Cooperative scheduling method and system for virtual power plants

Abstract

This invention relates to the field of intelligent power management technology, and discloses a collaborative scheduling method and system for virtual power plants. The method includes the following steps: S1, collecting real-time data from each distributed power source, load, and energy storage system within the virtual power plant, and performing ultra-short-term forecasting to obtain forecast parameters; S2, dynamically calculating the dynamic operating boundary based on the real-time status of the energy storage system; S3, generating a pre-schedule plan through collaborative decision-making using a multi-objective fuzzy satisfaction function one day before the current operating day; S4, updating the boundary and forecast parameters on a rolling basis within the current operating day, and generating real-time scheduling instructions through model prediction; S5, monitoring the deviation between the actual output of each resource and the real-time scheduling instructions in real time, and initiating a collaborative deviation compensation mechanism for power balancing when the deviation exceeds a threshold. This invention enables refined collaborative scheduling of different types of distributed resources in complex environments with high uncertainty.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent power management, in particular to a collaborative scheduling method and system for a virtual power plant. BACKGROUND

[0002] With the deepening of the "double carbon" strategy, the proportion of new energy represented by wind power and photovoltaic power in the power system continues to increase, and the power grid structure is undergoing a transition from traditional centralized to clean and low-carbon, safe and efficient new power systems. Under this background, as an intelligent management system that aggregates distributed resources such as distributed power sources, energy storage systems, controllable loads, electric vehicles, and participates in the coordinated optimization of power grid operation and power market through advanced information communication technology, the virtual power plant has become a key technical path to improve the ability of the power grid to absorb renewable energy and ensure the balance of power supply and demand.

[0003] Currently, virtual power plant technology has entered the large-scale demonstration stage from the concept verification stage in China. The current mainstream mode mainly focuses on the preliminary aggregation and instruction response of resources, and most of them build software platforms with basic functions such as resource monitoring, load regulation, and demand response in terms of technical architecture, and preliminarily realize the "observable and measurable" of distributed resources. In terms of scheduling strategy, existing systems usually use methods based on price signals or direct load control, and according to the adjustment instructions issued by the power grid, the aggregated resources are simply started or stopped or power is adjusted. However, these modes still rely more on administrative means or fixed price incentive mechanisms, the target of scheduling optimization is relatively single, mainly focusing on meeting the instantaneous power balance demand of the power grid, and most of them still follow the "top-down" instruction scheduling logic, regarding the heterogeneous resources within the aggregate as a whole that can be uniformly scheduled, lacking detailed consideration of the individual differences and autonomy of the resources. There is no systematic solution to how to coordinate the differentiated interest demands of different subjects within the resources and how to deal with the severe fluctuations of both sides of the source and load under high proportion of new energy access.

[0004] Further, although the virtual power plant has developed rapidly, it still faces several key technical difficulties in actual operation, which restricts the maximization of its effectiveness:

[0005] First, it is difficult to achieve multi-objective collaborative optimization under complex operating environment. Virtual power plants aggregate photovoltaic, energy storage, flexible load and other resources, whose response speed, regulation capacity, cost structure and interest demands are different. The scheduling needs to meet the goals of power grid peak shaving, economy, high proportion of renewable energy consumption, user comfort and equipment life extension, etc. These goals often restrict each other, for example, frequent charging and discharging of energy storage for economic benefits may damage its life. Simple single-objective optimization or traditional methods such as weighted sum are difficult to find a fair and efficient optimal solution among all objectives, especially unable to handle the dynamic trade-off between objectives.

[0006] Secondly, the output of new energy has high uncertainty, which is easy to cause scheduling deviation. The output of photovoltaic is intermittent and volatile due to weather influence, and the load demand also has randomness. The existing prediction technology is difficult to be completely accurate, so the current plan based on point prediction often deviates from the actual situation. Especially in the scenario of high proportion of new energy, the uncertainty is amplified, and the system is difficult to balance the scheduling.

[0007] Thirdly, the fine consideration of the whole life cycle loss of the energy storage system is insufficient. The existing scheduling model usually simplifies the energy storage as an ideal energy body, ignores its charging and discharging efficiency attenuation, cycle life loss and other dynamic characteristics, and fails to include the constraint of battery health state on the adjustable capacity into the optimization boundary. This not only may cause the scheduling instruction to exceed the safe operation range of the energy storage device, aggravate its aging, but also cause distortion in economic evaluation due to the failure to quantify the attenuation cost, so as to realize the real balance between short-term income and long-term asset value. SUMMARY

[0008] The present application aims to provide a collaborative scheduling method and system for a virtual power plant, which can realize fine collaborative scheduling of different types of distributed resources in a complex environment with high uncertainty, and can improve the operation economy and efficiency of the virtual power plant while ensuring the reliability of the power grid.

[0009] To achieve the above-mentioned purpose, the present application provides the following basic scheme.

[0010] Scheme one

[0011] The collaborative scheduling method for a virtual power plant comprises the following steps:

[0012] S1, real-time acquisition of the output data of each distributed power source in the virtual power plant, the power consumption data of each load, and the state data of the energy storage system, and based on this, super-short-term prediction is carried out to obtain prediction parameters for the next N hours, the prediction parameters at least including prediction values and corresponding uncertainty intervals; N is an integer between 1 and 6;

[0013] S2, based on the real-time state of the energy storage system, dynamically calculating its dynamic operation boundary considering life attenuation, the dynamic operation boundary including the safe charging and discharging boundary based on the battery health state;

[0014] S3, based on the prediction parameters, establishing and solving a first optimization model to maximize the overall operation satisfaction of the virtual power plant as the target to generate a pre-scheduling plan for the next 24 hours on the day before the current operation day; wherein the overall operation satisfaction is the collaborative decision result of the fuzzy satisfaction function based on multiple targets; the multiple targets include the grid demand matching degree, the total operation cost, the energy storage attenuation cost and the load comfort degree;

[0015] S4, within the current running day, taking the pre-scheduling plan as a reference, the following steps are executed at a fixed time interval:

[0016] S41, according to the latest running data, updating the dynamic running boundary and prediction parameters of each resource;

[0017] S42, under the constraint of the updated boundary, a second optimization model with shortened optimization time domain is established and solved, real-time scheduling instructions of the current period are generated, and are issued to each resource for execution;

[0018] S5, real-time monitoring of the deviation of the actual output of each resource from the real-time scheduling instructions, when the deviation exceeds a threshold, starting a collaborative deviation compensation mechanism based on the real-time adjustable capacity of each resource to balance power.

[0019] Scheme two

[0020] A collaborative scheduling system for a virtual power plant has a computer program stored thereon, and the computer program, when executed by a processor, implements the steps of the collaborative scheduling method for the virtual power plant according to scheme one.

[0021] The working principle and advantages of the present application are that:

[0022] The collaborative scheduling method and system for a virtual power plant can achieve fine collaborative scheduling of different types of distributed resources in a complex environment with high uncertainty, and can improve the operation economy and efficiency of the virtual power plant while ensuring the reliability of the power grid. The key points are:

[0023] First, the present scheme breaks through the limitations of traditional scheduling and can achieve fine collaborative scheduling; and the uncertainty of new energy output and the whole life cycle loss of the energy storage system are finely considered and processed.

[0024] Traditional scheduling methods often regard the virtual power plant as a black box or a single entity, and use top-down command control, ignoring the heterogeneity and autonomy of internal resources. In view of this, the present scheme specially designs the concept of "dynamic running boundary", models the energy storage, load and other resources as intelligent agents with their own states and constraints, and can accurately model the resource characteristics.

[0025] Particularly for energy storage systems, the scheme does not simplify it as an ideal energy container, but establishes a safe charging and discharging boundary linked with real-time health status, and quantifies its dynamic attenuation cost, so that the scheduling model can actively avoid the operation strategy harmful to the battery life, not only can improve the economy of scheduling decision (explicitly long-term asset depreciation cost and into the optimization goal), but also helps to enhance the physical feasibility and operation safety of the system, prevent the risk of equipment failure caused by excessive call. In addition, through the rolling optimization based on model predictive control and dynamic boundary negotiation mechanism, the scheme can realize real-time fine tuning under the current plan macro guidance, effectively smooth the fluctuations caused by renewable energy and load prediction error, has better robustness and self-adaptation ability.

[0026] Secondly, the scheme can realize multi-objective collaborative optimization in complex operation environment.

[0027] Firstly, the scheme specially applies the "fuzzy cooperative game" theory to the multi-objective optimization of virtual power plant, aiming to maximize the overall "satisfaction" rather than a single economic indicator, which is an important change in concept. Most of the existing technologies use weighted summation to handle multiple objectives, but the weight setting often depends on experience and can easily lead to poor performance of a certain target and lower overall efficiency. The scheme seeks Pareto optimality through the geometric mean of the satisfaction function, requiring all sub-targets to achieve a certain level of satisfaction. This design of collaborative decision-making mechanism overcomes the inherent defects of traditional methods, and its application in virtual power plant dispatching field has a certain originality.

[0028] Secondly, the scheme embeds the long-term loss problem of energy storage attenuation in the mathematical model of short-term operation optimization of power system through dynamic cost factor, and combines it with model predictive control framework to form a closed-loop optimization system considering life loss, which can effectively coordinate instantaneous decision and long-term consequences.

[0029] Finally, through the closed loop from collaborative deviation compensation to benefit distribution based on contribution, the former adjusts the output of each participating subject (such as energy storage, distributed power supply, controllable load) when there is a deviation between actual output and plan, to ensure that the overall collaborative effect meets the expectation; the latter distributes the benefits according to the actual contribution (such as adjustment amount, response speed, accuracy, etc.) of each subject after the collaboration is completed, reflecting "more pay, more reward", the combination of the two can link technical scheduling with market incentive mechanism, ensuring the stability of the collaborative alliance and the enthusiasm of the participating subjects. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 Method flowchart for embodiment one of the collaborative scheduling method and system for virtual power plant of the present application. DETAILED DESCRIPTION

[0031] The application will be described in further detail below with reference to the embodiments:

[0032] Embodiment One

[0033] The embodiment is basically as shown in the accompanying drawings: a collaborative scheduling method for a virtual power plant, comprising the following steps: Figure 1

[0034] S1, real-time acquisition of output data of each distributed power source in the virtual power plant, power consumption data of each load, and state data of the energy storage system, and based thereon, super-short-term prediction is performed to obtain prediction parameters for the next N hours, the prediction parameters at least including predicted values and corresponding uncertainty intervals; N is an integer between 1 and 6. In this embodiment, N is 4, i.e. the period of super-short-term prediction is 4 hours. In actual application, the value of N can be adjusted according to actual prediction requirements.

[0035] In this embodiment, real-time data is collected through terminal devices such as smart meters, photovoltaic inverters, energy storage converters (PCS), battery management systems (BMS) installed on the user side, and the data collection frequency can be set as needed, for example, for fast-changing quantities such as power, voltage, and current, the collection frequency is set to once every 15 seconds; for slow-changing quantities such as SOC and temperature, it can be set to once every 1-5 minutes. The collected raw data needs to be processed through data cleaning (such as removing outliers and filling missing values), format standardization, and can be stored in a time series database (such as InfluxDB) or a relational database (such as MySQL) for subsequent calling.

[0036] The output data of the distributed power source includes real-time active / reactive power, voltage, current, and inverter state of each distributed photovoltaic power station. The power consumption data of each load includes the total load power of each industrial and commercial user and electric vehicle charging pile, the current state and adjustable range of interruptible load. The state data of the energy storage system includes real-time state of charge (SOC), battery voltage, current, temperature, internal resistance, state of health (SOH) estimation value, and cumulative cycle number.

[0037] Specifically, the super-short-term prediction is realized by using an Attention-BiLSTM prediction model based on an attention mechanism, and the model is optimized by using a quantum genetic algorithm to obtain the prediction parameters.

[0038] ​In this embodiment, prediction models are trained separately for load and distributed power generation (photovoltaics). Taking photovoltaic prediction as an example, the model input features include: historical 72-hour photovoltaic output sequence, historical meteorological data (irradiance, temperature, humidity), numerical weather prediction (NWP) data for the next 24 hours, and time features (hourly, daily type). Historical data from the past year (divided into training and validation datasets in an 8:2 ratio) is used for model training. A quantum genetic algorithm (QGA) is employed to optimize the hyperparameters of the BiLSTM network (such as the number of hidden layer nodes, learning rate, and dropout rate) to minimize the root mean square error (RMSE) of the prediction, thereby obtaining the optimal model parameters.

[0039] In ultra-short-term forecasting, a rolling forecast for the next 4 hours is automatically triggered every 15 minutes. The latest real-time data and NWP data are input into a pre-trained LSTM model, which outputs a sequence of power prediction values ​​for the next 16 time points (at 15-minute intervals). .

[0040] The uncertainty interval in the prediction parameters is calculated using the Conformal Prediction method, and robust optimization constraints are provided for the first and second optimization models. This process includes the following sub-steps:

[0041] S101, Using the validation dataset, calculate the prediction error of the prediction model on each sample. . Let i be the actual value of the i-th sample. Let be the predicted value of the prediction model for the i-th sample. It reflects the degree of deviation between the model's predicted values ​​and the actual values.

[0042] S102, the absolute value of the prediction error for all samples. Arrange them in ascending order so that an appropriate error value can be selected to determine the radius of uncertainty based on the sorting results.

[0043] S103, for a new prediction point t, select the sorted point in order. The absolute value of the positional error is used as the radius of uncertainty. Where n is the number of samples in the validation dataset. It is the significance level (e.g., let's assume...) (This indicates a 90% confidence level). This indicates rounding up. The absolute value of the error selected at this position can ensure that, with a certain level of confidence, the actual value of the new predicted point falls within the subsequently determined interval.

[0044] S104, the uncertainty interval for generating this prediction point is... .

[0045] This is the prediction value of the prediction model for the new prediction point t.

[0046] Ultimately, the obtained prediction parameters consist of two parts: the predicted value sequence. And the corresponding uncertainty interval sequence. The upper and lower limits of the uncertainty interval can then be used as part of the constraints of the optimization problem (i.e., to provide robust optimization constraints for the first and second optimization models), forcing the scheduling scheme to remain feasible and safe under any possible adverse scenarios, thereby enhancing the virtual power plant's ability to cope with uncertainty.

[0047] S2, based on the real-time state of the energy storage system, dynamically calculate its dynamic operating boundary considering lifetime degradation, the dynamic operating boundary including the safe charge and discharge boundary based on the battery health state.

[0048] The safe charge / discharge boundary is defined as follows: , is a function of the current state of charge (SOC), cumulative equivalent cycle count, average depth of discharge, and battery temperature of the energy storage system, and is calculated online using a decay model.

[0049] Specifically, the safe charge / discharge boundaries are not fixed, but dynamically adjusted according to the real-time state of the energy storage. In this embodiment, the boundaries can be set according to the following rules: the basic boundary is set to [20%, 90%]. If the average battery temperature remains above 35°C, both the upper and lower boundaries are reduced by 5%, becoming [25%, 85%], to reduce high-rate charge / discharge and slow down degradation; if the battery health state (SOH) is below 80%, the basic boundary is further reduced to [30%, 80%] to extend the remaining service life.

[0050] When the actual SOC of the energy storage system exceeds the safe charge / discharge boundary, a higher degradation cost coefficient is assigned to it in the optimization model. In this embodiment, the coefficient is calculated using the following empirical model:

[0051] ;

[0052] Where I(t) is the current charging and discharging current, reflecting the intensity of charging and discharging; Rated capacity, used to normalize the current for easier unified analysis; CycleCount is the cumulative number of cycles, CycleLife is the rated cycle life; exp is an abbreviation for exponential function, representing exponential operation with the natural constant e as the base. Coefficient , , , The degradation cost coefficient is obtained by fitting data from accelerated battery aging tests. It is used to quantify the equivalent economic loss to battery life caused by a unit charge-discharge operation at the current moment.

[0053] S3, the day before the current operating day, based on the predicted parameters, with the goal of maximizing the overall operational satisfaction of the virtual power plant, establish and solve the first optimization model to generate a pre-schedule plan for the next 24 hours; wherein, the overall operational satisfaction is the result of collaborative decision-making based on a fuzzy satisfaction function of multiple objectives; the multiple objectives include grid demand matching degree, total operating cost, energy storage attenuation cost and load comfort.

[0054] The fuzzy satisfaction function is in the form of a weighted geometric average of multiple sub-objective satisfaction functions, i.e. ,in, Let the satisfaction level be that of the i-th sub-goal. The weights are assigned to x, which is the scheduling decision variable (such as energy storage charging and discharging power, load adjustment amount, etc.).

[0055] Specifically, in this embodiment, .

[0056] in, Cost satisfaction is inversely proportional to total operating costs; the lower the cost, the closer the satisfaction level is to 1.

[0057] The satisfaction level of matching power grid demand is inversely proportional to the deviation between the actual response volume and the dispatch instructions.

[0058] Satisfaction with energy storage degradation is inversely proportional to the total degradation cost.

[0059] The total attenuation cost equals ; The attenuation cost coefficient, It represents the absolute value of the battery's charging and discharging power, reflecting the intensity of charging and discharging.

[0060] The load comfort satisfaction is inversely proportional to the total load reduction and the degree of deviation from its comfort range.

[0061] The NSGA-II multi-objective genetic algorithm is used to solve the problem. The algorithm outputs a set of Pareto optimal solutions, from which the scheduler selects a compromise solution as the current plan, i.e., the pre-scheduling plan, based on the current strategy.

[0062] S4. Within the current operating day, referencing the pre-scheduled plan, execute the following steps on a rolling basis at fixed time intervals:

[0063] S41, based on the latest operational data, update the dynamic operational boundaries and prediction parameters of each resource;

[0064] S42, under the updated boundary constraints, establish and solve the second optimization model that shortens the optimization time domain, generate the real-time scheduling instructions for the current period, and issue them to each resource for execution.

[0065] The second optimization model employs a Model Predictive Control (MPC) framework, and the shortened optimization time domain is 2 to 6 hours. In this embodiment, the optimization time domain is the next 4 hours (16 points), meaning that during each optimization, the situation for the next 4 hours is analyzed and optimized based on the predicted data. The control time domain is 1 hour (4 points), meaning that the actual time range for executing scheduling decisions is the next 1 hour, with a focus on optimizing resource power settings within this time period.

[0066] Specifically, in this embodiment, the latest forecast parameters (such as load forecast, renewable energy output forecast, etc.) and the actual state of resources (especially the state of charge (SOC) of energy storage) are obtained every 15 minutes. Then, using the existing pre-scheduling plan as a reference trajectory, an optimization problem is solved again.

[0067] Furthermore, the optimization problem to be solved again is a simplified version of the first optimization model, which can be solved using a faster quadratic programming (QP) solver. The result obtained is the precise power setpoint for each resource (such as energy storage, distributed power sources, controllable loads, etc.) for the next hour, which is used to guide actual operation scheduling.

[0068] S5, monitor the deviation between the actual output of each resource and the real-time scheduling command in real time. When the deviation exceeds the threshold (e.g., 5%), initiate a collaborative deviation compensation mechanism based on the real-time adjustable capability of each resource to balance power.

[0069] The cooperative deviation compensation mechanism includes the following operations:

[0070] The real-time contribution capacity of each resource is calculated based on the power deficit or surplus and the real-time adjustable margin of each resource; wherein the real-time adjustable margin is calculated based on the difference between the current actual output of the resource and its maximum output, and the degree of proximity of the current state of the resource to the dynamic operating boundary.

[0071] Specifically, for each resource i, calculate its real-time adjustable margin. .in, It is the maximum output of resource i. It is the actual effort currently being put in. This is the rated output, and the formula reflects how much output adjustment space the resources have.

[0072] For energy storage, if its SOC (State of Charge) is close to the dynamic operating boundary (i.e., the safe charge / discharge boundary), its margin is reduced. This is because when energy storage is close to the boundary, continued charging and discharging may damage its lifespan or exceed the safe range, so its adjustable margin is reduced to avoid over-utilization.

[0073] Among them, the normalized distance D of the energy storage's SOC relative to the upper or lower limit of the safe charge / discharge boundary is used. When D is less than 0.1, it is determined that the boundary is approaching, and a reduction factor is applied. Its adjustability margin is reduced. The reduction factor is calculated using a sigmoid function. .

[0074] Based on the magnitude of the real-time contribution capability, the power value that needs to be compensated is proportionally allocated to one or more resources to share the burden.

[0075] Specifically, let the total deviation be... The compensation amount allocated to resource i is By prioritizing the use of resources with greater margins, we can utilize resource adjustment capabilities more efficiently and quickly eliminate deviations.

[0076] After completing a demand response or ancillary service, the contribution of each resource in the collaborative deviation compensation mechanism is calculated based on the actual compensation amount, response speed, and response accuracy, and the Shapley value method is used for benefit allocation.

[0077] Specifically, after a response event concludes, the system calculates a contribution score for each resource j that participated in the response. ; This represents the actual compensation amount for resource j. For the response speed of resource j, Let be the response accuracy of resource j; a, b, and c are the weights of the corresponding indicators. The actual compensation amount reflects the amount of resource output adjustment, the response delay reflects the timeliness of resource response, and the response accuracy represents the accuracy of resource execution instructions. Combining these three dimensions can comprehensively measure the contribution of resources.

[0078] The Shapley value method is used for profit distribution. The Shapley value method is a method used in cooperative game theory to fairly distribute the total profit of an alliance. Its core is to calculate the average marginal contribution of each participant across all possible combinations of cooperation. In practice, an approximate algorithm can be used to calculate the average marginal contribution of each resource across all possible combinations of participation. The profit allocated to resource j is... ,in, This represents the total benefit of the event. This method fairly reflects the actual value of each resource within the collaborative alliance, ensuring the enthusiasm of participating entities.

[0079] This embodiment also provides a collaborative scheduling system for virtual power plants, which stores a computer program. When the computer program is executed by a processor, it implements the steps of the collaborative scheduling method for virtual power plants described above.

[0080] This embodiment provides a collaborative scheduling method and system for virtual power plants, which can achieve fine-grained collaborative scheduling of different types of distributed resources in complex environments with high uncertainty. It can improve the operational economy and efficiency of virtual power plants while ensuring the reliability of the power grid.

[0081] In addition, to verify the application effect of this solution, a case study of power dispatching during the summer peak season in an industrial park in Xinjin District was selected for simulation analysis. The participating resources included: Photovoltaics: Rooftop photovoltaic system in the park, with a total capacity of 2MW; Energy Storage: A 1MW / 2MWh lithium iron phosphate battery energy storage system; Load: An injection molding workshop within the park (interruptible load, with a maximum power reduction of 500kW) and a central air conditioning system (flexible load, with a power reduction of 300kW).

[0082] The grid event is set as follows: On the previous day (the 14th), the virtual power plant platform received a demand response instruction issued by the power grid dispatch center, requesting a reduction of a total of 1000kW of load power between 12:00 and 14:00 on the 15th to alleviate the pressure on the regional power grid.

[0083] Step 1: Current collaborative optimization (afternoon of the 14th).

[0084] Based on the weather forecast (sunny, strong radiation), the photovoltaic output is predicted to reach a peak of 1.8MW at noon on the 15th. The load forecast shows that the base load at noon is 3.5MW.

[0085] The energy storage status is currently 70% SOC and 92% SOH, with a stable number of cycles recently. Based on this data, the dynamic safety boundary for energy storage during the midday period is calculated to be [25%, 85%], and a moderate degradation cost factor is generated.

[0086] The current optimization aims to "maximize overall satisfaction".

[0087] The optimized result (pre-schedule plan) is as follows: 12:00-14:00: Photovoltaic: Full power generation of 1.8MW, all of which is uploaded to the grid.

[0088] Energy storage: Starting at 12:00, discharge at 500kW power for 2 hours to reduce SOC from 70% to 50%.

[0089] This decision is made because the SOC is always far from the safety boundary, and the attenuation cost is controllable.

[0090] Load: 500kW reduction in injection molding workshop and 200kW reduction in central air conditioning.

[0091] Total reduction: 500kW (energy storage) + 500kW (injection molding) + 200kW (air conditioning) = 1200kW > 1000kW, leaving a 200kW margin to cope with uncertainties.

[0092] Step Two: Intraday Rolling Optimization and Real-Time Collaboration (Noon on the 15th)

[0093] 12:00: Plan initiated. Everything normal. 12:45: Emergency. Weather radar detected a temporarily formed cloud cluster moving towards the park. The forecasting model immediately updated its "forecast parameters," predicting that photovoltaic output would plummet from 1.8MW to 1.2MW within the next 30 minutes, disrupting the power balance (PV output decreased by 600kW, but the grid's 1000kW reduction directive still needs to be fulfilled).

[0094] The energy storage system currently has a state of charge (SOC) of 60% and an output power of 500kW. The injection molding workshop has already reduced its output by 500kW as planned. The air conditioning system has also been reduced by 200kW.

[0095] Calculations show that if energy storage is allowed to increase its output to make up for the photovoltaic deficit, its SOC will drop rapidly and reach the lower limit of the dynamic safety boundary (25%), at which point the degradation cost factor will increase sharply.

[0096] Initiate the collaborative deviation compensation mechanism. Send a negotiation request to the park's central air conditioning system: "Are you willing to add an additional 100kW load reduction for 25 minutes on top of the existing load? The platform will pay an additional incentive fee." The air conditioning system, based on its indoor temperature comfort model, determines that the operation is feasible and automatically confirms the request. It then issues a new dispatch instruction: Air conditioning: additionally reduce load by 100kW (total reduction reaches 300kW). Energy storage: maintain the original planned 500kW discharge power to avoid entering the high-degradation risk zone.

[0097] Total reduction: 500kW (energy storage) + 500kW (injection molding) + 300kW (air conditioning) = 1300kW. After the photovoltaic power is reduced by 600kW, the net on-grid power still meets the grid's requirement of a reduction of 1000kW.

[0098] In summary, in the event of unforeseen circumstances (such as deteriorating weather), this solution can ensure that scheduling targets are met through coordinated scheduling. Furthermore, in the face of emergencies, it does not rely on a single resource to bear the burden, but rather mobilizes load resources for coordinated compensation through a "negotiation mechanism" to achieve multi-party game theory. At the same time, through dynamic boundaries and attenuation cost factors, it successfully avoids energy storage operating under unhealthy conditions, achieving optimal life-cycle cost.

[0099] Example 2

[0100] The cooperative scheduling method for virtual power plants, based on Embodiment 1, further includes S6, the adaptive learning step:

[0101] Historical operating data, including prediction bias and energy storage attenuation model bias, are collected periodically, and the historical operating data is used to update and train the prediction model used to generate the prediction parameters and the attenuation model.

[0102] Specifically, the steps for collecting the prediction bias data include:

[0103] We collect ultra-short-term load forecasts and photovoltaic output forecasts every 15 minutes over the past week, along with their corresponding actual measurements. We also collect meteorological data (actual irradiance, temperature, etc.) at the corresponding forecast times. From this, we can calculate the absolute forecast error at each time point: |forecast value - actual value|, which serves as the forecast deviation data.

[0104] The steps for collecting deviation data of the energy storage attenuation model include:

[0105] Collect detailed records of each charge-discharge cycle of the energy storage system over the past week, including: initial SOC, final SOC, average charge / discharge power, and average temperature. Perform a complete capacity calibration of the energy storage system weekly (through full charge / discharge testing or online evaluation with specialized equipment) to obtain the actual measured capacity degradation value. Based on the expected capacity degradation value calculated by the degradation model (e.g., based on cumulative ampere-hour throughput multiplied by a degradation coefficient), the prediction deviation of the degradation model can be calculated as: |expected degradation value - measured degradation value|, which serves as the energy storage degradation model deviation data.

[0106] The collected raw historical operational data is cleaned to remove invalid data caused by communication interruptions or equipment failures. The data is then aligned by time and constructed into a sample format suitable for model training. For example, for a prediction model, each sample contains "input features (historical data, weather forecasts)" and "labels (actual values)".

[0107] In the update training, the prediction model is updated using either incremental learning or periodic full retraining strategies. The incremental learning strategy includes calculating the average prediction error (MAE) over the past week weekly. If this error exceeds a preset threshold for two consecutive weeks (e.g., 15% higher than the baseline error), an incremental learning iteration is immediately triggered. Incremental learning only uses new data from the most recent weeks to fine-tune the existing model parameters to adapt to short-term changes in data distribution, resulting in lower computational overhead.

[0108] The scheduled full retraining strategy includes automatically triggering a full retraining of the prediction model on the first weekend of each month. Using all available historical data from the past year (or two years) as the training set, the model is retrained to capture long-term seasonality and pattern changes, which incurs significant computational overhead.

[0109] The decay model is updated quarterly, and includes the following steps: summarizing all decay model prediction bias data for the current quarter, and using a Bayesian update method to update key parameters in the decay model (such as the coefficients in Example 1). , , The correction is then performed. Specifically, the existing parameters are considered as the prior distribution, and the measured deviation data for this quarter are used as observational evidence. The posterior distribution of the parameters is calculated using Bayes' theorem. The mean of the posterior distribution is the new corrected parameter.

[0110] Optionally, the attenuation model can also be updated by fitting, using "cumulative equivalent cycle count" or "cumulative charge and discharge energy" as independent variables and "measured capacity attenuation" as dependent variable, to refit the function form of the original attenuation model.

[0111] This embodiment provides a collaborative scheduling method and system for virtual power plants. Compared with Embodiment 1, it can periodically update and train the prediction model and attenuation model using historical operating data, thereby obtaining more accurate prediction and attenuation models. This enables more reasonable arrangements in scheduling decisions and equipment maintenance, reducing system failures and power outages caused by inaccurate predictions or incorrect estimations of equipment status, and thus improving the reliability of the entire power system.

[0112] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics of the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A collaborative scheduling method for a virtual power plant, characterized in that, The method comprises the following steps: S1, real-time acquisition of output data of each distributed power source in the virtual power plant, power consumption data of each load, and state data of the energy storage system, and based thereon, super-short-term prediction is performed to obtain prediction parameters for N hours in the future, the prediction parameters at least including predicted values and corresponding uncertainty intervals; N is an integer between 1 and 6; S2, based on the real-time state of the energy storage system, a dynamic operation boundary considering life attenuation is dynamically calculated, the dynamic operation boundary including a safe charging and discharging boundary based on a battery health state; The safe charging and discharging boundary is defined as: is a function of the current state of charge SOC, the cumulative equivalent cycle number, the average discharge depth and the battery temperature of the energy storage system, and is calculated online by a decay model. The safe charging and discharging boundary is dynamically adjusted according to the real-time state of the energy storage, and is set according to the following rules: the basic boundary is set to [20%, 90%], if the average temperature of the battery continuously exceeds 35℃, the upper and lower boundaries are both inwardly contracted by 5%, becoming [25%, 85%], so as to reduce high-rate charging and discharging and slow down the attenuation; if the battery health state SOH is lower than 80%, the basic boundary is further contracted to [30%, 80%], so as to prolong the remaining service life; When the actual SOC of the energy storage system exceeds the safe charging and discharging boundary, a higher attenuation cost coefficient is assigned to it in the optimization model , which is used to quantify the equivalent economic loss caused by unit charging and discharging operation at the current time to the battery life, and the coefficient is calculated through the following empirical model: ； Wherein, I(t) is the current charging and discharging current; is the rated capacity; CycleCount is the cumulative cycle count, CycleLife is the rated cycle life; exp is the abbreviation of exponential function, indicating the exponential operation with the natural constant e as the base; the coefficient , , , is obtained by fitting the battery accelerated aging test data; S3, on the day before the current operation day, based on the prediction parameters, a first optimization model is established and solved to maximize the overall operation satisfaction of the virtual power plant, and a pre-scheduling plan for 24 hours in the future is generated; wherein the overall operation satisfaction is a collaborative decision result based on a fuzzy satisfaction function of multiple targets; the multiple targets include a grid demand matching degree, a total operation cost, an energy storage attenuation cost and a load comfort degree; The fuzzy satisfaction function is in the form of a weighted geometric mean of multiple sub-target satisfaction functions, that is: ； wherein, is the satisfaction degree of the i-th sub-target, is its weight, x is the dispatching decision variable; is the cost satisfaction degree, which is inversely proportional to the total operation cost, the lower the cost, the closer the satisfaction degree to 1; is the grid demand matching satisfaction degree, which is inversely proportional to the deviation of the actual response amount and the dispatching instruction; is the energy storage attenuation satisfaction degree, which is inversely proportional to the total attenuation cost; The total attenuation cost is equal to ; is the attenuation cost coefficient, is the absolute value of the battery charge and discharge power; S4, in the current operation day, taking the pre-scheduling plan as a reference, the following steps are executed at fixed time intervals: S41, according to the latest operation data, the dynamic operation boundary and the prediction parameters of each resource are updated; S42, under the constraint of the updated boundary, a second optimization model with a shortened optimization time domain is established and solved to generate real-time scheduling instructions for the current period, and the instructions are issued to each resource for execution; S5, the actual output of each resource and the deviation of the real-time scheduling instructions are monitored in real time, and when the deviation exceeds a threshold, a collaborative deviation compensation mechanism based on the real-time adjustable capacity of each resource is started to balance the power.

2. The collaborative scheduling method for virtual power plant according to claim 1, wherein, In S4, the second optimization model adopts a model predictive control (MPC) framework, and the shortened optimization time domain is 2 to 6 hours.

3. The method for collaborative scheduling of virtual power plant according to claim 1, wherein, In S5, the collaborative deviation compensation mechanism includes the following operations: According to the power shortage or surplus and the real-time adjustable margin of each resource, the real-time contribution capacity of each resource is calculated; wherein the real-time adjustable margin is calculated based on the difference between the current actual output of the resource and the maximum output, and the closeness between the current state of the resource and the dynamic operation boundary; According to the size of the real-time contribution capacity, the power value to be compensated is proportionally allocated to one or more resources to jointly bear.

4. The collaborative scheduling method for virtual power plant according to claim 3, wherein, After completing a demand response or auxiliary service, the contribution degree of each resource in the collaborative deviation compensation mechanism is calculated based on the actual compensation amount, response speed and response accuracy, and the Shapley value method is used for benefit distribution.

5. The method for collaborative scheduling of virtual power plant according to claim 1, wherein, The ultra-short-term prediction is realized by using a bidirectional long short-term memory network prediction model based on an attention mechanism, and hyperparameters of the model are optimized by using a quantum genetic algorithm to obtain the prediction parameters.

6. The method for collaborative scheduling of virtual power plant according to claim 1, wherein, Uncertainty intervals in the prediction parameters are calculated by using a Conformal Prediction method, and robust optimization constraints are provided for the first optimization model and the second optimization model.

7. The method for collaborative scheduling of virtual power plant according to claim 1, wherein, S6, an adaptive learning step, is further included. Historical operation data, including prediction bias and energy storage attenuation model bias, are periodically collected, and the prediction model used to generate the prediction parameters and the attenuation model are updated and trained by using the historical operation data.

8. A collaborative scheduling system for a virtual power plant, characterized in that, A computer program is stored thereon, and the computer program is executed by a processor to realize steps of the collaborative scheduling method for a virtual power plant according to any one of claims 1 to 7.

Images