Virtual power plant regulation capability test method and system based on deep belief network

Abstract

The invention discloses a virtual power plant regulation capability test method and system based on a deep belief network, and the method comprises the steps: deploying an intelligent terminal system at an aggregation resource side in a virtual power plant, and collecting resource operation data in real time; dynamically loading market rule parameters of the target transaction variety through a rule configuration engine; the method comprises the following steps: preprocessing historical scheduling data of a virtual power plant, collecting resource-level data and virtual power plant-level data, cleaning abnormal data, and interpolating and filling missing values; calculating an adjustment capability index based on the preprocessed data; inputting the indexes and the rule parameters into a pre-trained deep belief network DBN model, and outputting a virtual power plant overall regulation capability predicted value and a marginal contribution weight of each aggregation resource; generating a virtual power plant admission judgment report and a resource optimization suggestion according to a comparison result of the predicted value and the market rule parameter; according to the invention, differential evaluation of the virtual power plant regulation capability indexes is realized, and the access test requirements under diversified power market rules are met.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of virtual power plant regulation capacity testing, and particularly relates to a virtual power plant regulation capacity testing method and system based on a deep belief network. BACKGROUND

[0002] As an important supporting technology of new power systems, virtual power plants have been rapidly developed at home and abroad in recent years. With the increase of renewable energy penetration and the expansion of distributed energy, virtual power plants aggregate distributed energy storage, electric vehicles, adjustable loads and other resources to participate in power market frequency regulation, peak regulation and demand response, becoming a key means to improve the flexibility of power grids and promote clean energy consumption.

[0003] Currently, the testing of virtual power plant regulation capacity mainly relies on a static index system and a preset test scenario. In the testing process, a fixed threshold method or a rule engine is usually used to check key parameters such as adjustable capacity and response rate, for example, by setting power ramping rate and steady-state accuracy to evaluate the performance of a single resource.

[0004] The core defect of the existing testing method is that it cannot dynamically adapt to the differentiated requirements of regulation capacity indicators under different market rules. For example, the frequency regulation market has strict requirements on response speed, while the peak regulation market pays more attention to sustainable output capacity. However, the traditional static testing algorithm lacks flexibility and cannot automatically adjust the testing logic and evaluation standards according to the trading varieties. SUMMARY

[0005] The purpose of the present application is to provide a virtual power plant regulation capacity testing method based on a deep belief network, which can dynamically adapt to the differentiated requirements of regulation capacity indicators under different market rules. On the other hand, a virtual power plant regulation capacity testing system based on a deep belief network is provided.

[0006] Technical solution: The virtual power plant regulation capacity testing method provided by the present application comprises the following steps:

[0007] (1) An intelligent terminal system is deployed on the aggregation side of the virtual power plant to collect resource operation data in real time, wherein the operation data at least includes power value, communication state and device rated parameter;

[0008] (2) The market rule parameters of the target trading variety are dynamically loaded through a rule configuration engine, wherein the market rule parameters include upper limit of response time, lower limit of regulation accuracy and adjustable capacity demand;

[0009] (3) The historical dispatching data of the virtual power plant is preprocessed to collect resource-level data and virtual power plant-level data, and abnormal data is cleaned and missing values are interpolated and filled;

[0010] (4) calculating an adjustment capability index based on the preprocessed data, the adjustment capability index including an adjustable capacity, an adjustment rate, a response time, a duration, and an adjustment accuracy;

[0011] (5) inputting the adjustment capability index and market rule parameters into a pre-trained deep belief network (DBN) model, the DBN model including an input layer, a hidden layer composed of multiple layers of restricted Boltzmann machines (RBMs), and a dynamically configured output layer;

[0012] (6) outputting, by the DBN model, a virtual power plant overall adjustment capability prediction value, the prediction value including an adjustable capacity, an adjustment rate, and a response time, and calculating a marginal contribution weight of each aggregated resource;

[0013] (7) generating a virtual power plant access determination report and a resource optimization suggestion according to a comparison result of the prediction value and the market rule parameters.

[0014] By deploying an intelligent terminal system on the aggregated resource side, not only real-time data collection is realized, but also local preprocessing is performed through edge computing, significantly improving the timeliness and reliability of data collection; by dynamically loading market parameters through a rule configuration engine, differentiating requirements of different trading varieties can be flexibly adapted, supporting second-level rule updating, ensuring that the test standard is synchronized with the market in real time; historical data are deeply preprocessed, abnormal values are eliminated by using an intelligent cleaning algorithm, missing values are filled in combination with spatiotemporal correlation analysis, and data quality is greatly improved; based on a multi-dimensional index calculation system, adjustment performance is comprehensively evaluated from multiple angles such as time, accuracy, and capacity, providing complete feature dimensions for subsequent analysis; through the deep feature extraction capability of the deep belief network model, complex coupling relationships between resources are accurately captured, and nonlinear modeling of adjustment capability is realized; the model output not only predicts the overall performance, but also quantifies the marginal contribution of each resource, providing accurate weight reference for resource scheduling; finally, a report is generated, which not only contains access determination, but also provides optimization suggestions, forming a complete closed loop of "evaluation-decision-optimization", significantly improving the market response capability and operation efficiency of the virtual power plant.

[0015] Preferably, the intelligent terminal system of step 1 further includes the following steps when performing data collection:

[0016] Before deployment, communication protocol compatibility detection is performed on each aggregated resource in the virtual power plant to ensure that the terminal supports the MQTT protocol;

[0017] When collecting data in real time through the edge computing node, the SHA-256 digital signature algorithm is used to ensure data integrity, and the signature timestamp is synchronized with the clock of the scheduling instruction;

[0018] When it is detected that the communication delay exceeds 200 ms or the data packet loss rate is > 5%, the redundant channel switching mechanism is automatically triggered, and the data in this period is marked as a low confidence interval.

[0019] Abnormal fluctuation monitoring is performed on the collected power data, and when the coefficient of variation of the adjustment rate exceeds 0.15, a reverse time window scanning algorithm is started to locate the abnormal source to a specific resource port.

[0020] The optimization design of the intelligent terminal system significantly improves the reliability and security of data collection: in the deployment stage, the system interconnection is ensured through communication protocol compatibility detection, the integrity and timeliness of data transmission are guaranteed through digital signature and timestamp synchronization technology; the communication quality is intelligently monitored and the redundant channel is automatically switched to effectively deal with network abnormal situations, and the confidence interval marking mechanism is used to ensure data reliability; the power abnormal fluctuation monitoring and precise positioning algorithm are innovatively introduced, which can quickly identify and track the abnormal source, providing a high-quality data basis for subsequent data analysis, and comprehensively enhancing the accuracy and robustness of virtual power plant operation state perception.

[0021] Preferably, the rule configuration engine of step 2 realizes dynamic binding of market rules and test indicators through template files in XML or JSON format, which contain response time thresholds, adjustment accuracy calculation methods and access standards; the template files store rule parameters in a hierarchical structure, including:

[0022] Market foundation layer: defines transaction variety identifier, rule effective time range and version control information;

[0023] Technical indicator layer: stores numerical constraint conditions and unit of measurement of upper limit of response time and lower limit of adjustment accuracy;

[0024] Business logic layer: contains segmented function expression of adjustable capacity demand and weight coefficient configuration item;

[0025] The rule configuration engine automatically checks the syntax compliance of the template file when loading, and performs boundary value detection on numerical parameters.

[0026] The innovative design of the rule configuration engine realizes intelligent adaptation and dynamic management of market rules: the standardized template file format (XML / JSON) is used to realize flexible binding of rules and indicators, supporting fast configuration of different transaction varieties; the hierarchical storage structure divides market rules into three levels of basic attributes, technical indicators and business logic, which not only guarantees the integrity of the rule system but also improves the refinement of parameter management; the built-in syntax checking and boundary detection mechanism effectively ensures the compliance and reasonableness of rule configuration, preventing incorrect parameter input, so that the virtual power plant can respond to the dynamic change needs of various power market rules in real time, significantly improving the adaptability and scalability of the test system.

[0027] Preferably, the adjustment capacity indicator of step 4 includes:

[0028] (4.1)Adjustable capacity C adj : the absolute value of the difference between the minimum load average value and the maximum load average value that meets the normal production demand of the user in a given time period, which is calculated as follows:

[0029] Determine the target upper limit P targ，max and lower limit P targ，min of the power of the virtual power plant being tested within the test time period

[0030] Test day one runs according to the target upper limit P targ，max , and record the actual power curve P act，max (k) within the test time period

[0031] Test day two runs according to the target lower limit P targ，min , and record the actual power curve P act，min (k) within the test time period

[0032] Calculate the adjustment capacity:

[0033] C adj (k) = P act，max (k) - P act，min (k)

[0034] In the formula, C adj (k) is the adjustment capacity of the virtual power plant at time k, with units of watts (W), kilowatts (kW), or megawatts (MW);

[0035] (4.2)Adjustment rate v adj : the adjustment capacity completed per unit of time, including upward adjustment rate and downward adjustment rate, which is calculated as follows:

[0036] Determine two steady-state target power values of the virtual power plant being tested within the upward adjustment rate test time period, the smaller power P 1,targ and the larger power P 2,targ ;

[0037] Determine two steady-state target power values of the virtual power plant being tested within the downward adjustment rate test time period, the larger power P 3,targ and the smaller power P 4,targ ;

[0038] The virtual power plant being tested completes an upward adjustment from P 1,targ to P 2,targ , records the time when the power of the virtual power plant being tested first exceeds and always exceeds 10% of P 1,targ , denoted as t1; records the time when the power of the virtual power plant being tested first reaches and maintains 90% of the target power, denoted as t2;

[0039] The up-regulation time calculation formula is: t up = t2-t1;

[0040] The tested virtual power plant completes a downward regulation once from P 3,targ to P 4,targ , records the time when the power of the tested virtual power plant is first lower than and always lower than 10% of P 3,targ during the regulation process, and marks it as t3; records the time when the power of the tested virtual power plant is first reaches and maintains at 110% of the target power, and marks it as t4;

[0041] The down-regulation time calculation formula is: t down = t4-t3;

[0042] The up-regulation rate calculation formula is: v up = (P 2,targ -P 1,targ ) / t up ;

[0043] The down-regulation rate calculation formula is: v down = (P 3,targ -P 4,targ ) / t down ;

[0044] In the formula, v up is the up-regulation rate, and v down is the down-regulation rate;

[0045] (4.3) Response time t resp : the time difference between the time when the instruction issued by the grid dispatching mechanism is executed and the time when the actual power output of the virtual power plant changes by an amplitude exceeding the allowable range of the load steady-state deviation and no longer returns in the trend, and the calculation method is:

[0046] During the up-regulation rate test of the virtual power plant, record the time when the instruction received by the tested virtual power plant is executed, and mark it as t1 and the actual power P 1,act ;

[0047] Set the 10% deviation of the target regulation power P 2,targ from the actual power P 1,act as the regulation dead zone, and record the time t5 when the actual power jumps out of the regulation dead zone;

[0048] The difference between t5 and t1 is the response time of the up-regulation: t resp,up = t5-t1;

[0049] During the down-regulation rate test of the virtual power plant, record the time when the control signal received by the tested virtual power plant is executed, and mark it as t3 and the actual power P 3,act ;

[0050] Set target regulation power P 4,targ With actual power P 3,act 10% of the deviation is the regulation dead zone, record the time t6 when the actual power jumps out of the regulation dead zone;

[0051] The difference between t6 and t3 is the response time of downward regulation: t resp,down = t6-t3;

[0052] (4.4) Duration T dur : the duration of the resource running stably within the target power ±5% range, which is calculated as:

[0053] When the actual power enters the target power ±5% range, start the duration timer;

[0054] If the power exceeds the allowed range for more than 10 seconds, it is determined that the regulation is interrupted, and the current cumulative time T dur is recorded;

[0055] (4.5) Regulation accuracy ε acc : the ratio of the absolute value of the deviation between the actual power and the target power in the regulation steady state stage to the target power, which is calculated as:

[0056] Determine the target power curve P targ (k) of the virtual power plant, recommend using the quantity-price curve simulation result as the target power curve;

[0057] The tested virtual power plant runs according to the target power, and records the actual power P act (k);

[0058] Regulation accuracy calculation:

[0059]

[0060] In the formula, ε acc is the number of sampling points in the steady state regulation stage.

[0061] By constructing a multi-dimensional regulation capacity evaluation index system, the comprehensive and accurate evaluation of the virtual power plant operation performance is realized: the double-day comparison test method is used to calculate the adjustable capacity, which accurately reflects the actual regulation potential of the virtual power plant; the two-way regulation rate is calculated by dynamically capturing the power change process, which objectively evaluates the rapid response capability; the response time measurement method based on the regulation dead zone judgment truly reflects the response speed of the system to the dispatching instruction; the duration calculation method with fault tolerance mechanism is introduced, which scientifically evaluates the stability of the system operation; the regulation accuracy is calculated by the relative value of steady-state deviation, which effectively quantifies the power control accuracy. This set of index system comprehensively describes the regulation characteristics of virtual power plant from multiple dimensions such as capacity, speed and accuracy, providing a scientific basis for subsequent intelligent analysis and optimization decision.

[0062] Preferably, the DBN model of step 5 is constructed by the following method:

[0063] The input layer dimension is configured as 5N+M, where N is the number of aggregated resources, 5 corresponds to the power value, communication state, device rated parameter and dynamic regulation capacity index of each resource, and M is the number of market rule parameters;

[0064] The hidden layer adopts a three-layer restricted Boltzmann machine RBM cascade structure, and the node numbers of each layer are configured in a dimension reduction ratio of 128-64-32, and the contrast divergence algorithm CD-k is used for unsupervised pre-training between layers;

[0065] The output layer dynamically selects the activation function according to the application scenario: when predicting continuous regulation capacity indicators, a linear neuron regression layer is used and a mean square error MSE loss function is configured; when market access classification is performed, a Softmax classification layer is used and a cross-entropy loss function is configured.

[0066] Through the innovative architecture design of deep belief network (DBN), high-precision intelligent prediction of virtual power plant regulation capacity is realized: the input layer adopts multi-dimensional feature fusion design, considering both aggregated resource characteristics and market rule parameters to ensure the integrity of model input; through the cascade structure of multiple layers of restricted Boltzmann machine and the contrast divergence pre-training algorithm, the deep features of data are effectively extracted and the initial parameters of the network are optimized; the output layer innovatively adopts a dynamically configurable structure, which can not only accurately predict continuous regulation indicators through regression analysis, but also realize market access evaluation through classification judgment, so that the model has multi-task processing capability. This intelligent modeling method significantly improves the accuracy and adaptability of virtual power plant regulation capacity evaluation, providing reliable technical support for power market decision-making.

[0067] Preferably, the marginal contribution weight of step 6 is calculated by the back propagation chain rule, and the formula is:

[0068]

[0069] wherein, is the kth layer weight, is the k+1th layer error term, is the kth layer activation value.

[0070] The marginal contribution weight of each aggregated resource is calculated by the back propagation chain rule, which realizes the accurate quantification of the internal resource value of the virtual power plant: based on the error back propagation mechanism of the deep belief network, the relationship between the network weight and the activation value is decomposed layer by layer, and the actual contribution of each aggregated resource to the overall adjustment ability is accurately traced. This weight calculation method based on mathematical derivation not only improves the objectivity and accuracy of resource evaluation, but also dynamically reflects the actual value change of each resource under different operating conditions, providing a scientific basis for resource optimization configuration and dispatching strategy of the virtual power plant, effectively improving the resource utilization efficiency and market competitiveness.

[0071] Preferably, the resource optimization suggestion of step 7 includes ranking according to the marginal contribution weight, and the high weight resource is preferentially involved in the dispatching.

[0072] Through the intelligent ranking mechanism based on the marginal contribution weight, the optimal dispatching strategy of the internal resources of the virtual power plant is realized: the marginal contribution degree of each resource calculated by the DBN model is taken as the core basis, and the resource priority ranking is automatically generated, ensuring that the high-quality resources with outstanding adjustment ability and high reliability are preferentially involved in market response. This data-driven optimization suggestion not only significantly improves the overall adjustment performance and market income of the virtual power plant, but also dynamically adapts to the demand changes of different operating scenarios, maximizes the resource utilization efficiency, and provides scientific and intuitive dispatching decision support for the operator, effectively enhancing the competitive advantage of the virtual power plant in the electricity market.

[0073] In the second aspect, the virtual power plant adjustment capacity test system of the present application comprises:

[0074] The data acquisition module acquires the power data, communication state and device rated parameters of the aggregated resources in real time through the Internet of Things gateway, and stores the historical and real-time data using a time series database;

[0075] The rule configuration module stores the market rule template containing the upper limit of response time, the lower limit of adjustment accuracy and the demand for adjustable capacity, and dynamically loads it to the test process;

[0076] The index calculation module calculates the adjustable capacity, adjustment rate, response time, duration and adjustment accuracy in real time based on the preprocessed resource level and virtual power plant level data;

[0077] The DBN model service module is deployed on a GPU server, includes an implicit layer composed of multiple layers of restricted Boltzmann machines and a dynamically configured output layer, and provides an online prediction interface and calculation of marginal contribution weights of aggregated resources.

[0078] The report generation module compares the predicted value with the market rule parameter, and outputs a report including an access determination result, a predicted value confidence interval and resource optimization suggestions.

[0079] By constructing a complete virtual power plant regulation capacity test system, a full-process closed-loop management from data acquisition to intelligent decision-making is realized: the data acquisition module realizes real-time aggregation and storage of multi-source heterogeneous data relying on the Internet of Things architecture, providing a reliable data foundation for the system; the rule configuration module supports flexible definition and dynamic loading of market rules, ensuring that test standards and market needs are synchronized in real time; the index calculation module comprehensively evaluates the operation characteristics of the virtual power plant through multi-dimensional quantitative analysis; the DBN model service module realizes accurate prediction of regulation capacity and quantification of resource value based on deep learning technology; and the report generation module intelligently integrates and analyzes the results to output an evaluation report with decision-making guidance value. The modules work together to form a complete technology chain of "perception-computation-analysis-decision", significantly improving the test efficiency and decision-making scientificity of the virtual power plant participating in the electricity market.

[0080] Preferably, the data acquisition module includes:

[0081] Edge computing terminal: deployed on each aggregated resource side, completes real-time data acquisition and abnormal data cleaning;

[0082] Intelligent execution terminal: built-in regulation instruction analysis module and power closed-loop control algorithm, real-time receiving of optimization instructions output by the DBN model service module and driving of resource execution.

[0083] Through the innovative design of the edge computing terminal and the intelligent execution terminal, an integrated closed-loop control system of "perception-decision-execution" of the virtual power plant is constructed: the edge computing terminal realizes data acquisition and preprocessing on the resource side, greatly improving the data timeliness and reliability; the intelligent execution terminal deeply integrates instruction analysis and closed-loop control functions, ensuring accurate and rapid execution of optimization strategies. This distributed intelligent architecture not only reduces the computing load of the central system, but also realizes millisecond-level response of control instructions, enabling the virtual power plant to dynamically maintain the optimal operating state and significantly improving the overall regulation performance and market response capability.

[0084] Preferably, the data acquisition module adopts a linear interpolation algorithm to process communication interruption or data loss, the DBN model service module provides a second-level response through a RESTful API, and the output layer supports dynamic switching between a regression task and a classification task, the regression task is used for outputting predicted values of adjustable capacity, adjustment rate and response time, and the classification task is used for generating an access decision result.

[0085] Through intelligent data processing and model service design, the robustness and practicality of the system are significantly improved: the data acquisition module uses a linear interpolation algorithm to intelligently repair missing data, ensuring data integrity; the DBN model service module realizes a second-level response through an efficient API interface, meeting real-time decision-making needs; the innovative dynamic output layer architecture enables the model to flexibly switch between regression prediction and classification determination, supporting both accurate adjustment capacity quantitative evaluation and rapid access decision generation, realizing integrated intelligent services of evaluation and decision-making, and greatly enhancing the efficiency and reliability of virtual power plants participating in power market testing.

[0086] In a third aspect, the present application also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program capable of being loaded and executed by the processor to implement the deep belief network-based virtual power plant adjustment capacity testing method.

[0087] In a fourth aspect, the present application also provides a computer readable storage medium storing a computer program, wherein the computer program is executed by a processor to implement the deep belief network-based virtual power plant adjustment capacity testing method.

[0088] Advantages: Compared with the prior art, the present application has the following significant advantages: 1. Different market parameters of different transaction varieties are dynamically loaded through a rule configuration engine, and combined with the adaptive output layer design of the deep belief network model, the differential evaluation of the virtual power plant adjustment capacity index is realized, meeting the access testing needs under diversified power market rules; 2. Based on multi-dimensional historical data preprocessing and the nonlinear modeling capability of the deep belief network, the key indexes such as the overall adjustable capacity and adjustment rate of the virtual power plant are accurately predicted, and the marginal contribution weight of each aggregated resource is calculated through the back propagation algorithm, providing data support for dispatching priority allocation; 3. Through real-time data integrity verification, communication redundancy switching and abnormal fluctuation positioning mechanism of the intelligent terminal, combined with historical data cleaning and interpolation filling, the influence of data noise on the test results is significantly reduced, ensuring the stability and reliability of the adjustment capacity evaluation; 4. The prediction results are automatically compared with the market rules to generate a report containing access decision and resource optimization suggestions, forming a "evaluation-prediction-optimization" closed loop, and improving the comprehensive competitiveness of the virtual power plant participating in market transactions. BRIEF DESCRIPTION OF DRAWINGS

[0089] Figure 1A flowchart of the method of the present application;

[0090] Figure 2 A flowchart of the regulation rate detection algorithm of the present application;

[0091] Figure 3 A DBN network structure diagram of the present application;

[0092] Figure 4 A schematic diagram of marginal contribution calculation of the present application;

[0093] Figure 5 An adjustable capacity report visualization example of the present application. DETAILED DESCRIPTION

[0094] The technical solutions of the present application will be further described below in combination with the drawings.

[0095] As Figure 1 shown, the embodiment of the present application provides a virtual power plant regulation capacity test method based on a deep belief network, which comprises the following steps:

[0096] (1) Deploy an intelligent terminal system on the resource side to collect the operation data of the aggregated resources in the virtual power plant in real time, including power values, communication states and equipment rated parameters.

[0097] (1.1) After the terminal collects the power data in real time, a lightweight Kalman filtering algorithm is used to eliminate transient noise and pre-calculate the regulation capacity index. The calculation delay needs to be controlled within 50 ms to ensure real-time performance. The breakpoint resume mechanism is supported, and when a network interruption is detected, the data is cached to the local SD card, and after recovery, it is automatically synchronized to the cloud time series database.

[0098] Deploy an intelligent execution terminal on the control side of the virtual power plant, integrate a PID closed-loop control algorithm, and receive the optimization instructions issued by the DBN model in real time. After analysis, the regulation signal is generated to drive the resource execution. The redundant instruction checking is supported: if the main channel delay is > 200 ms, automatically switch to the standby channel to ensure the real-time performance of the instructions.

[0099] Test the communication protocol adaptability, verify whether the terminal is compatible with the resource side protocol, and add a protocol conversion gateway to the incompatible equipment.

[0100] Test the local calculation performance, test the calculation delay and communication stability of the terminal under the full load operation state of the resource, and ensure that the index calculation delay is < 50 ms and the packet loss rate is < 1%.

[0101] (1.2) System docking

[0102] The data acquisition module is connected with the existing system of the virtual power plant through the MQTT protocol, supports real-time data stream transmission; the rule configuration module interacts with the power market trading platform through the RESTful API, dynamically obtains the market rule template; the report generation module is connected with the user management platform, uses the HTTPS protocol to transmit the access report, and ensures data security.

[0103] The real-time power data is packaged in the IEEE C37.118.2 standard format, including timestamp, device ID, power value and communication state; the market rule template includes fields: transaction variety, response time value, adjustment accuracy calculation method, access standard, and checks the legality through JSON Schema, ensures that the required fields are complete and the data type is correct.

[0104] The data acquisition module uses TLS1.3 encryption transmission, the edge computing terminal integrates SM4 national encryption algorithm to encrypt the local storage data, and the intelligent execution terminal uses two-way digital certificate authentication to ensure that the instruction source is reliable.

[0105] (1.3) Test request and qualification verification

[0106] The server receives the virtual power plant test request instruction sent by the client, and the instruction includes the device topology of the target virtual power plant and the grid connection license; the device topology marks the communication address, type and rated power of all aggregated resources.

[0107] Perform automatic qualification verification process: extract the certificate number through OCR technology, and verify online with the energy supervision database to verify the validity of the certificate; check the integrity of the device topology, which must include the communication address of the energy storage unit, new energy unit and load node; if the verification fails, generate a reason code and generate a detailed report.

[0108] (1.4) Dynamic test scheme generation

[0109] According to the resource composition of the target virtual power plant, select test items from the candidate test library;

[0110] Communication interface test: verify HTTP / HTTPS protocol stack compatibility;

[0111] Operation platform test: simulate end-to-end response delay measurement of grid dispatching instructions;

[0112] Adjustable capacity test: inject step power instructions, record the deviation rate of the actual output curve and the DBN predicted value;

[0113] Based on the market rules, automatically set the test threshold (response time ≤ 30 seconds), and generate a test script.

[0114] Test instruction issuing and execution: issuing test instructions through HTTP proxy service deployed on resource terminal, which contains target power, duration and sampling interval.

[0115] Abnormality monitoring mechanism: if communication delay > 200 ms or power mutation (such as ± 20% of rated value) is detected, the test is automatically suspended and an alarm is triggered, and it is resumed after the fault is eliminated.

[0116] (2) Dynamically load market rule parameters of target transaction varieties through rule configuration engine, which at least include upper limit of response time, lower limit of adjustment accuracy and adjustable capacity demand;

[0117] According to the target transaction variety, dynamically load test rules through the rule engine; bind the corresponding test index calculation rules.

[0118] (3) Data preprocessing, collect at least 6 months of historical scheduling data of virtual power plant, including resource level data and virtual power plant level data. Clean abnormal data and interpolate missing values.

[0119] (3.1) Collect at least 6 months of historical scheduling data of virtual power plant, including:

[0120] Resource level data: real-time power of each resource, communication state (0 / 1), maximum / minimum power limit.

[0121] Virtual power plant level data: actual output curve of virtual power plant, market access result (qualified / unqualified).

[0122] (3.2) Data cleaning: remove data segments with communication interruption (communication state is 0) or power exceeding equipment rated range (such as P 实际 > P max ); linear interpolation method is used to fill in missing data

[0123] (3.3) Abnormal value processing: for power mutation, sliding window mean filtering is used, the formula is:

[0124]

[0125] (4) Calculate each index, including adjustable capacity, adjustment rate, response time, duration, adjustment accuracy.

[0126] Collect resource power data through Internet of Things gateway, and calculate the following indexes:

[0127] (4.1) Adjustable capacity test method:

[0128] Determine the upper limit P targ，max and lower limit P targ，min of the target power of the virtual power plant being tested within the test time period;

[0129] Test day one according to target upper limit P targ，max Run, record actual power curve P act，max (k) during test period.

[0130] Test day two according to target lower limit P targ，min Run, record actual power curve P act，min (k) during test period.

[0131] Calculate regulation capacity:

[0132] C adj (k) = P act，max (k) - P act，min (k)

[0133] In the formula: C adj (k) is the regulation capacity of virtual power plant k at the moment, unit is watt (W), kilowatt (kW) or megawatt (MW) and the like.

[0134] (4.2) Regulation rate test method, as shown in Figure 2 :

[0135] Determine two steady-state target power values of the virtual power plant being tested within the upward regulation rate test period, the smaller power P 1,targ and the larger power P 2,targ .

[0136] Determine two steady-state target power values of the virtual power plant being tested within the downward regulation rate test period, the larger power P 3,targ and the smaller power P 4,targ .

[0137] The virtual power plant being tested completes an upward regulation from P 1,targ to P 2,targ , records the moment when the power of the virtual power plant being tested is first higher than and always higher than 10% of P 1,targ during the regulation process, denoted as t1; records the moment when the power of the virtual power plant being tested is first reaches and maintains at 90% of the target power, denoted as t2;

[0138] The upward regulation time calculation formula is: t up = t2-t1

[0139] The virtual power plant being tested completes a downward regulation from P 3,targ to P 4,targ , records the moment when the power of the virtual power plant being tested is first lower than and always lower than P 3,targthe moment when the power of the tested virtual power plant reaches and maintains 110% of the target power for the first time, denoted as t4;

[0140] The downward adjustment time calculation formula is: t down = t4-t3

[0141] The upward adjustment rate calculation formula is: v up = (P 2,targ -P 1,targ ) / t up

[0142] The downward adjustment rate calculation formula is: v down = (P 3,targ -P 4,targ ) / t down

[0143] In the formula: v up is the upward adjustment rate, and v down is the downward adjustment rate.

[0144] (4.3) Response time test method:

[0145] During the upward adjustment rate test of the virtual power plant, record the moment when the tested virtual power plant receives the instruction execution, denoted as t1 and the actual power P 1,act

[0146] Set the target adjustment power P 2,targ and the 10% deviation of the actual power P 1,act as the adjustment dead zone, and record the moment t5 when the actual power jumps out of the adjustment dead zone;

[0147] The difference between t5 and t1 is the response time of upward adjustment: t resp,up = t5-t1

[0148] During the downward adjustment rate test of the virtual power plant, record the moment when the tested virtual power plant receives the control signal, denoted as t3 and the actual power P 3,act ;

[0149] Set the target adjustment power P 4,targ and the 10% deviation of the actual power P 3,act as the adjustment dead zone, and record the moment t6 when the actual power jumps out of the adjustment dead zone

[0150] The difference between t6 and t3 is the response time of downward adjustment: t resp,down = t6-t3

[0151] (4.4) Duration:

[0152] The length of time that the resource remains stable at the target power under the regulation instruction. When the actual power enters the range of ±5% of the target power, the duration timer is started; if the power exceeds the allowed range for more than 10 seconds, it is determined that the regulation is interrupted, and the current cumulative time T is recorded dur

[0153] (4.5) Regulation accuracy: the degree of deviation of the actual power from the target power.

[0154] Determine the target power curve P of the virtual power plant targ (k), recommend the results simulated by the quantity-price curve as the target power curve;

[0155] The tested virtual power plant operates according to the target power, and the actual power P is recorded act (k)

[0156] The regulation accuracy calculation formula is:

[0157]

[0158] In the formula: ε acc is the regulation accuracy of the given target power.

[0159] (5) Input the aggregate resource regulation capability index and market rule parameter into the pre-trained deep belief network (DBN) model, and the DBN model includes an input layer, a hidden layer composed of multiple layers of restricted Boltzmann machines (RBM), and a dynamically configured output layer. The specific steps are as follows.

[0160] (5.1) Network structure design

[0161] Input layer: contains aggregate resource indicators (5 types x N resources) and market rule parameters (such as upper limit of response time), with a total dimension of 5N+M (M is the number of rule parameters).

[0162] Hidden layer: composed of three layers of restricted Boltzmann machines (RBM) stacked, with node numbers of 128, 64, and 32 respectively, and each layer of RBM extracts features through unsupervised learning.

[0163] Output layer: dynamically configured according to business scenarios.

[0164] (5.2) Unsupervised pre-training

[0165] Training data: normalized resource-level feature matrix (dimension 5N).

[0166] Training parameters: contrast divergence (CD-1) algorithm, learning rate 0.01, iteration 100 times. After each layer of RBM is trained, the hidden layer output is used as the input of the next layer.

[0167] (5.3) Supervised fine-tuning

[0168] Regression task: using mean square error (MSE) loss function, Adam optimizer (learning rate 0.001), 200 iterations.

[0169] Classification task: using cross-entropy loss function, Softmax output layer, threshold set to 0.5 (probability ≥ 0.5 is determined as "qualified").

[0170] (6) Output the virtual power plant overall regulation capacity prediction value through the DBN model, including the adjustable capacity Regulation rate And response time And calculate the marginal contribution weight of each aggregated resource.

[0171] Input real-time indicators and market parameters into the pre-trained DBN model to output prediction results, the specific steps are as follows:

[0172] (6.1) Deep belief network (DBN) modeling, as shown in Figure 3 :

[0173] Input layer: aggregated resources C adj ,v adj ,t resp ,T dur ,ε acc And access constraints of trading varieties (such as upper limit of response time t req , lower limit of regulation accuracy ε min )

[0174] Hidden layer: 3 layers of restricted Boltzmann machine (RBM), node number is 128-64-32 respectively, each layer of RBM extracts features through unsupervised learning;

[0175] RBM energy function:

[0176]

[0177] Where v is the visible layer (input data), h is the hidden layer, and w ij is the weight matrix;

[0178] Output layer: dynamically configured according to business scenarios:

[0179] Regression task: 1 linear neuron, virtual power plant regulation capacity index

[0180] Classification task: softmax layer outputs "qualified" or "unqualified".

[0181] (6.2) Marginal fitting algorithm:

[0182] The Shapley value method is used to quantify the marginal contribution of each resource to the overall capacity of the virtual power plant, and the steps are as follows:

[0183] As Figure 4 shown, define the output target and calculate the gradient: by backpropagation, calculate the partial derivative of the output target with respect to each input feature, which is the marginal contribution of the resource, and the calculation formula of the marginal contribution value is:

[0184]

[0185] wherein, is the weight of the kth layer, is the error term of the k+1th layer, is the activation value of the kth layer.

[0186] (6.3) In a specific embodiment, for example, the marginal contribution value of the adjustable capacity of the virtual power plant is calculated

[0187]

[0188] wherein W (l) represents the weight matrix of the lth layer, σ'(z) is the derivative of the activation function, is the weighted input of the jth neuron of the lth layer.

[0189] Marginal contribution normalization:

[0190]

[0191] Regression mode: predict the adjustable capacity of the virtual power plant and the confidence interval.

[0192] Classification mode: output the admission probability.

[0193] According to the real-time weight analysis, the priority of the resource is adjusted, the resource aggregation strategy is adjusted, the high-weight resource is preferentially involved in the scheduling, and the response efficiency of the virtual power plant is improved.

[0194] (7) According to the comparison result of the predicted value and the market rule parameter, a virtual power plant admission decision report and a resource optimization suggestion are generated.

[0195] Report content: performance indicators of each resource, predicted value and confidence interval of the overall capacity of the virtual power plant, and marginal contribution analysis result, as Figure 5 shown.

[0196] Based on the similar inventive concept, the embodiments of the present application also provide a virtual power plant regulation capacity test system corresponding to the virtual power plant regulation capacity test method, comprising:

[0197] The data acquisition module acquires power data, communication states and equipment rated parameters of the aggregated resources in real time through the Internet of Things gateway, and stores historical and real-time data in a time series database;

[0198] The rule configuration module stores market rule templates containing upper limit of response time, lower limit of adjustment accuracy and adjustable capacity demand, and dynamically loads the templates to the test process;

[0199] The index calculation module calculates adjustable capacity, adjustment rate, response time, duration and adjustment accuracy in real time based on preprocessed resource-level and virtual power plant-level data;

[0200] The DBN model service module is deployed on a GPU server, contains an implicit layer composed of multiple layers of restricted Boltzmann machines and a dynamically configured output layer, provides an online prediction interface and marginal contribution weight calculation of the aggregated resources, and outputs an admission decision result, a prediction value confidence interval and resource optimization suggestions.

[0201] The report generation module compares the predicted value with market rule parameters, and outputs a report containing an admission decision result, a prediction value confidence interval and resource optimization suggestions.

[0202] The data acquisition module comprises:

[0203] The edge computing terminal is deployed on each aggregated resource side, and completes real-time data acquisition and abnormal data cleaning;

[0204] The intelligent execution terminal is internally provided with an adjustment instruction analysis module and a power closed-loop control algorithm, receives the optimization instructions output by the DBN model service module in real time, and drives the resources to execute.

[0205] The data acquisition module adopts a linear interpolation algorithm to process communication interruption or data loss, the DBN model service module provides a second-level response through a RESTful API, and the output layer supports dynamic switching between regression tasks and classification tasks, the regression tasks are used to output predicted values of adjustable capacity, adjustment rate and response time, and the classification tasks are used to generate an admission decision result.

[0206] The application further discloses an electronic device.

[0207] Specifically, the electronic device can be a computer device such as a desktop computer, a notebook computer, a palm computer, and a cloud server. The computer device can include, but is not limited to, a processor and a memory. The processor and the memory can be connected through a bus or other means. The processor can be a central processing unit (CPU). The processor can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, graphics processing units (GPUs), embedded neural-network processing units (NPUs) or other dedicated deep learning co-processors, discrete gates or transistor logic, discrete hardware components, or combinations thereof.

[0208] The memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs and modules. The processor performs various functional applications and data processing of the processor by running the non-transitory software programs, instructions and modules stored in the memory. The memory can include a program storage area and a data storage area, wherein the program storage area can store application programs required by the control unit and at least one function; and the data storage area can store data created by the processor and the like. In addition, the memory can include a high-speed random access memory, and can also include a non-transitory memory. In some embodiments, the memory can optionally include a memory remotely arranged relative to the processor, and these remote memories can be connected to the processor through a network. Examples of the network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0209] The application further discloses a computer-readable storage medium.

[0210] Specifically, the computer-readable storage medium is used to store a computer program, and the computer program is executed by the processor to implement the method in the above method embodiments.

[0211] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments of the method of the present application can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the above-mentioned embodiments of the method. The storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM), a random access memory (RAM), a flash memory, a hard disk drive (HDD) or a solid-state drive (SSD), etc. The storage medium can also include a combination of the above-mentioned types of memories.

Claims

1. A method for testing regulation capability of a virtual power plant based on a deep belief network, characterized in that, The method comprises the following steps: (1) aggregating resource-side intelligent terminal systems in a virtual power plant, collecting resource operation data in real time, wherein the operation data at least includes power value, communication state and device rated parameter; (2) dynamically loading market rule parameters of target transaction varieties through a rule configuration engine, wherein the market rule parameters include upper limit of response time, lower limit of adjustment accuracy and adjustable capacity demand; (3) preprocessing historical scheduling data of the virtual power plant, collecting resource-level data and virtual power plant-level data, cleaning abnormal data and interpolating missing values; (4) calculating adjustment capacity indicators based on the preprocessed data, wherein the adjustment capacity indicators include adjustable capacity, adjustment rate, response time, duration and adjustment accuracy; (5) inputting the adjustment capacity indicators and market rule parameters into a pre-trained deep belief network (DBN) model, wherein the DBN model includes an input layer, a hidden layer composed of multiple layers of restricted Boltzmann machines (RBMs) and a dynamically configured output layer; (6) outputting a virtual power plant overall adjustment capacity prediction value through the DBN model, wherein the prediction value includes adjustable capacity, adjustment rate and response time, and calculating the marginal contribution weight of each aggregated resource; (7) generating a virtual power plant access determination report and resource optimization suggestion according to the comparison result of the prediction value and the market rule parameters.

2. The virtual power plant regulation capability test method according to claim 1, wherein, The intelligent terminal system of step 1 further comprises the following steps when performing data collection: Before deployment, detect the communication protocol compatibility of each aggregated resource in the virtual power plant to ensure that the terminal supports MQTT protocol; When collecting data in real time through an edge computing node, use SHA-256 digital signature algorithm to ensure data integrity, and synchronize the signature timestamp with the scheduling instruction clock; When detecting that the communication delay exceeds 200 ms or the data packet loss rate is > 5%, automatically trigger the redundant channel switching mechanism, and mark the data in this period as a low confidence interval; Implement abnormal fluctuation monitoring on the collected power data, and when the coefficient of variation of the adjustment rate exceeds 0.15, start the reverse time window scanning algorithm to locate the abnormal source to a specific resource port. 3.The virtual power plant regulation capability test method according to claim 1, wherein, The rule configuration engine of step 2 dynamically binds market rules and test indicators through template files in XML or JSON format, wherein the template files contain response time threshold, adjustment accuracy calculation method and access standard; the template files store rule parameters in a hierarchical structure, including: Market base layer: define transaction variety identifier, rule effective time range and version control information; Technical indicator layer: store numerical constraint conditions and unit of measurement of upper limit of response time and lower limit of adjustment accuracy; Business logic layer: contains segmented function expression of adjustable capacity demand and weight coefficient configuration item; The rule configuration engine automatically checks the syntax compliance of the template file when loading, and performs boundary value detection on numerical parameters. 4.The virtual power plant regulation capability test method according to claim 1, wherein, The adjustment capacity indicators of step 4 include: (4.1) Adjustable capacity C adj : The absolute value of the difference between the minimum load average and the maximum load average that meets the normal production demand of the user in a given time period, which is calculated as follows: determining an upper limit P of the target power of the virtual power plant under test for the test period targ，max and a lower limit P targ，min ; Test day one at target upper limit P targ，max Run, record actual power curve P over test period act，max (k); Test day two according to target lower limit P targ，min Run, record actual power curve P over test period act，min (k); Calculate the adjustment capacity: C adj (k) = P act，max (k) - P act，min (k) In the formula, C adj (k) is the regulation capacity of the virtual power plant k at the moment, in units of watts (W), kilowatts (kW), or megawatts (MW); (4.2) Adjusting rate v adj : Adjusting capacity completed in unit time, including upward adjusting rate and downward adjusting rate, which is calculated in the following way: determining two steady state target power values for the virtual power plant under test for the up-regulation rate test time period, a lower power P 1,targ and a higher power P 2,targ ; determining two steady state target power values for the virtual power plant under test for the down regulation rate test time period, a larger power P 3,targ and a smaller power P 4,targ ; The tested virtual power plant makes an upward adjustment in the direction from P 1,targ to P 2,targ at a time when the power of the tested virtual power plant is first higher than and always higher than 10% of P 1,targ during the adjustment process, recorded as t1; and records the time when the power of the tested virtual power plant first reaches and maintains 90% of the target power, recorded as t2; The up-regulation time calculation formula is: t up = t2-t1; The tested virtual power plant performs one downward adjustment in the direction from P 3,targ to P 4,targ . The time at which the power of the tested virtual power plant first falls below and remains below 10% of P 3,targ during the adjustment is recorded, denoted as t3; the time at which the power of the tested virtual power plant first reaches and remains at 110% of the target power is recorded, denoted as t4; The down-regulation time calculation formula is: t down = t4 - t3; The up-regulation rate calculation formula is: v up = (P 2,targ - P 1,targ ) / t up ; The down-regulation rate calculation formula is: v down = (P 3,targ - P 4,targ ) / t down ; where: v up v is the up-regulation rate down v is the down-regulation rate (4.3) Response time t resp The time difference between the moment when the instruction from the grid dispatching agency is executed and the moment when the actual power output of the virtual power plant changes beyond the allowed range of load steady-state deviation and no longer returns in the trend. The calculation method is: During the virtual power plant in upward regulation rate test, record the moment when the instruction received by the tested virtual power plant is executed, denoted as t1 and the actual power P 1,act ; Set target adjustment power P 2,targ With actual power P 1,act The 10% deviation is the adjustment dead zone. Record the moment t5 when the actual power jumps out of the adjustment dead zone. The difference between t5 and t1 is the response time for up-regulation: t resp,up = t5 - t1; During the virtual power plant in down-regulation rate test, record the time when the tested virtual power plant receives the control signal, denoted as t3 and the actual power P 3,act ; Set target adjustment power P 4,targ With actual power P 3,act The 10% deviation is the adjustment dead zone. Record the moment t6 when the actual power jumps out of the adjustment dead zone. The difference between t6 and t3 is the response time of the down-regulation: t resp,down = t6 - t3; (4.4) Duration T dur : the duration of time that the resource is operating steadily within the target power ± 5%, calculated as: When the actual power enters the range of target power ± 5%, start the duration timer; If the power exceeds the allowable range for more than 10 seconds, it is determined that the adjustment is interrupted, and the current cumulative time T is recorded dur ; (4.5) Regulation accuracy ε acc : the ratio of the absolute value of the deviation of the actual power from the target power in the regulation steady state phase to the target power, which is calculated as follows: Determining a virtual power plant target power curve P targ (k), recommending to simulate the clearing result with a quantity-price curve as the target power curve; The tested virtual power plant is operated according to the target power, the actual power P is recorded act (k); Adjustment accuracy calculation: In the formula, ε acc is the number of sampling points in the steady-state regulation stage. 5.The virtual power plant regulation capability test method according to claim 1, wherein, The DBN model of step 5 is constructed in the following way: The input layer dimension configuration is 5N+M, wherein N is the number of aggregated resources, 5 corresponds to the power value, communication state, device rated parameter and dynamic adjustment capability index of each resource, and M is the number of market rule parameter; The hidden layer adopts a cascade structure of three layers of restricted Boltzmann machine (RBM), the node numbers of each layer are configured in a dimension reduction ratio of 128-64-32, and the contrast divergence algorithm (CD-k) is adopted between layers for unsupervised pre-training; The output layer dynamically selects an activation function according to an application scenario: when predicting a continuous type adjustment capability index, a linear neuron regression layer is adopted and a mean square error (MSE) loss function is configured; when performing market access classification, a Softmax classification layer is adopted and a cross-entropy loss function is configured. 6.The virtual power plant regulation capability test method according to claim 1, wherein, The marginal contribution weight of step 6 is calculated by a back propagation chain rule, and the formula is: wherein, is the kth layer weight, is the k+1th layer error term, is the kth layer activation value. 7.The virtual power plant regulation capability test method according to claim 1, wherein, The resource optimization suggestion of step 7 includes ordering according to the marginal contribution weight, and high weight resources are preferentially involved in scheduling.

8. A deep belief network-based virtual power plant regulation capability test system, characterized in that, Comprise: The data acquisition module acquires power data, communication state and device rated parameter of the aggregated resources in real time through the Internet of Things gateway, and stores historical and real-time data using a time series database; The rule configuration module stores market rule templates including upper limit of response time, lower limit of adjustment accuracy and adjustable capacity demand, and dynamically loads them to the test process; The index calculation module calculates adjustable capacity, adjustment rate, response time, duration and adjustment accuracy in real time based on preprocessed resource level and virtual power plant level data; The DBN model service module is deployed on a GPU server, and comprises a hidden layer composed of multiple layers of restricted Boltzmann machine and a dynamically configured output layer, and provides an online prediction interface and calculation of marginal contribution weight of the aggregated resources; The report generation module compares the predicted value with the market rule parameter, and outputs a report including access determination result, prediction value confidence interval and resource optimization suggestion.

9. The virtual power plant regulation capability testing system of claim 8, wherein, The data acquisition module comprises: Edge computing terminal: deployed on each aggregated resource side, completes real-time data acquisition and abnormal data cleaning; Intelligent execution terminal: built-in adjustment instruction analysis module and power closed-loop control algorithm, receives the optimization instruction output by the DBN model service module in real time and drives the resource to execute.

10. The virtual power plant regulation capability testing system of claim 8, wherein, The data acquisition module adopts a linear interpolation algorithm to process communication interruption or data loss, the DBN model service module provides a second-level response through a RESTful API, and the output layer supports dynamic switching between regression tasks and classification tasks, the regression task is used to output the predicted value of adjustable capacity, adjustment rate and response time, and the classification task is used to generate the access determination result.

Images