Virtual power plant regulation and control method and system based on dynamic parameter identification and hierarchical collaboration

Abstract

The invention discloses a virtual power plant regulation and control method and system based on dynamic parameter identification and hierarchical collaboration, and belongs to the technical field of virtual power plant dynamic regulation and control, and the method comprises the steps: collecting multi-dimensional data of building distributed resources, the building distributed resources comprise roof photovoltaic equipment, EV charging piles and energy storage equipment, integrating the multi-dimensional data through federal learning; building distributed resources are aggregated according to power generation characteristics and power generation economy of the building distributed resources, external characteristic parameters of the virtual power plant are generated, and an energy and operation reserve optimization model of the virtual power plant is constructed; power generation parameters of an RLS algorithm optimization model are updated online, the uncertainty of virtual power plant operation is considered, a layered time sequence cooperation framework is constructed to dynamically adjust the virtual power plant energy and operation reserve optimization model, regulation and control instructions are generated and sent to distributed resource equipment of all buildings, and the operation state and power output of the equipment are controlled.

Description

Technical Field

[0001] The present invention belongs to the technical field of virtual power plant dynamic regulation, and more specifically, relates to a virtual power plant regulation method and system based on dynamic parameter identification and hierarchical collaboration. Background Art

[0002] With the wide application of distributed energy sources, such as the large-scale deployment of rooftop photovoltaic devices, EV charging piles, and energy storage devices in buildings, the traditional power system regulation methods face many challenges. As an effective means of integrating distributed energy resources, virtual power plants have received extensive attention.

[0003] However, there are certain deficiencies in the existing technologies for the regulation of virtual power plants. On the one hand, for the multi-dimensional data processing of building distributed resources, traditional methods are often difficult to effectively integrate while ensuring data privacy. The data formats, acquisition frequencies, and data volumes of different types of devices vary greatly. If a centralized data processing method is adopted, it will involve a large amount of data transmission and storage, which not only increases costs but also poses a risk of data leakage, and the existing technologies lack effective distributed data integration methods. On the other hand, in the process of virtual power plant resource aggregation, previous studies often only consider a single factor, such as simply aggregating based on power generation characteristics or power generation economy, without comprehensive consideration, resulting in inaccurate external characteristic parameters of the generated virtual power plant and making it difficult for the constructed optimization model to adapt to the complex and changing actual operating conditions.

[0004] In addition, there are many uncertain factors in the operation of virtual power plants, such as the fluctuation of the power generation power of photovoltaic devices caused by changes in light intensity, the randomness of user charging demands, etc. Most of the existing regulation methods are based on static models and lack effective response mechanisms for these uncertainties. They cannot dynamically adjust the optimization model according to the real-time changing power generation parameters, resulting in a large deviation between the regulation instructions and the actual operating conditions and making it difficult to achieve precise control of the device operating state and power output. Moreover, the existing technologies do not make full use of the feedback information in the regulation process and fail to timely and effectively adjust the optimization model based on the actual operating data, thus affecting the overall operating efficiency and stability of the virtual power plant. Summary of the Invention

[0005] To solve the deficiencies in the existing technologies, the present invention provides a virtual power plant regulation method and system based on dynamic parameter identification and hierarchical collaboration.

[0006] The present invention adopts the following technical solutions.

[0007] The first aspect of the present invention provides a virtual power plant regulation method based on dynamic parameter identification and hierarchical collaboration, including the following steps: Collect multi-dimensional data of building distributed resources, where the building distributed resources include rooftop photovoltaic devices, EV charging piles, and energy storage devices, and integrate their multi-dimensional data through federated learning; Aggregate building distributed resources according to their power generation characteristics and power generation economy, generate external characteristic parameters of the virtual power plant, and construct an optimization model for the energy and operation reserve of the virtual power plant; online update the power generation parameters of the optimization model for the energy and operation reserve of the virtual power plant through the RLS algorithm, consider the uncertainty of the operation of the virtual power plant, construct a hierarchical time-series coordination framework, and dynamically adjust the optimization model for the energy and operation reserve of the virtual power plant in combination with the real-time power generation parameters updated by the RLS algorithm; According to the optimization model for the energy and operation reserve of the virtual power plant dynamically adjusted by the hierarchical time-series coordination framework, generate control instructions and send them to each building distributed resource device, control the operation status and power output of the device, and monitor the operation status of the virtual power plant in real time, and adjust the optimization model for the energy and operation reserve of the virtual power plant according to the feedback information of the actual operation data.

[0008] Optionally, the aggregating the building distributed resources according to their power generation characteristics and power generation economy, generating external characteristic parameters of the virtual power plant, and constructing an optimization model for the energy and operation reserve of the virtual power plant includes: Regard each device as a separate cluster, where the devices include rooftop photovoltaic devices, EV charging piles, and energy storage devices; Calculate the power generation characteristic matching degree between different devices according to the power generation characteristic curve of each device; Aggregate the devices based on the power generation characteristic matching degree between every two devices to generate external characteristic parameters of the virtual power plant; Construct an optimization model for the energy and operation reserve of the virtual power plant based on the device aggregation result and the external characteristic parameters of the virtual power plant, including a benchmark loss function, power balance constraints, energy storage system constraints, EV constraints, PV output constraints, and grid interaction power constraints, and its benchmark loss function is:

[0009] Wherein, is the grid interaction cost, is the operation cost of the th EV charging pile, is the penalty cost for curtailed light.

[0010] Optionally, the external characteristic parameters of the virtual power plant include rated power, lower limit of economic power, upper limit of economic power, dead zone of unit control error, upward ramp rate, downward ramp rate, upper limit of secondary frequency regulation upward adjustment, lower limit of secondary frequency regulation downward adjustment, minimum time interval between unit shutdown and startup, and cost-capacity function.

[0011] Optionally, the online update of the power generation parameters of the virtual power plant energy and operation reserve optimization model by the RLS algorithm includes: Based on the multi-dimensional data related to photovoltaic equipment, the RLS algorithm is used to dynamically correct the photovoltaic output prediction model and online update the efficiency of the photovoltaic inverter; Based on the multi-dimensional data related to EV charging piles and energy storage devices, the RLS algorithm is used to online update the energy storage SOC constraint and the EV frequency modulation time constant.

[0012] Optionally, considering the uncertainty of the operation of the virtual power plant, the construction of a hierarchical time-series collaborative framework includes: The hierarchical time-series collaborative framework includes a day-ahead layer, an intraday layer, and a real-time layer; In the day-ahead layer, based on the historical data of photovoltaic output, EV behavior, and the grid and market, and the photovoltaic output prediction model, a photovoltaic-EV joint scenario is generated, and the high-risk scenarios that meet the preset risk threshold are selected and injected into the optimization model. The economic losses of the high-risk scenarios are quantified by conditional value-at-risk constraints, and a benchmark scheduling plan is formulated; In the intraday layer, based on the periodically collected photovoltaic output prediction data, the scheduling plan is dynamically adjusted through rolling optimization, and the distribution parameters of the photovoltaic output prediction error are updated by Bayesian online learning; In the real-time layer, based on the second-by-second updated measured photovoltaic output data and the grid frequency modulation demand signal, the distribution parameters of the photovoltaic output prediction error are corrected in real time by event-triggered Bayesian online learning, the standby capacity call strategy is dynamically adjusted, and the charging and discharging priorities of the charging piles and energy storage devices are optimized according to the time-of-use electricity price and the real-time load demand.

[0013] Optionally, the day-ahead layer includes: Based on the historical data of photovoltaic output, EV behavior, and the grid and market, and the photovoltaic output prediction model, a photovoltaic-EV joint scenario is generated, and the high-risk scenarios that meet the preset risk threshold are selected and injected into the optimization model. Cluster analysis is performed on the historical photovoltaic output data and EV behavior data, and the photovoltaic output and EV available capacity are used as feature vectors to generate a set of photovoltaic-EV joint scenarios; Calculate the net load deviation of each scenario, and select the high-risk scenarios that meet the preset risk threshold; Introduce the conditional value-at-risk constraint into the objective function of the virtual power plant energy and operation reserve optimization model, and minimize the conditional value-at-risk. The conditional value-at-risk constraint formula is:

[0014] Add the conditional value-at-risk constraint to the benchmark loss function to generate a benchmark scheduling plan:

[0015] Among them, is the quantile threshold; is the confidence level; is the expectation operator; is the benchmark loss function.

[0016] Optionally, the intraday layer includes: Based on the periodically collected photovoltaic power output prediction data, dynamically adjust the scheduling plan through rolling optimization, and the rolling optimization time window is ; Add a real-time correction term to the benchmark objective function:

[0017] Among them, is the adjustment cost coefficient; is the penalty term for the adjustment amount of the scheduling plan; And update the distribution parameters of the photovoltaic power output prediction error through Bayesian online learning.

[0018] Optionally, the real-time layer includes: Based on the second-level updated measured photovoltaic power output data and the grid frequency regulation demand signal, real-time correct the distribution parameters of the photovoltaic power output prediction error through event-triggered Bayesian online learning, and dynamically adjust the reserve capacity calling strategy, including: Determine the grid frequency deviation according to the grid frequency regulation demand for calculating the reserve power demand:

[0019] Among them, is the grid frequency regulation deviation; is the reserve power demand; is the frequency regulation demand; The calculation formula for the reserve capacity allocation constraint is:

[0020] Among them, is the th standby discharge power of the building charging pile; is the standby discharge power of the energy storage device.

[0021] Optionally, optimizing the charging and discharging priorities of building charging piles and energy storage devices according to time-of-use electricity prices and real-time load demands includes: Receive the time-of-use electricity price signal. During peak hours, the building charging piles and energy storage devices give priority to discharging. During off-peak hours, the building charging piles and energy storage devices give priority to charging; Monitor the real-time load demand and predict the load demand. If the real-time load demand is greater than the predicted load demand, call the energy storage device to discharge and reduce the power purchase demand; if the real-time load demand is less than the predicted load demand, give priority to consuming the surplus PV power to reduce PV curtailment.

[0022] The second aspect of the present invention provides a virtual power plant regulation system based on dynamic parameter identification and hierarchical cooperation. Based on the virtual power plant regulation method described in the first aspect of the present invention, the system includes: A data acquisition and integration module for collecting multi-dimensional data of building distributed resources and integrating them through federated learning; a resource aggregation and modeling module for aggregating building distributed resources according to generation characteristics and economy, generating external characteristic parameters of the virtual power plant, and constructing an energy and operating reserve optimization model; a parameter update and dynamic adjustment module for online updating the generation parameters through the RLS algorithm and dynamically adjusting the optimization model in combination with the hierarchical time series cooperation framework; an optimization regulation and status monitoring module for generating regulation instructions and sending them to the building distributed resource devices to control their operating status and power output, and at the same time, real-time monitoring the operating status of the virtual power plant and adjusting the optimization model according to the feedback information.

[0023] Compared with the prior art, the beneficial effects of the present invention at least include: (1) Through the federated learning technology, a balance between privacy protection and efficient data integration is achieved when collecting multi-dimensional data of building distributed resources. The fusion of multi-dimensional data provides comprehensive input for subsequent optimization, solves the model deviation problem caused by single data dimension in traditional methods, and improves the refinement and scenario adaptability of the scheduling strategy; (2) The resource dynamic aggregation mechanism based on generation characteristics and economy can flexibly adapt to the spatio-temporal heterogeneity of distributed resources, match the grid scheduling requirements by generating external characteristic parameters of the virtual power plant, and improve the resource utilization rate; (3) Introduce the recursive least squares method to online identify and update the generation parameters, which can real-time correct the model errors caused by equipment aging and environmental fluctuations; (4) The hierarchical optimization combined with the hierarchical time series cooperation framework can cover medium- and long-term risk prediction and quickly respond to the second-level grid frequency modulation demand. This dynamic adjustment mechanism can effectively improve the adaptability of the virtual power plant to uncertainties. Description of the Drawings

[0024] Figure 1 It is a schematic flowchart of the method provided according to the embodiments of the present invention. Detailed Embodiments

[0025] To make the objectives, technical solutions and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the spirit of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the protection scope of the present invention.

[0026] To more clearly introduce the prominent substantive features of the present invention and the significant progress brought to the prior art, the following introduces an application example of implementing the present invention.

[0027] The following will detail the embodiments of the present invention in conjunction with the accompanying drawings. The application example specifically includes: In Embodiment 1 of the present invention, a virtual power plant regulation method based on dynamic parameter identification and hierarchical collaboration is provided. As Figure 1 shown, it includes the following steps: Step 1: Collect multi-dimensional data of building distributed resources and integrate the multi-dimensional data through federated learning. Preferably, in Step 1, the building distributed resources include rooftop photovoltaic devices, EV charging piles, and energy storage devices. Preferably, in Step 1, the multi-dimensional data of the building distributed resources includes: The V2G status of EV charging piles, user travel patterns, rooftop photovoltaic output curves, energy storage SOC, and building cooling / heating load data.

[0028] Preferably, in Step 1, integrating the multi-dimensional data through federated learning includes: Deploy computing nodes locally in each building, use the collected EV charging pile and user travel data to train the EV user behavior model. After training, only upload the parameters of the EV user behavior model to the VPP central node to avoid the leakage of raw data and achieve privacy protection for users.

[0029] Step 2: Aggregate the building distributed resources according to their power generation characteristics and power generation economy, generate the external characteristic parameters of the virtual power plant, and construct an optimization model for the energy and operation reserve of the virtual power plant. Preferably, in Step 2, aggregating the building distributed resources according to their power generation characteristics and power generation economy includes: Step 2.1: Regard each device as a separate cluster. The devices include rooftop photovoltaic devices, EV charging piles, and energy storage devices.

[0030] Step 2.2: Calculate the power generation characteristic matching degree between different devices according to the power generation characteristic curves of each device. Further preferably, Step 2.2 includes: The degree of coincidence of power generation characteristics between different devices is calculated by the following formula:

[0031] In the formula, is the degree of coincidence between device and ; is the regulation potential at the regulation cycle point in the power generation characteristic curve of device ; is the regulation potential at the regulation cycle point in the power generation characteristic curve of device ; is the number of regulation cycle points for comparison; using this formula, the degree of coincidence of two power generation characteristic curves can be simply compared. The higher the degree of coincidence, the is smaller.

[0032] Step 2.3: Aggregate the devices based on the degree of coincidence of power generation characteristics between every two devices; Further preferably, the step 2.3 includes: Aggregate the devices according to the agglomerative hierarchical clustering method, and its objective function is as follows:

[0033] Wherein, is the number of all combinations of taking any two numbers from the number of ; is the number of resources within the group after aggregation; is the load regulation potential of the load device ; is the total planned regulation amount.

[0034] It should be noted that the technical effects achieved by the agglomerative hierarchical clustering through the above objective function include: 1) The virtual power plant formed after aggregation can meet the expected regulation amount of the power grid company; 2) The characteristic parameters of the aggregated resource device groups are similar, which is convenient for centralized control; 3) The number of load devices aggregated by the virtual power plant can be minimized, that is, the number of devices called is small, reducing the impact on users.

[0035] Preferably, in the step 2, the external characteristic parameters of the virtual power plant include: Rated power, lower limit of economic power, upper limit of economic power, unit control error dead zone, upward ramp rate, downward ramp rate, upper limit of secondary frequency regulation (AGC) upward adjustment, lower limit of secondary frequency regulation (AGC) downward adjustment, minimum time interval between unit shutdown and startup, and cost-capacity function.

[0036] Step 2.4, construct an optimization model for virtual power plant energy and operating reserve, including: Benchmark loss function:

[0037] Among them, is the grid interaction cost, representing the cost of purchasing or selling electricity between the building and the grid:

[0038] Among them, is the power interaction value between the building and the grid. A positive value indicates electricity purchase, and a negative value indicates electricity sale; is the time-of-use electricity price. When purchasing electricity, it is billed at and when selling electricity, it is settled at Usually, ; is the operating cost of the EV charging pile:

[0039] Among them, is the charging and discharging power of the EV charging pile at time . Discharging is negative; , and are the equipment characteristic parameters of the EV charging pile, obtained by fitting experimental or historical data.

[0040] is the penalty cost for curtailment of light:

[0041] Among them, is the penalty coefficient for curtailment of light, in yuan per kilowatt, set according to market rules; is the curtailment of light power.

[0042] Constraint conditions: (1) Power balance constraint

[0043] The sum of the grid power purchase, PV output, EV charging and discharging power, and energy storage device charging and discharging power should be equal to the building load demand; (2) Energy storage system constraint SOC Dynamic Equation:

[0044] Wherein, is the state of charge of the energy storage device at time, representing the proportion of the current remaining power of the energy storage battery to the total capacity; is the state of charge of the energy storage at time; is the charging efficiency; is the discharging efficiency; is the charging power at time; is the discharging power at time; is the rated capacity of the energy storage; is the time interval; SOC Safety Range:

[0045] Charge and Discharge Power Limit:

[0046] (3) EV Constraint SOC Dynamic Equation:

[0047] Wherein, is the state of charge of the electric vehicle at time, representing the proportion of the current remaining power of the energy storage battery to the total capacity; is the state of charge of the electric vehicle at time; is the charging efficiency of the electric vehicle; is the discharging efficiency of the electric vehicle; is the charging power of the electric vehicle at time; is the discharging power of the electric vehicle at time; is the battery capacity of the electric vehicle; is the time interval; SOC Safety Range:

[0048] Charge and Discharge Power Limit:

[0049]

[0050] Wherein, is the upper limit of the charging power of the EV charging pile, is the upper limit of the discharge power of the EV charging pile; User travel demand:

[0051] Among them, is the off-grid time of the electric vehicle, is the minimum off-grid SOC required by the user.

[0052] (4) Photovoltaic output constraint

[0053]

[0054] Among them, is the maximum theoretical photovoltaic output (kW), calculated from irradiance and temperature.

[0055] (5) Grid interaction power constraint

[0056] Among them, is the maximum power selling capacity, is the maximum power purchasing capacity.

[0057] Step 3, online update the generation parameters of the virtual power plant energy and operation reserve optimization model constructed in Step 2 through RLS algorithm parameter identification; Preferably, the said Step 3 includes: Step 3.1, based on the multi-dimensional data related to the photovoltaic equipment collected in Step 1, dynamically correct the photovoltaic output model by using the recursive least squares method (RLS), and the photovoltaic output model is expressed as:

[0058] Among them, is the photovoltaic power output, is the predicted photovoltaic power, is the overflow photovoltaic power.

[0059] More preferably, dynamically correcting the photovoltaic output model by using the recursive least squares method (RLS) includes: Initialize the parameter estimation value and the initial covariance matrix ; At each time step , obtain new measurement data (actual photovoltaic output) and input data such as variables affecting photovoltaic output like solar irradiance, temperature, etc.), and calculate the gain matrix according to the RLS formula:

[0060] Among them, is the gain matrix at the -th time step, which is used to adjust the amplitude of parameter update; is the covariance matrix at the -th time step, which reflects the uncertainty of parameter estimation; is the input data vector at the -th time step, which contains various factors affecting the photovoltaic output; is the forgetting factor, and its value range is , which is used to adjust the forgetting speed of the algorithm for historical data; Update the parameter estimation value and the covariance matrix. The updated parameter estimation value is:

[0061] Among them, is the updated parameter estimation value at the -th time step; is the parameter estimation value at the -th time step; The updated covariance matrix is:

[0062] Among them, is the updated covariance matrix at the -th time step; is the identity matrix.

[0063] It should be noted that considering the deviation between the actual photovoltaic output and the predicted value due to factors such as dust occlusion and weather changes, the present invention dynamically corrects the photovoltaic output model through the recursive least squares method, making the predicted value of the photovoltaic output model closer to the actual value, improving the accuracy of the model, and providing a more reliable basis for further updating power generation parameters such as the efficiency of the photovoltaic inverter in step 3.

[0064] Step 3.2, according to the relevant data of the EV charging pile and the energy storage device collected in step 1, use the RLS algorithm to online update the energy storage SOC constraint and the EV frequency modulation time constant.

[0065] It should be noted that the RLS update trigger conditions for the photovoltaic output model include prediction error trigger and environmental mutation trigger. The prediction error threshold can be set to 5%. Exemplarily, the environmental mutation trigger includes the irradiance mutation rate exceeding the threshold. The RLS update trigger conditions for the energy storage SOC constraint include SOC trajectory deviation trigger and charge-discharge rate anomaly trigger. Exemplarily, the SOC trajectory deviation trigger includes the absolute error of the actual SOC deviating from the predicted trajectory exceeding 3%. The RLS update trigger conditions for the EV frequency modulation time constant include response delay trigger and power fluctuation trigger. Exemplarily, the response delay trigger includes the absolute error between the actual frequency modulation response time and the model value exceeding 0.2 s. The priority of the RLS algorithm for online updating the power generation parameters of the virtual power plant energy and operation reserve optimization model is photovoltaic update > energy storage update > EV update.

[0066] Step 4: Considering the uncertainties in the operation of the virtual power plant, construct a hierarchical time-series coordination framework to dynamically adjust the virtual power plant energy and operation reserve optimization model. Preferably, the hierarchical time-series coordination framework includes a day-ahead layer, an intra-day layer, and a real-time layer. More preferably, in the day-ahead layer, based on relevant historical data and prediction models, generate a photovoltaic-EV joint scenario, screen out extremely high-risk scenarios for injection into the optimization model, and quantify the extreme risks through conditional value at risk (CVaR) constraints to formulate a benchmark scheduling plan. Specifically, the historical data includes photovoltaic output, EV behavior, and historical data of the power grid and the market. More specifically, the power grid and market data includes time-of-use electricity price historical records, load demand data, and extreme event records. The generation of the photovoltaic-EV joint scenario based on historical data and prediction models and the screening of extremely high-risk scenarios include: Use the improved K-means algorithm to perform clustering analysis on the historical photovoltaic output data and EV behavior data collected in Step 1, and use the photovoltaic output and EV available capacity as feature vectors to generate groups of photovoltaic-EV joint scenarios:

[0067] where is the photovoltaic output at time and is the EV available capacity at time Calculate the net load deviation of each scenario, and screen out high-risk scenarios, which are calculated by the following formula:

[0068] where is the th scenario at The net load deviation at a moment, which is used to measure the imbalance degree of power supply and demand in this scenario; is the load demand at a moment; Select the top scenarios with the largest deviation as extremely high-risk scenarios and inject them into the optimization model. Exemplarily, 10% can be selected as the value of.

[0069] Specifically, the quantification of extreme risks through conditional value at risk (CVaR) constraints and the formulation of a benchmark scheduling plan include: Introduce the conditional value at risk (CVaR) constraint into the objective function of the virtual power plant energy and operating reserve optimization model, and minimize the conditional value at risk. The CvaR constraint formula is:

[0070] Among them, is the quantile threshold, representing the value at risk (CvaR); is the confidence level, with a range from 0 to 1. Exemplarily, the present invention selects the confidence level as 95%; is the expectation operator; is the benchmark loss function; Add the CvaR constraint to the benchmark loss function to generate a benchmark scheduling plan:

[0071] Further preferably, in the intraday layer, based on more recent prediction data, the scheduling plan is dynamically adjusted through rolling optimization; Specifically, the prediction data is updated every 15 minutes, and the scheduling plan is dynamically corrected to cope with real-time fluctuations. The formula for the rolling optimization time window is:

[0072] Among them, is the rolling window length. Exemplarily, it is set to 4 hours; is the current time point; On the basis of the day-ahead layer objective function, add a real-time correction term:

[0073] Among them, is the adjustment cost coefficient, with the unit of yuan / kW; is the penalty term for the adjustment amount of the scheduling plan.

[0074] Specifically, in the intraday layer, it also includes updating the prediction error distribution parameters through Bayesian online learning based on the photovoltaic output prediction error data of the past few hours. The photovoltaic output prediction error calculation formula is:

[0075] wherein, is the prediction error, is the photovoltaic output at time ; the prediction error follows a Gaussian distribution ; The updating of the prediction error distribution parameters through Bayesian learning includes:

[0076]

[0077] wherein, and are respectively the mean and variance (prior) of the historical error distribution, and are respectively the mean and variance (likelihood) of the real-time error within the current time window, and are the updated posterior distribution parameters.

[0078] Further preferably, in the real-time layer, the photovoltaic prediction error distribution is corrected in real time through high-frequency Bayesian learning. Combining the grid frequency regulation demand and the real-time error distribution, the spare capacity calling strategy is dynamically adjusted, and according to the time-of-use electricity price and the real-time load demand, the charging and discharging priorities of charging piles and energy storage devices are optimized; Specifically, the real-time correction of the photovoltaic prediction error distribution through high-frequency Bayesian learning includes: Bayesian learning is performed using the instantaneous error data of an extremely short time window to obtain the real-time error distribution, and the photovoltaic prediction is corrected in real time. The extremely short time window is within 1 minute.

[0079] It should be noted that both the real-time layer and the intraday layer adopt Bayesian online learning technology to correct the photovoltaic output prediction error distribution. The difference lies in the goal, data time window, update frequency, and parameter adjustment logic. The goal of Bayesian learning in the intraday layer is to periodically correct the prediction model and optimize the subsequent rolling scheduling plan; its data time window uses the cumulative error data of the past few hours (such as 4 hours) to reflect the prediction deviation trend of a longer time period; its update frequency is low-frequency update, synchronized with the rolling optimization period; its parameter adjustment logic is that the prior distribution is adjusted based on the posterior distribution parameters of the previous rolling optimization, and the likelihood function is adjusted based on the statistics of the error data within the current time window. The goal of Bayesian learning in the real-time layer is to instantaneously respond to real-time fluctuations; its data time window uses the instantaneous error data of an extremely short time window; its update frequency is high-frequency update (per second - per minute), synchronized with the real-time control period; its parameter adjustment logic is that the prior distribution is adjusted based on the posterior distribution parameters of the previous real-time update, and the likelihood function is adjusted based on the statistics of the latest instantaneous error.

[0080] In a further preferred embodiment, the parameters updated in the real-time layer can be used as the prior input for the optimization of the next cycle in the intra-day layer, forming a data closed-loop.

[0081] Specifically, in combination with the power grid frequency regulation requirements and the real-time error distribution, the dynamic adjustment of the reserve capacity calling strategy includes: Determine the power grid frequency deviation according to the power grid frequency regulation requirements for calculating the reserve power demand:

[0082] Wherein, is the power grid frequency regulation deviation, with the unit of Hz; is the reserve power demand, with the unit of MW; is the frequency regulation demand, determined according to the power grid rules, with the unit of MW / Hz; The calculation formula for the reserve capacity allocation constraint is:

[0083] Wherein, is the reserve discharge power of the th building charging pile, with the unit of kW; is the reserve discharge power of the energy storage device.

[0084] Specifically, the optimization of the charging and discharging priorities of the building charging piles and the energy storage devices according to the time-of-use electricity price and the real-time load demand includes: (1) Electricity price trigger rule: Peak hours: Priority of building charging piles: discharging > charging; Priority of energy storage devices: discharging > charging; Valley hours: Priority of building charging piles: charging > discharging; Priority of energy storage devices: charging > discharging.

[0085] (2) Load demand response rule: If the real-time load : Call the energy storage device to discharge to reduce the power purchase demand; If the real-time load : Give priority to consuming the surplus power of photovoltaic to reduce the abandonment of light.

[0086] It should be noted that in view of the uncertainty problems existing in the operation of virtual power plants in the prior art, the present invention proposes a hierarchical time-sequence coordination framework, which realizes the optimal balance of economy, reliability and flexibility of virtual power plants in an uncertain environment through a hierarchical optimization strategy, providing core regulation capabilities for new power systems.

[0087] Specifically, the day-ahead layer in the hierarchical time-series collaborative framework is used for long-term risk perception and benchmark plan formulation, the intra-day layer is used for short-term rolling correction and prediction accuracy improvement, and the real-time layer is used for instantaneous response and dynamic resource allocation; the error parameters of the real-time layer are fed back to the intra-day layer, and the correction model of the intra-day layer is transmitted to the day-ahead layer, forming a global optimization closed-loop, realizing data closed-loop feedback, from day-ahead risk perception to real-time second-level response, covering the entire process of "prediction-optimization-control" of the virtual power plant.

[0088] Step 5: Generate a control command according to the virtual power plant energy and operating reserve optimization model dynamically adjusted by the hierarchical time-series collaborative framework, and send it to each building distributed resource device to control the operating state and power output of the device, and monitor the operating state of the virtual power plant in real time, and adjust the virtual power plant energy and operating reserve optimization model according to the feedback information of the actual operating data.

[0089] In Embodiment 2 of the present invention, a virtual power plant regulation system based on dynamic parameter identification and hierarchical collaboration is provided. Based on the virtual power plant regulation method based on dynamic parameter identification and hierarchical collaboration provided in Embodiment 1, the system includes: A data acquisition and integration module for acquiring multi-dimensional data of building distributed resources and integrating them through federated learning; a resource aggregation and modeling module for aggregating building distributed resources according to power generation characteristics and economy, generating external characteristic parameters of the virtual power plant, and constructing an energy and operating reserve optimization model; a parameter update and dynamic adjustment module for online updating power generation parameters through the RLS algorithm and dynamically adjusting the optimization model in combination with the hierarchical time-series collaborative framework; an optimization regulation and state monitoring module for generating a control command and sending it to the building distributed resource device to control its operating state and power output, and at the same time monitoring the operating state of the virtual power plant in real time and adjusting the optimization model according to the feedback information.

[0090] The present disclosure may be a system, method, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions thereon for causing a processor to implement various aspects of the present disclosure.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: the specific implementation manners of the present invention can still be modified or equivalently replaced, and any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the protection scope of the claims of the present invention.

Claims

1. A virtual power plant control method based on dynamic parameter identification and hierarchical collaboration, characterized in that: The steps include: Collect multi-dimensional data of distributed building resources, including rooftop photovoltaic equipment, EV charging piles, and energy storage equipment, and integrate their multi-dimensional data through federated learning; Aggregate building distributed resources according to their power generation characteristics and power generation economy, generate virtual power plant external characteristic parameters and build a virtual power plant energy and operation reserve optimization model; update the power generation parameters of the virtual power plant energy and operation reserve optimization model online through the RLS algorithm, consider the uncertainty of virtual power plant operation, build a hierarchical time series coordination framework, and dynamically adjust the virtual power plant energy and operation reserve optimization model in combination with the real-time power generation parameters updated by the RLS algorithm; According to the virtual power plant energy and operating reserve optimization model dynamically adjusted according to the hierarchical timing coordination framework, control instructions are generated and sent to the distributed resource equipment in each building to control the operating status and power output of the equipment, and the operating status of the virtual power plant is monitored in real time. The virtual power plant energy and operating reserve optimization model is adjusted according to the feedback information of actual operating data.

2. A virtual power plant control method based on dynamic parameter identification and hierarchical collaboration according to claim 1, characterized in that: The method of aggregating building distributed resources according to their power generation characteristics and power generation economy, generating virtual power plant external characteristic parameters and constructing a virtual power plant energy and operation reserve optimization model includes: Each device is considered as a separate cluster, including rooftop photovoltaic devices, EV charging stations, and energy storage devices; Calculate the degree of consistency between the power generation characteristics of different devices based on the power generation characteristic curve of each device; Aggregate the devices based on the matching degree of power generation characteristics between every two devices to generate external characteristic parameters of the virtual power plant; Based on the equipment aggregation results and the external characteristic parameters of the virtual power plant, a virtual power plant energy and operation reserve optimization model is constructed, including the benchmark loss function, power balance constraints, energy storage system constraints, EV constraints, photovoltaic output constraints and grid interaction power constraints. The benchmark loss function is: in, is the grid interaction cost, For the The operating cost of an EV charging station, Penalty cost for abandoning light.

3. A virtual power plant control method based on dynamic parameter identification and hierarchical collaboration according to claim 2, characterized in that: The external characteristic parameters of the virtual power plant include rated power, economic power lower limit, economic power upper limit, unit control error dead zone, upward climbing rate, downward climbing rate, secondary frequency regulation upward limit, secondary frequency regulation downward limit, minimum time interval between unit shutdown and startup, and cost-capacity function.

4. A virtual power plant control method based on dynamic parameter identification and hierarchical collaboration according to claim 3, characterized in that: The online updating of the power generation parameters of the virtual power plant energy and operation reserve optimization model by the RLS algorithm includes: Based on multi-dimensional data related to photovoltaic equipment, the RLS algorithm is used to dynamically correct the photovoltaic output prediction model and update the photovoltaic inverter efficiency online; Based on the multi-dimensional data related to EV charging piles and energy storage devices, the RLS algorithm is used to update the energy storage SOC constraint and EV frequency modulation time constant online.

5. A virtual power plant control method based on dynamic parameter identification and hierarchical collaboration according to claim 4, characterized in that: Considering the uncertainty of virtual power plant operation, building a hierarchical timing coordination framework includes: The hierarchical time series coordination framework includes a day-ahead layer, an intra-day layer, and a real-time layer; In the day-ahead layer, PV-EV joint scenarios are generated based on PV output, EV behavior, historical data of the grid and the market, and the PV output forecast model. High-risk scenarios that meet the preset risk threshold are selected and injected into the optimization model. The economic losses of the high-risk scenarios are quantified through conditional risk value constraints, and a benchmark dispatch plan is formulated. In the intra-day layer, based on the periodically collected PV output forecast data, the dispatch plan is dynamically adjusted through rolling optimization, and the PV output forecast error distribution parameters are updated through Bayesian online learning; In the real-time layer, based on the actual measured PV output data updated in seconds and the grid frequency regulation demand signal, event-triggered Bayesian online learning is used to correct the PV output forecast error distribution parameters in real time, dynamically adjust the backup capacity call strategy, and optimize the charging and discharging priority of charging piles and energy storage equipment according to time-of-use electricity prices and real-time load demand.

6. A virtual power plant control method based on dynamic parameter identification and hierarchical collaboration according to claim 5, characterized in that: The day-ahead layer includes: Based on the historical data of PV output, EV behavior, power grid and market and the PV output prediction model, a PV-EV joint scenario is generated. High-risk scenarios that meet the preset risk threshold are selected and injected into the optimization model. Cluster analysis is performed on historical PV output data and EV behavior data. PV output and EV available capacity are used as feature vectors to generate Combined PV-EV scenarios; Calculate the net load deviation of each scenario and filter out high-risk scenarios that meet the preset risk threshold; The conditional risk value constraint is introduced into the objective function of the virtual power plant energy and operating reserve optimization model to minimize the conditional risk value. The conditional risk value constraint formula is: Add the conditional risk value constraint to the benchmark loss function to generate a benchmark scheduling plan: in, is the quantile threshold; is the confidence level; is the expectation operator; is the baseline loss function.

7. A virtual power plant control method based on dynamic parameter identification and hierarchical collaboration according to claim 6, characterized in that: The intraday layer includes: Based on the periodically collected photovoltaic output forecast data, the scheduling plan is dynamically adjusted through rolling optimization. The rolling optimization time window is ; Add a real-time correction term based on the baseline objective function: in, To adjust the cost factor; It is the penalty item of scheduling plan adjustment; And the photovoltaic output prediction error distribution parameters are updated through Bayesian online learning.

8. A virtual power plant control method based on dynamic parameter identification and hierarchical collaboration according to claim 7, characterized in that: The real-time layer includes: Based on the measured PV output data updated in seconds and the grid frequency modulation demand signal, the PV output forecast error distribution parameters are corrected in real time through event-triggered Bayesian online learning, and the reserve capacity call strategy is dynamically adjusted, including: Determine the grid frequency deviation based on the grid frequency regulation requirements and use it to calculate the standby power requirements: in, It is the frequency deviation of the power grid; For standby power requirements; For frequency modulation needs; The calculation formula of the spare capacity allocation constraint is: in, For the The standby discharge power of each building charging pile; It is the standby discharge power of the energy storage device.

9. A virtual power plant control method based on dynamic parameter identification and hierarchical collaboration according to claim 8, characterized in that: The optimization of charging and discharging priorities of building charging piles and energy storage equipment according to time-of-use electricity prices and real-time load demand includes: Receive time-of-use electricity price signals. During peak hours, building charging piles and energy storage equipment are discharged first, and during off-peak hours, building charging piles and energy storage equipment are charged first. Monitor the real-time load demand and the predicted load demand. If the real-time load demand is greater than the predicted load demand, call on the energy storage equipment to discharge and reduce the demand for electricity purchases. If the real-time load demand is less than the predicted load demand, give priority to absorbing the surplus photovoltaic power and reduce the abandonment of light.

10. A virtual power plant control system based on dynamic parameter identification and hierarchical collaboration, based on a virtual power plant control method based on dynamic parameter identification and hierarchical collaboration according to any one of claims 1 to 9, characterized in that: The system includes: The data collection and integration module is used to collect multi-dimensional data of building distributed resources and integrate them through federated learning. The resource aggregation and modeling module is used to aggregate building distributed resources according to power generation characteristics and economic efficiency, generate external characteristic parameters of virtual power plants and build energy and operation reserve optimization models. The parameter update and dynamic adjustment module is used to update power generation parameters online through the RLS algorithm, and dynamically adjust the optimization model in combination with the hierarchical timing collaborative framework. The optimization control and status monitoring module is used to generate control instructions and send them to building distributed resource equipment to control their operating status and power output, while monitoring the operating status of the virtual power plant in real time and adjusting the optimization model according to feedback information.

Images