packet loss rate exceeding a preset threshold is detected, the backup link switching mechanism is immediately activated to switch to the redundant link for data transmission. At the same time, a data supplementation mechanism is triggered to supplement the lost packet data within the preset time period to ensure data continuity. Finally, the accessed data is marked by the source according to the collection area. Core area data is marked with a core identifier, and related area data is marked with a related identifier. At the same time, the collection node number and collection timestamp are attached to provide dual identification basis of region dimension and node dimension for subsequent data processing and feature mining.

[0027] Specifically, the process of generating a full-domain energy flow data map includes: employing a multi-source interface adaptation mechanism to perform protocol conversion and format parsing for different data source communication protocols (including wired and wireless communication protocols) and data formats (including structured and semi-structured data), building a unified data access gateway to achieve seamless access and centralized aggregation of multi-dimensional energy flow data, and simultaneously performing real-time integrity verification on the accessed data to ensure the compatibility and integrity of the data access; then entering the data processing stage, during redundancy removal, synchronously linking data source information (including acquisition device number, acquisition timestamp, and data credibility rating), filtering and removing data with unreliable sources and repetition exceeding the threshold according to preset duplication judgment rules, retaining only non-duplicated core data with reliable sources; the anomaly correction stage combines equipment rating... Operating thresholds and historical data fluctuation ranges assist in identifying abnormal data. Interpolation completion is used to correct single-point abnormal data, and trend fitting is used to correct continuous abnormal data. Standardization and normalization processing uses a unified data conversion algorithm to convert data with different dimensions and numerical ranges into a preset standard format. After processing, a complete data quality report containing data pass rate, anomaly rate, and reliability is generated. Finally, in the feature mining stage, the equipment operating attributes (including equipment model, operating years, and maintenance records) are associated with data features. Time series analysis algorithms are used to extract data time series correlation features, and association rule algorithms are used to mine cross-data source feature correlation relationships. This generates a full-domain energy flow data map that includes both basic data and the correlation logic between equipment operating features and data, adapting to the precise calling requirements of subsequent modules.

[0028] Specifically, the dynamic modeling mechanism for energy-load correlation includes the following steps: Based on historical wind and solar power output data and historical load data in the global energy flow data map, the basic parameters of the dynamic energy-load correlation model are initialized, including the initial correlation threshold, parameter update cycle, and initial value of feature weights. At the same time, the constraint range for adjusting the model parameters is set to avoid parameter drift. In the feature capture stage, the extracted spatiotemporal features of distributed photovoltaic and wind power output (temporal features include time-period fluctuations and periodic changes, and spatial features include regional distribution and cluster linkage) are weighted and coupled with the energy-load correlation parameters. Features are filtered according to the correlation threshold, and core features of wind and solar power output with a correlation higher than the threshold with load demand are captured first, while low-correlation redundant features are weakened. During the prediction process, every time the preset update cycle is reached, the parameters of the energy-load correlation model are dynamically adjusted in combination with the real-time collected wind and solar power output data and load data. The correlation weights and thresholds are corrected so that the prediction results not only conform to the temporal and spatial characteristics of wind and solar power output itself, but also accurately adapt to the load demand correlation logic. At the same time, the basis and results of each parameter adjustment are recorded to form a model parameter update log.

[0029] Specifically, the coupling prediction process includes: based on a global energy flow data map, distinguishing the data acquisition dimensions of user-side rigid loads and flexible loads; acquiring fixed-dimensional data such as voltage, current, and continuous operating time for rigid loads, and additionally acquiring dynamic-dimensional data such as adjustable power range, response delay time, and switching status for flexible loads; establishing independent feature libraries (storing single load type features) and joint feature libraries (storing the correlation features between the two types of loads) for the two types of loads respectively, and storing feature data in time segments; through a dynamic modeling mechanism of energy load correlation, employing a correlation rule mining algorithm to mine the linkage response law between the two types of loads and wind and solar power output, and clarifying the flexible load correlation mechanism. The adjustable window period of the load (the period when the load rate is lower than the preset value and there is no mandatory operation requirement), the adjustable range and the response delay parameters are defined to clarify the fixed demand period, peak and valley occurrence patterns and non-adjustable constraints of rigid loads. During coupled prediction, data from independent feature libraries and joint feature libraries are called simultaneously, taking into account the time-series variation patterns of the load itself and the correlation characteristics of wind and solar power output. Flexible loads are divided into multiple prediction intervals according to the adjustable range, and the probability distribution of each interval is marked. For rigid loads, time-series prediction algorithms are used to accurately predict the peak and valley periods, duration and load intensity. The prediction results of the two types of loads are linked and calibrated with the wind and solar power output prediction results to complete the coupled prediction process.

[0030] Specifically, the calibration prediction process includes: establishing a dual verification system. The first verification is based on a historical data deviation database, retrieving predicted and actual data under similar operating conditions within a preset period, calculating the average deviation rate as a verification benchmark, and comparing the preliminary prediction result with the benchmark deviation rate. The second verification is based on real-time data sampling values, collecting real-time data of wind and solar power output and load at preset sampling intervals, and comparing the instantaneous deviation between the real-time data and the preliminary prediction result. After entering the calibration stage, for the wind and solar power output prediction curve, the deviation caused by meteorological factors is corrected by combining real-time meteorological data (wind speed, light intensity, cloud cover), and the prediction value is adjusted using the meteorological coefficient correction method. For the load demand prediction scheme, the deviation caused by the response delay of flexible load is corrected by shifting the prediction period according to the actual response delay time, while correcting the adjustable amplitude deviation. After calibration, the wind and solar power output prediction curve and the load demand prediction scheme are aligned on a minute-level time axis, and the supply and demand gap period, gap intensity, surplus period, and surplus amount of the two are marked. A prediction accuracy rating and deviation range description are added to provide accurate and traceable data support for subsequent optimization modules.

[0031] Specifically, the process of constructing a multi-objective optimization system that integrates the entire life cycle value includes: breaking down the life cycle loss of the energy storage unit into subdivided dimensions such as charge-discharge cycle loss, operating temperature loss, environmental aging loss, and maintenance loss; clarifying the loss calculation formula and influencing factors for each subdivided dimension; incorporating the weighted sum of the losses of each dimension into the multi-objective optimization system; and setting a loss control upper limit; combining the entire process of virtual power plant operation, clarifying the calculation items of the comprehensive operation benefits, such as power generation revenue, grid dispatch compensation revenue, operation and maintenance costs, loss costs, and default costs; and establishing a comprehensive benefit accounting model, in which the revenue item is calculated according to the real-time electricity market price and the standard agreed upon by the dispatch instructions, and the cost item is calculated according to the actual operation and maintenance expenditure and the quantitative value of loss.

[0032] Specifically, the multi-objective optimization system consists of three core objectives: the upper layer is the objective layer, which sets three core objectives: the comprehensive operational efficiency of the virtual power plant, the accuracy of grid dispatch command tracking, and the lifespan loss of energy storage units, and clarifies the priority constraints and weight adjustment range of each objective; the middle layer is the criterion layer, which refines the evaluation indicators and accounting standards corresponding to each core objective; and the lower layer is the indicator layer, which breaks down the specific parameters and calculation logic corresponding to each criterion, and finally constructs a multi-objective optimization system that takes into account both short-term dispatch efficiency and long-term operational benefits and covers multiple dimensions of objectives and constraints.

[0033] Specifically, the key decision variables include: establishing correlation functions with the three core objectives for the energy storage cluster charging and discharging strategy, the controllable load flexible regulation scheme, and the power interaction mode with the main grid and microgrid, respectively. The correlation function for the energy storage cluster charging and discharging strategy clarifies the impact coefficients of charging and discharging power and duration on energy storage losses, operational efficiency, and dispatch accuracy. The correlation function for the controllable load flexible regulation scheme clarifies the impact coefficients of load switching ratio and timing on efficiency and dispatch accuracy. The correlation function for the power interaction mode clarifies the impact coefficients of interaction power and time period on revenue and grid adaptability. The values ​​of each impact coefficient are determined through regression analysis algorithms. Combining the equipment parameters in the global energy flow data map, the adjustment step size and limit threshold of each decision variable are set. The energy storage charging and discharging strategy is implemented according to the energy storage equipment. The preset proportional adjustment step size of the rated power is set, and the limit threshold corresponds to the upper limit of the charging and discharging power and the upper limit of the depth of the equipment; the controllable load control scheme sets the adjustment step size according to the preset proportional of the total load, and the limit threshold corresponds to the adjustable range of the flexible load; the power interaction mode sets the adjustment step size according to the preset proportional of the line transmission capacity, and the limit threshold corresponds to the line transmission limit and the range of dispatching instructions, so as to avoid the variable adjustment exceeding the equipment operating capacity and external constraints; a variable coordination mechanism is established, and variable linkage adjustment rules are set. When the energy storage charging and discharging power changes, the power interaction mode with the main grid / microgrid is adjusted synchronously to ensure power balance; the controllable load control scheme dynamically adapts according to the energy storage operating status (charging / discharging, remaining capacity) to achieve multi-variable collaborative optimization and avoid system imbalance caused by the adjustment of a single variable.

[0034] Specifically, the process of generating the energy flow scheduling instruction set includes: conducting a risk assessment on the prediction results output by the energy load prediction module, setting two major risk assessment indicators—deviation rate and fluctuation amplitude—and classifying them into three risk levels: low, medium, and high. Conditions where the wind and solar power output prediction deviation exceeds a preset value or where load demand fluctuations are large are marked as high-risk, serving as an important auxiliary basis for optimization calculations; extracting constraint parameters from the full-domain energy flow data map, classifying them into three categories according to constraint type: safety constraints, equipment constraints, and scheduling constraints. Safety constraints include hard indicators such as energy storage equipment voltage and current thresholds and line transmission capacity limits; equipment constraints include equipment capability indicators such as energy storage charging and discharging power range and flexible load adjustability limits; and scheduling constraints include… External indicators such as the scope of power grid dispatch instructions and power market trading rules are included. Constraint priorities are set, with safety constraints having the highest priority, followed by equipment constraints, and then dispatch constraints. During optimization calculations, the feasible domain of decision variables is first defined based on high-priority constraints, eliminating variable combinations that exceed the constraint range. Then, the values ​​of decision variables are adjusted according to the risk level. Under high-risk conditions, the adjustment range of variables is reduced and redundancy is increased; under low-risk conditions, the variable values ​​are optimized to improve efficiency. The optimal variable combination is determined through multiple rounds of iterative calculations, generating an energy flow dispatch instruction set that meets various constraint requirements, can cope with predicted risks, and adapts to multi-objective needs. Each instruction clearly defines the execution object, execution parameters, execution time period, and fault tolerance range.

[0035] Specifically, the edge-side lightweight AI preprocessing mechanism includes: loading a pre-defined instruction compliance verification rule base, performing compliance verification on each item of the received energy flow scheduling instruction set, checking whether the instruction parameters conform to the rated operating capacity, adjustable range, and timing constraints of the corresponding device of the edge management node, marking instructions with excessive parameters or timing conflicts as non-compliant and feeding them back to the optimization module; and grouping the verified instructions by device affiliation (energy storage cluster, controllable load, distributed power source) and control priority, with instructions belonging to the same device and having the same priority. Instructions at the edge level are grouped together to generate a list of grouped instructions. Instructions are transmitted using a combination of local caching and real-time push. The edge first caches the grouped instructions in the local storage unit to avoid instruction loss, and then pushes the corresponding instructions to each edge management node in real time through the edge communication link. During the push process, an instruction verification code is attached. After receiving the code, the edge management node returns a verification code receipt to confirm that the instruction has not been tampered with. Simultaneously, the instruction reception time, verification result, group information, transmission status, and receipt status are uploaded to the system master node to establish a traceability ledger for the entire instruction transmission process, enabling querying and traceability of each link in the instruction transmission.

[0036] Specifically, the hierarchical feedback process includes: synchronously acquiring three types of data through edge sensing devices and communication modules: instruction execution status data, including information such as execution completion rate, execution deviation, and whether there is abnormal interruption; operating condition data, including parameters such as real-time power, temperature, and remaining capacity of the equipment; and safety monitoring information, including signals such as overload, over-temperature, and fault alarms. The acquired data is first classified according to its timeliness, with safety monitoring information and instruction execution status data, which have high real-time requirements, classified as Level 1 data, and operating condition statistical data, which have high latency tolerance, classified as Level 2 data. Level 1 data is given priority in the processing flow, while Level 2 data is processed in batches according to a preset cycle. Through a lightweight AI preprocessing mechanism at the edge, threshold judgment and trend analysis are used to preprocess the Level 1 data. The system employs analytical methods for rapid anomaly diagnosis, marking abnormal information such as equipment malfunctions, parameter exceedances, and excessive command execution deviations, attaching diagnostic conclusions and preliminary handling suggestions, and generating emergency feedback data packets. It also performs statistical analysis on secondary data, extracting information such as operating condition change trends, energy consumption statistics, and execution efficiency, generating regular feedback data packets. Different feedback channels are used to push data according to data type and timeliness requirements. Primary emergency feedback data packets are prioritized and pushed to the system master node and corresponding modules via high-speed communication links to ensure rapid feedback of core anomaly information. Secondary regular feedback data packets are pushed periodically via ordinary communication links, synchronously updating the overall energy flow data map. This hierarchical data feedback provides timely and accurate support for dynamic system adjustment and optimization strategy iteration.

[0037] This embodiment uses a park microgrid connected to a liquid-cooled energy storage system as an application scenario. The park microgrid covers a core area and related radiation areas. The core area aggregates distributed photovoltaic power stations, distributed wind power stations, several sets of liquid-cooled energy storage clusters (including liquid-cooled temperature control units, energy storage cell groups, and local monitoring modules), industrial flexible loads (such as adjustable loads on production lines), and residential rigid loads (such as household electricity loads). Related areas connect distributed small power sources and commercial loads (such as air conditioning loads in shopping malls). The system is equipped with a global energy flow data module, an energy load prediction module, a multi-objective collaborative optimization module, and an edge intelligent control feedback module. Each module communicates with edge management nodes via industrial Ethernet. Edge management nodes are deployed in the liquid-cooled energy storage clusters, load aggregation units, and wind and solar power stations, forming a three-layer architecture of "cloud-based coordination - edge execution - energy storage terminal linkage." The liquid-cooled energy storage system serves as the core control unit, undertaking energy buffering and peak-shaving tasks.

[0038] 1.1 Multi-dimensional energy flow data access Based on the microgrid topology of the park, the distribution density of wind and solar power plants, and the load intensity, the core area boundary is defined as A (the park's main body), and the associated radiation area boundary is defined as B (the surrounding collaborative power consumption area of ​​the park). Data acquisition nodes are determined to include wind and solar power output acquisition nodes, load monitoring nodes, liquid-cooled energy storage operation and maintenance nodes, electricity price acquisition nodes, and meteorological acquisition nodes. A node list and equipment access ledger are generated, and the data sampling frequency of each node is specified: the sampling frequency of wind and solar power output data is 1 time / minute, the sampling frequency of liquid-cooled energy storage system operating parameters (cell voltage, liquid cooling temperature, charging and discharging power, etc.) is 1 time / 10 seconds, the sampling frequency of rigid load data is 1 time / 5 minutes, the sampling frequency of flexible load data is 1 time / 2 minutes, and the sampling frequency of electricity price and meteorological data is 1 time / 15 minutes.

[0039] A wired + wireless dual-mode transmission link was established. The wired link uses fiber optic transmission (for critical data in the core area), while the wireless link uses industrial-grade LoRa wireless communication technology (for distributed nodes in the associated area). The multi-dimensional energy flow data is labeled with source identifiers according to the collection area. Data in the core area is labeled as A-node number-timestamp t, and data in the associated area is labeled as B-node number-timestamp t. The timestamp t is generated using a unified clock synchronization mechanism (synchronization accuracy ≤1ms) to ensure data timing consistency. For data from liquid-cooled energy storage systems, a separate identifier is used to distinguish between cell data and liquid-cooled temperature control data, providing accurate data classification for subsequent map generation and optimized scheduling.

[0040] 1.2 Data Processing and Generation of Global Energy Flow Data Maps The system adapts to different data sources' communication protocols and data formats. Data from the liquid-cooled energy storage system is accessed through its remote monitoring platform interface, which uses the IEC61850-90-7 protocol (a communication protocol applicable to energy storage systems). Data from wind and solar power plants uses the IEC61850 protocol, and load data uses the Modbus protocol. A protocol conversion module completes multi-protocol adaptation, and a format parsing module converts heterogeneous data into a unified JSON intermediate format. A unified data access gateway is built to achieve data aggregation. The gateway supports data caching and breakpoint resume functions to prevent the loss of critical data from the liquid-cooled energy storage system.

[0041] The data processing step is performed according to the following procedure: (1) Redundancy removal: Associate the data source information (collection device number, credibility rating), set the repetition threshold α (α is 3 times), when the same data is collected repeatedly more than α times within the time window Δt (Δt=1 minute), and the credibility rating is lower than the preset level λ (λ=B level), the low credibility repeated data is removed and the high credibility data is retained; for the high frequency sampling data of the liquid-cooled energy storage system, additional redundancy judgment conditions are set. If the fluctuation of the three consecutive sampling data is ≤±0.5%, the first data is retained and the subsequent redundant data is removed.

[0042] (2) Anomaly correction: For single-point abnormal data, the interpolation completion method is used. A linear interpolation function y=kx+b (k is the slope and b is the intercept) is constructed based on the normal data points adjacent to the abnormal data, and the completed value is calculated. For continuous abnormal data (the number of consecutive abnormalities is ≥5 times), the trend fitting method is used. A quadratic trend function f(x)=ax is constructed based on the historical data of the same period and similar working conditions. 2 +bx+c (x is the time variable, a, b, and c are fitting coefficients) to generate corrected data; for the liquid cooling temperature data of the liquid cooling energy storage system, if anomalies occur, additional cell temperature data is used for auxiliary correction to ensure the rationality of the liquid cooling system operating parameters.

[0043] (3) Standardization and normalization: A linear normalization algorithm is used to map data with different dimensions and different numerical ranges to the interval [0,1]. The calculation formula is x'=(x-x_min) / (x_max-x_min), where x is the original data, x_max and x_min are the historical maximum and minimum values ​​of this type of data, respectively, and x' is the standardized data. For the charging and discharging power data of the liquid-cooled energy storage system, the interval [0,1] is used separately to ensure that the dimensions are consistent with other energy data.

[0044] (4) Map generation: Associate the operating attributes of the associated equipment (liquid-cooled energy storage system model, cell cycle count, liquid-cooled system maintenance records, wind and solar equipment operating years, load equipment type) with the processed data features, extract the time-series correlation features of the data through the time-series analysis algorithm (ARIMA algorithm), and use the association rule mining algorithm (Apriori algorithm) to mine the cross-data source feature correlation relationship (such as the correlation relationship between liquid-cooled energy storage charging and discharging power and photovoltaic output, liquid-cooled temperature), and generate a full-domain energy flow data map containing basic data information, correlation relationship and feature label. The map is updated according to the time granularity Δt (Δt=5 minutes) and pushed synchronously to the liquid-cooled energy storage remote monitoring platform.

[0045] 2.1 Modeling and Training of the Dynamic Model of Energy-Load Correlation (1) Data preparation: Historical wind and solar power output data (PV output P_pv, wind power output P_w), historical load data (rigid load P_lr, flexible load P_lf), and historical operation data of liquid-cooled energy storage system (charge and discharge power P_e, cell temperature T_b, liquid cooling temperature T_c) were extracted from the global energy flow data map. The time span was T (T=1 year). The data were divided into training set (η=70%), validation set (ζ=20%), and test set (1-η-ζ=10%), where η+ζ<1. The data were preprocessed to remove abnormal data caused by extreme weather (rainstorms, typhoons) to ensure the validity of the training data.

[0046] (2) Model initialization: An LSTM neural network was selected to construct a dynamic model of energy load correlation. Based on the training set data, the basic parameters of the model were initialized, including the initial correlation threshold γ (γ=0.6), the model weight matrix W (input layer-hidden layer weight matrix W1, hidden layer-output layer weight matrix W2), and the bias term b (hidden layer bias b1, output layer bias b2). The constraint range for adjusting the model parameters was set as [W_min, W_max] (W_min=-0.5, W_max=0.5) and [b_min, b_max] (b_min=-0.1, b_max=0.1) to avoid parameter drift. The model input dimension was 8 (P_pv, P_w, P_lr, P_lf, P_e, T_b, T_c, time features), and the output dimension was 2 (predicted wind and solar power output, ...). (3) Feature extraction and coupling: Multidimensional features are extracted from the data in the training set. Spatiotemporal features (temporal features: time period fluctuation features, daily / weekly cycle change features; spatial features: regional distribution features, cluster linkage features) are extracted from the wind and solar power output data. Operational status features (charging and discharging status, temperature change rate) are extracted from the liquid-cooled energy storage data. Demand features (time period demand intensity, response delay features) are extracted from the load data. The feature vector X=[P_pv,X_pvt,X_pvs,P_w,X_wt,X_ws,P_lr,P_lf,X_lr,X_lf,P_e,T_b,T_c] is constructed. The feature vector X is weighted and coupled with the energy load correlation parameter. The calculation formula is X'=W1·X+b1. Features are selected according to the correlation threshold γ. Core features X'≥γ are retained (such as photovoltaic power output time period features, liquid-cooled energy storage charging and discharging power, rigid load demand intensity), and low correlation redundant features are weakened.

[0047] (4) Model training and iteration: With the goal of minimizing prediction bias, construct the loss function L=∑(y̅-y) 2 / N+λ·∑||W|| 2(Where y is the actual value, y̅ is the predicted value, N is the number of samples, and λ is the regularization coefficient, λ=0.001, used to avoid overfitting); The adaptive moment estimation (Adam) algorithm is used to update the model parameters. The learning rate μ is initially set to 0.001 and dynamically decays according to the number of iterations (decaying to 0.9 of the original value every 100 iterations). The iteration formulas are W=W-μ·∇L(W,b) and b=b-μ·∇L(W,b); After every iteration number k (k=50), the empirical... The validation set is used to verify the model's performance. The mean absolute error (MAE) of the validation set is calculated as: MAE = ∑|y̅_v - y_v| / N_v (where y̅_v is the predicted value of the validation set, y_v is the actual value of the validation set, and N_v is the number of samples in the validation set). If the MAE is lower than the threshold ε (ε = 0.05), the iteration is stopped; otherwise, the parameters are adjusted until the model converges. During the training process, test data from a liquid-cooled energy storage system are used periodically to verify the model's ability to fit the energy storage-related features, ensuring that the model is adapted to the energy storage regulation requirements.

[0048] (5) Model deployment and update: Deploy the trained and converged model to the energy load prediction module, set the parameter update cycle T1 (T1=7 days), and at each cycle T1, combine real-time wind and solar power output data, load data and real-time operation data of liquid-cooled energy storage system (P_e, T_b, T_c), dynamically adjust the correlation weight and threshold γ through the sliding window method to ensure that the model adapts to changes in operating conditions; at the same time, push the model update log to the liquid-cooled energy storage remote monitoring platform to realize the traceability of model status.

[0049] 2.2 Coupling Prediction and Result Calibration Calculation (1) Coupled prediction: Based on the global energy flow data map, establish an independent feature library for rigid loads (storing features such as voltage, current, and continuous operating time), an independent feature library for flexible loads (storing features such as adjustable range, response delay time, and switching status), and a joint feature library (storing the correlation features between the two types of loads), and a liquid-cooled energy storage-load correlation feature subset (storing the linkage data between energy storage charging and discharging status and load demand); through the dynamic model of energy load correlation, explore the linkage response law between the two types of loads and wind and solar power output and liquid-cooled energy storage operation status, clarify the adjustable window period and adjustable range of flexible loads and the fixed demand period of rigid loads, and determine the influence coefficient of liquid-cooled energy storage charging and discharging status on load prediction.

[0050] (2) Load forecasting calculation: The feature library data is called to divide the flexible load into multiple forecast intervals (3 levels: low load interval, medium load interval, and high load interval) according to the adjustable range. The probability distribution parameters θ (θ1, θ2, θ3, satisfying θ1+θ2+θ3=1) of each interval are marked. The predicted value is calculated by weighting the interval as P̅_lf=∑θ_i·P_lf_i (i=1,2,3, P_lf_i is the load value of each interval). For rigid load, the time series forecasting algorithm is used based on historical time series changes. The peak and trough periods, durations, and load intensity P̅_lr are predicted regularly. The two types of load prediction results are linked with the wind and solar power output prediction results P̅_pv and P̅_w, and the liquid-cooled energy storage predicted charge and discharge power P̅_e for calibration. The calibration formula is P̅_l=P̅_lr+P̅_lf+δ·(P̅_pv+P̅_w-P̅_lr-P̅_lf-P̅_e), where δ is the deviation correction coefficient (δ=0.8-1.2, dynamically adjusted according to the remaining capacity of liquid-cooled energy storage).

[0051] (3) Dual verification calibration: The first verification is based on the historical data deviation database. It retrieves the predicted data and actual data under the same working conditions within the preset period (30 days) and calculates the average deviation rate μ=∑|y̅_h-y_h| / y_h (y̅_h is the historical predicted value and y_h is the historical actual value). It compares the preliminary prediction result with μ. If the deviation rate exceeds μ, the correction coefficient δ is adjusted. The second verification is based on the real-time data sampling value. Real-time data is collected at the sampling interval Δt1 (Δt1=1 minute) and the instantaneous deviation Δ=|y̅-y_r| (y_r is the real-time data) is calculated. If Δ exceeds the threshold Δ_max (Δ_max=10%), it triggers immediate calibration and corrects the prediction result. For liquid-cooled energy storage data, a temperature deviation verification is added separately to ensure that the prediction parameters of the liquid-cooled system are reasonable.

[0052] 3.1 Construction and Modeling of Multi-Objective Optimization System The process of constructing a multi-objective optimization system that integrates the entire lifecycle value is as follows: (1) Target layer modeling: Three core objectives are set, and specific objective functions are constructed in combination with the characteristics of liquid-cooled energy storage systems: Virtual power plant overall operation benefit target F1: F1=α1·R_pow+α2·R_disp-α3·C_op-α4·C_loss-α5·C_cool Where R_pow is the revenue from wind, solar and energy storage power generation (R_pow=P_pv·p_pv+P_w·p_w+P_e·p_e, where p_pv, p_w and p_e are the grid connection tariffs for photovoltaic, wind power and energy storage respectively), R_disp is the revenue from grid dispatch compensation (R_disp=ΔP·p_disp, where ΔP is the dispatch response power and p_disp is the dispatch compensation tariff), C_op is the operation and maintenance cost (including operation and maintenance costs for wind, solar, load and energy storage), C_loss is the energy storage loss cost (C_loss=k_l·P_e·t, where k_l is the loss coefficient and t is the operating time), C_cool is the liquid cooling energy consumption cost of the liquid-cooled energy storage system (C_cool=P_cool·p_e·t, where P_cool is the power of the liquid cooling system), and α1, α2, α3, α4 and α5 are weighting coefficients (α1+α2+α3+α4+α5=1). (2) Energy storage unit lifetime loss target F2: F2=β1·∑(N_i / N_max_i)+β2·∑(ΔT_b·k_T) Where N_i is the number of charge-discharge cycles of the liquid-cooled energy storage cell, N_max_i is the rated cycle life of the cell, ΔT_b is the temperature fluctuation value of the cell, k_T is the temperature loss coefficient, β1 and β2 are the loss weights (β1+β2=1), and the objective is to minimize F2. (3) Target F3 for tracking accuracy of power grid dispatch instructions: F3 = 1 - |P_out - P_inst| / P_inst, where P_out is the actual output of the virtual power plant (including the combined output after wind, solar, energy storage and load regulation), and P_inst is the output of the power grid dispatch instructions. The target is to maximize F3 (i.e. minimize the tracking deviation). (4) Modeling of criteria layer and indicator layer: The criteria layer constructs the evaluation indicators corresponding to each target. The evaluation indicators for benefit targets include the rate of return achievement (actual return / expected return), cost control rate (actual cost / budgeted cost), and liquid cooling energy consumption rate (P_cool / P_e); the evaluation indicators for loss targets include the cell cycle loss rate (N_i / N_max_i) and the cell temperature loss rate (ΔT_b / T_b_max, where T_b_max is the rated maximum temperature of the cell); the evaluation indicators for accuracy targets include the deviation rate (|P_out-P_inst| / P_inst) and response delay (actual response time / rated response time); the indicator layer decomposes the specific parameters corresponding to each criterion and clarifies the calculation logic.

[0053] (5) Weight allocation: The weights ω1, ω2 and ω3 of the three core objectives are determined by the Analytic Hierarchy Process (AHP) to satisfy ω1+ω2+ω3=1, and the weight adjustment range [ω_min,ω_max] is set (ω_min=0.2, ω_max=0.5). For the operation scenario of liquid-cooled energy storage system, when the number of cell cycles is close to the rated value, ω2 (energy storage life loss weight) is dynamically increased to ensure the long-term operation of energy storage system.

[0054] 3.2 Key Decision Variables and Optimization Calculation (1) Setting of key decision variables: The charging and discharging strategy for liquid-cooled energy storage clusters is as follows: The decision variable is the charging and discharging power P_e (charging is negative and discharging is positive). The adjustment step size ΔP_e is set (ΔP_e=0.01P_enom, where P_enom is the rated power of energy storage). The constraint range is [P_e_min, P_e_max] (P_e_min=-P_enom, P_e_max=P_enom). The correlation function P_e=f(F1,F2) is established to characterize the impact of charging and discharging power on efficiency and cell lifespan. At the same time, the liquid cooling system power P_cool is correlated to ensure that P_cool is dynamically adjusted with P_e.

[0055] (2) Controllable load flexible control scheme: The decision variable is the load switching ratio k (0≤k≤1), the adjustment step size Δk (Δk=0.05), the constraint range [0,k_max] (k_max=0.3, i.e. the maximum adjustable load ratio is 30%), and the correlation function k=f(F1,F3) to ensure that the load control adapts to the scheduling instructions and benefit objectives.

[0056] (3) Power interaction mode: The decision variable is the power P_n interacting with the main network / microgrid, the adjustment step size ΔP_n (ΔP_n=0.01P_n_max, P_n_max is the maximum power interaction), the constraint range [P_n_min,P_n_max], the correlation function P_n=f(F1,F3), and the balance between power interaction, scheduling accuracy, and revenue target.

[0057] (4) Constraint setting: Constraint parameters are extracted from the global energy flow data map and divided into three categories based on the characteristics of the liquid-cooled energy storage system. Constraint priorities are set (safety constraints > equipment constraints > scheduling constraints): Safety constraints: Liquid-cooled energy storage cell voltage ∈ [U_b_min, U_b_max], cell temperature ∈ [T_b_min, T_b_max], liquid cooling system temperature ∈ [T_c_min, T_c_max], line transmission power ≤ P_lim; Equipment constraints: Liquid-cooled energy storage charging and discharging power ≤ P_enom, liquid cooling system power ≤ P_cool_enom, flexible load adjustable range ∈ [P_l_min, P_l_max], wind and solar power output ≤ rated output; Scheduling constraints: power interaction with the main grid / microgrid ≤ scheduling command range, response delay ≤ t_max (t_max = 30 seconds), charge / discharge switching delay ≤ t_switch (t_switch = 5 seconds, liquid-cooled energy storage specific constraint); (5) Optimization calculation: Genetic algorithm (GA) is used to solve the problem online. The population size is initialized to M (M=50), the number of iterations is G (G=100), and the fitness function is F=ω1·F1+ω2·(1-F2)+ω3·F3 (the minimum objective F2 is transformed into the maximum objective 1-F2). In each iteration, feasible solutions are selected according to the constraints (prioritizing the liquid-cooled energy storage temperature and voltage constraints). Selection is carried out through roulette wheel selection, single-point crossover (crossover probability p_c=0.8), and mutation (mutation probability p_m=0). .05) Update the population and calculate the fitness of each individual. When the number of iterations reaches G or the fitness function value fluctuates for 10 consecutive generations by ≤±0.001 (converging to the threshold F_m), stop the iteration, output the optimal combination of decision variables (P_e*, k*, P_n*), generate the energy flow scheduling instruction set, specify the charging and discharging power of the liquid-cooled energy storage system, the liquid-cooled temperature control value, and the control parameters of load and wind and solar equipment. The instruction format is compatible with the IEC61850-90-7 protocol and can be directly sent to the liquid-cooled energy storage remote monitoring platform.

[0058] 4.1 Lightweight AI Preprocessing and Command Transmission at the Edge The implementation process of the lightweight AI preprocessing mechanism at the edge is as follows: (1) Command verification: Load the preset command compliance verification rule base. The rule base includes the rated parameters of the liquid-cooled energy storage system (voltage, power, temperature range), load control authority, wind and solar power output limits, etc. Verify each item of the received energy flow scheduling command set, focusing on whether the liquid-cooled energy storage command parameters (P_e, T_c) are within the rated range [P_e_min, P_e_max] and [T_c_min, T_c_max], and whether the control timing is conflict-free (such as whether the charge and discharge switching time meets the t_switch constraint). Mark non-compliant commands as abnormal types (parameter over-limit, timing conflict), feed them back to the multi-objective collaborative optimization module, and push them to the liquid-cooled energy storage remote monitoring platform for alarm.

[0059] (2) Command grouping and transmission: For commands that pass verification, they are grouped according to equipment affiliation (liquid-cooled energy storage cluster, controllable load, wind and solar power station) and control priority (emergency control: such as handling abnormal energy storage temperature; routine control: such as daily charging and discharging) to generate a group command list; local caching + real-time push method is used for transmission. The edge end caches the command to the local storage unit (caching time ≥ 24 hours) and pushes the command to the liquid-cooled energy storage remote monitoring platform through industrial Ethernet. The interface protocol adopts IEC61850-90-7 and an additional command verification code (MD5 encryption) is attached. After receiving the command, the monitoring platform returns the verification code receipt to confirm that the command has not been tampered with; the entire process information of command transmission (issuance time, reception time, verification result) is uploaded to the system master node at the same time to establish a traceability ledger and realize the full traceability of liquid-cooled energy storage commands.

[0060] 4.2 Device-driven and hierarchical feedback calculation The device driving and hierarchical feedback process is as follows: (1) Device drive: After receiving the instruction, the edge management node drives the device to perform the operation according to the differentiated control logic. The liquid-cooled energy storage system adjusts the charging and discharging state according to the instruction power P_e*, and the liquid-cooled temperature control unit adjusts the cooling power according to the instruction temperature T_c*, matching the changes in cell temperature in real time; the flexible load starts or stops or adjusts the power according to the switching ratio k*, and the wind and solar power station adjusts the output according to the instruction; during the execution, the high-frequency acquisition of the device operating parameters (liquid-cooled energy storage is acquired once every 10 seconds, and other devices are acquired once every 1 minute) is used. If the temperature of the liquid-cooled energy storage cell exceeds the threshold T_b_max, a pause instruction is immediately triggered to start the liquid-cooled emergency cooling. The execution is restarted after the parameters return to normal.

[0061] (2) Data acquisition and processing: Real-time acquisition of equipment command execution status data (completion rate, deviation), operating condition data (wind and solar power output, load power, liquid-cooled energy storage P_e, T_b, T_c, remaining capacity SOC, etc.) and safety monitoring information (overload, over-temperature, voltage abnormality alarms), classified according to timeliness: Level 1 data (safety alarms, energy storage temperature abnormalities, execution abnormalities) has high real-time requirements and a sampling frequency of 10 seconds / time; Level 2 data (operating condition statistics, revenue accounting data) has high latency tolerance and a sampling frequency of 5 minutes / time; a separate encrypted transmission channel is established for liquid-cooled energy storage data to ensure data security.

[0062] (3) Layered feedback: Through the lightweight AI preprocessing mechanism at the edge, the first-level data is quickly diagnosed as abnormal (diagnosis delay ≤ 1 second), and an emergency feedback data packet is generated with additional diagnostic conclusions (such as the cause of energy storage overheating and disposal suggestions). The data is fed back to the multi-objective collaborative optimization module and the liquid-cooled energy storage remote monitoring platform via a high-speed link (transmission rate ≥ 100 Mbps). The second-level data is batch statistically analyzed to generate regular feedback data packets, which are pushed at 5-minute intervals and updated synchronously to the full-domain energy flow data map. The feedback data format is adapted to the IEC61850-90-7 protocol to ensure data interaction compatibility with the liquid-cooled energy storage monitoring platform and realize closed-loop control of "instruction issuance-execution-feedback-optimization".

[0063] The system operates collaboratively across the entire chain of "data acquisition-processing-prediction-optimization-execution-feedback," adapting to microgrid scenarios in industrial parks where liquid-cooled energy storage systems are integrated: The full-domain energy flow data module collects multi-dimensional data at a set frequency, processes it to generate a full-domain energy flow data map containing energy storage characteristics, and simultaneously pushes it to the energy storage monitoring platform; the energy load prediction module, based on a trained and converged LSTM model, combines energy storage data (P_e, T_b, T_c) to complete coupled prediction and dual-verification calibration; the multi-objective collaborative optimization module solves the problem using a genetic algorithm, generating scheduling instructions (including parameters such as P_e* and T_c*) adapted to the lifespan and efficiency of energy storage, and issues them according to the IEC61850-90-7 protocol; the edge intelligent control feedback module drives equipment execution, frequently collects energy storage operation data, and provides hierarchical feedback to the front-end modules for iterative parameter optimization. The entire process ensures that the liquid-cooled energy storage system operates collaboratively with wind, solar, and loads, achieving precise matching of energy supply and demand in the industrial park microgrid, balancing operational efficiency and long-term energy storage lifespan.

[0064] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An AI-based virtual power plant energy supply and demand matching system, characterized in that, include: The module includes a global energy flow data module, an energy load prediction module, a multi-objective collaborative optimization module, and an edge intelligent control feedback module. The full-domain energy flow data module acquires multi-dimensional energy flow data of the coverage area and related regions of the access virtual power plant, performs redundancy removal, anomaly correction and standardization processing, and generates a full-domain energy flow data map through data time sequence alignment and feature mining. The energy load prediction module introduces a dynamic modeling mechanism for energy load correlation. Based on the full-domain energy flow data map, it captures and predicts the spatiotemporal characteristics of distributed photovoltaic and wind power output. Simultaneously, it couples and predicts the temporal variation patterns and response characteristics of user-side rigid and flexible loads, calibrates the prediction results, and outputs wind and solar power output prediction curves and load demand prediction schemes. The multi-objective collaborative optimization module constructs a multi-objective optimization system that integrates the value of the entire life cycle. With the comprehensive operation benefits of the virtual power plant, the tracking accuracy of the grid dispatching instructions, and the life loss of the energy storage unit as the core objectives, the energy storage cluster charging and discharging strategy, the controllable load flexible regulation scheme, and the power interaction mode with the main grid and microgrid are used as key decision variables. Combined with the prediction results and the constraint parameters of the full-domain energy flow data map, an energy flow dispatching instruction set is generated. The edge intelligent control feedback module is based on a lightweight AI preprocessing mechanism at the edge. It receives the energy flow scheduling instruction set, transmits it to the edge management node, and drives each device to perform corresponding scheduling operations. It obtains the device instruction execution status, operating condition data and safety monitoring information, and then provides feedback in layers after analysis and processing at the edge.

2. The system according to claim 1, characterized in that, The specific process of acquiring multi-dimensional energy flow data of the coverage area and related regions of the virtual power plant includes: combining the power grid topology, wind and solar power station distribution and load density, delineating the boundaries between the core coverage area and the related radiation area of ​​the virtual power plant, determining the data acquisition nodes and corresponding equipment types that need to be connected within the area, generating a list of data acquisition nodes and an equipment access ledger; establishing a data transmission link, marking the multi-dimensional energy flow data by the acquisition area, and adding acquisition node numbers and acquisition timestamps.

3. The system according to claim 1, characterized in that, The specific process of generating the full-domain energy flow data map includes: performing protocol conversion and format parsing for communication protocols and data formats of different data sources, and building a unified data access gateway; then entering the data processing stage, synchronously associating data source information during redundancy removal, filtering and removing data with unreliable sources and repetition exceeding the threshold according to preset duplication judgment rules, and using interpolation completion method to correct single-point abnormal data and trend fitting method to correct continuous abnormal data in the anomaly correction stage, and standardization and normalization processing to convert data with different dimensions and different numerical ranges into preset standard formats; associating equipment operating attributes and data characteristics, extracting data time-series correlation characteristics, mining cross-data source feature correlation relationships, and generating the full-domain energy flow data map.

4. The system according to claim 1, characterized in that, The specific process of the energy-load correlation dynamic modeling mechanism includes: initializing the basic parameters of the energy-load correlation dynamic model based on historical wind and solar power output data and historical load data in the global energy flow data map, and setting the constraint range for adjusting the model parameters; in the feature capture stage, the extracted spatiotemporal features of distributed photovoltaic and wind power output are weighted and coupled with the energy-load correlation parameters, and features are selected according to the correlation threshold; during the prediction process, when the preset update cycle is reached, the parameters of the energy-load correlation dynamic model are dynamically adjusted in combination with real-time wind and solar power output data and load data, and the correlation weights and thresholds are corrected.

5. The system according to claim 1, characterized in that, The specific process of the coupled prediction includes: based on the global energy flow data map, establishing independent and joint feature libraries for the two types of loads through the data collection dimensions of user-side rigid and flexible loads; mining the linkage response law between the two types of loads and wind and solar power output through the dynamic modeling mechanism of energy load correlation; calling the data from the independent and joint feature libraries, and dividing the flexible load into multiple prediction intervals according to the adjustable range based on the load's own time-series variation law and the correlation characteristics of wind and solar power output, and marking the probability distribution of each interval; for the peak and valley periods, duration, and load intensity of the rigid load prediction, linking the prediction results of the two types of loads with the prediction results of wind and solar power output to complete the coupled prediction process.

6. The system according to claim 1, characterized in that, The specific process of the calibration prediction results includes: establishing a dual verification system. The first verification is based on the historical data deviation database, which retrieves the predicted data and actual data under the same working conditions within a preset period, calculates the average deviation rate as the verification benchmark, and compares the preliminary prediction results with the benchmark deviation rate. The second verification is based on real-time data sampling values, which collects real-time data of wind and solar power output and load at preset sampling intervals, and compares the instantaneous deviation between the real-time data and the preliminary prediction results.

7. The system according to claim 1, characterized in that, The specific process of constructing a multi-objective optimization system that integrates the value of the entire life cycle includes: incorporating the weighted sum of the losses of each dimension of the energy storage unit throughout its life cycle into the multi-objective optimization system and setting an upper limit for loss control; establishing a comprehensive benefit accounting model in conjunction with the entire operation process of a virtual power plant, and constructing the multi-objective optimization system through a hierarchical modeling approach.

8. The system according to claim 7, characterized in that, The upper layer of the multi-objective optimization system is the objective layer, which sets three core objectives: the comprehensive operating efficiency of the virtual power plant, the tracking accuracy of the power grid dispatch command, and the life loss of the energy storage unit, and determines the priority constraints and weight adjustment range of the core objectives; the middle layer is the criterion layer, which constructs the evaluation indicators and accounting standards corresponding to each core objective. The lower layer is the indicator layer, which breaks down the specific parameters and calculation logic corresponding to each criterion.

9. The system according to claim 1, characterized in that, The key decision variables specifically include: establishing correlation functions with the three core objectives for the energy storage cluster charging and discharging strategy, the controllable load flexible regulation scheme, and the power interaction mode with the main grid and microgrid, respectively; combining the equipment parameters in the full-domain energy flow data map, setting the adjustment step size and limit threshold of the key decision variables, establishing a variable coordination mechanism, and setting variable linkage adjustment rules.

10. The system according to claim 1, characterized in that, The specific process of generating the energy flow scheduling instruction set includes: conducting a risk assessment on the prediction results, extracting constraint parameters from the global energy flow data map, classifying them into three categories according to constraint type: safety constraints, equipment constraints, and scheduling constraints, and setting constraint priorities; determining the optimal variable combination through iterative calculation, and generating the energy flow scheduling instruction set.

11. The system according to claim 1, characterized in that, The lightweight AI preprocessing mechanism at the edge specifically includes: loading a preset instruction compliance verification rule library, performing compliance verification on each item of the energy flow scheduling instruction set, and grouping the verified instructions into dual groups according to device affiliation and control priority to generate a grouped instruction list.

12. The system according to claim 1, characterized in that, The specific process of the hierarchical feedback includes: classifying the timeliness of the acquired instruction execution status data, operating condition data and safety monitoring information; generating emergency feedback data packets or regular feedback data packets through the edge lightweight AI preprocessing mechanism; and using different feedback channels based on data type and timeliness requirements to synchronously update the full-domain energy flow data map.

Images