Energy storage power station multi-time scale optimization scheduling method and system fusing adaptive distribution robustness and model prediction control

Abstract

The invention discloses an energy storage power station multi-time-scale optimization scheduling method and system fusing adaptive distribution robustness and model prediction control, and the method comprises the following steps: obtaining the data of an energy storage power station, and solving a hierarchical planning model containing weekly-ahead distribution robustness optimization and day-ahead adaptive distribution robustness optimization; obtaining a robust day-ahead energy storage power scheduling plan and a charge state reference trajectory; respectively taking the day-ahead energy storage power scheduling plan and the charge state reference trajectory as a center reference and a terminal constraint of a model prediction controller, and performing rolling optimization solution in each control period by using the model prediction controller to obtain a corresponding power set value; and superposing the power set value with a fine tuning power amount to generate a final energy storage system output power instruction. On the premise of ensuring the robustness of the system, the decision conservative property is obviously reduced, and the economic benefit and the execution precision of the scheduling plan are improved.

Description

Technical Field

[0001] This invention relates to power system operation and optimization technology, specifically to a multi-timescale optimization scheduling method and system for energy storage power stations that integrates adaptive distributed rod and model predictive control. Background Technology

[0002] Against the backdrop of a global energy structure transitioning towards low-carbon and clean energy, renewable energy sources, represented by wind and solar power, are being integrated into modern power systems at an unprecedented scale and speed. However, these new energy sources inherently possess volatility, randomness, and intermittency, making their power output difficult to predict accurately. This poses unprecedented challenges to the real-time power balance, frequency stability, voltage control, and economic dispatch of the power grid. Traditional dispatch methods relying on deterministic forecasting often fall short when faced with significant errors between renewable energy output and predicted values. Such dispatch schemes may lead to substantial power deviations in actual operation, potentially even threatening the safe and stable operation of the entire power grid in severe cases.

[0003] Against this backdrop, Energy Storage Systems (ESS) are widely recognized as a core technology for mitigating renewable energy fluctuations, enhancing grid flexibility and resilience, and promoting high-proportion renewable energy consumption due to their advantages such as rapid power response, flexible bidirectional charging and discharging capabilities, and flexible deployment locations. ESS can efficiently transfer electrical energy over time, charging during periods of low electricity prices or high renewable energy generation, and discharging during periods of high electricity prices or high load. This not only achieves peak shaving and valley filling but also provides critical ancillary services to the grid, such as frequency regulation, voltage regulation, and backup power. Therefore, developing an efficient, reliable, and economical optimized dispatch strategy for energy storage power stations to maximize their lifecycle value in complex market environments while ensuring the safe operation of the grid has become a hot topic and a challenge for both power system research and the energy storage industry. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a multi-timescale optimization scheduling method and system for energy storage power stations that integrates adaptive bibliometric planning and model predictive control. This system can realize a multi-timescale energy storage optimization scheduling framework that dynamically and adaptively adjusts robustness and seamlessly connects long-term robust planning with real-time rolling control, so as to maximize the economic benefits of energy storage power stations while ensuring the safe and reliable operation of the system.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A multi-time-scale optimal scheduling method for energy storage power plants that integrates adaptive distributed rod and model predictive control includes the following steps: Data from energy storage power stations is acquired and a hierarchical planning model including pre-weekly sub-Blu-rod optimization and day-ahead adaptive sub-Blu-rod optimization is solved to obtain a robust day-ahead energy storage power scheduling plan and a state of charge reference trajectory. The daytime energy storage power scheduling plan and the state of charge reference trajectory are used as the central reference and terminal constraint of the model predictive controller, respectively. The model predictive controller is used to perform rolling optimization in each control cycle to obtain the corresponding power setpoint. The power setpoint is superimposed with the fine-tuning power to generate the final energy storage system output power command, which is then executed.

[0006] Furthermore, the energy storage power station data includes long-term forecast information. When solving the hierarchical planning model that includes pre-weekly sub-Blule bar optimization and day-ahead adaptive sub-Blule bar optimization, the pre-weekly sub-Blule bar optimization model is solved first using the long-term forecast information to obtain the energy storage plan and state of charge trajectory for the next week. The target result is selected from these as the boundary conditions and input data for the day-ahead adaptive sub-Blule bar optimization model. Then, the adaptive function of the day-ahead adaptive sub-Blule bar optimization model is calculated based on the forecast lead time, which is the time difference between the current scheduling time and the future execution time. Finally, the day-ahead adaptive sub-Blule bar optimization model is solved.

[0007] Furthermore, the mathematical expression of the objective function of the pre-week Brussels bar optimization model is as follows:

[0008] in, This represents the first-stage decision variables within the pre-week scheduling cycle. Its feasible region; For an uncertain random variable in the coming week, This is a function for the total weekly operating cost; It is a Wasserstein fuzzy set used to characterize long-term uncertainty, based on empirical distribution. Centered on, with radius The probability distribution sphere; Let P be the expected value under the true probability distribution P; The mathematical expression for the objective function of the current adaptive split-bar optimization model is as follows:

[0009] in, For the day-ahead decision variables in the first stage, Its feasible region; For an uncertain random variable in the next 24 to 48 hours, This is a function for the total daily operating cost; It is an adaptive Wasserstein fuzzy set, and its mathematical expression is as follows:

[0010] in, Represents all definitions in the support set The set of probability distributions on the x-axis. Let P represent a probability distribution. Representative predictive lead time An adaptive function.

[0011] Furthermore, the mathematical expression of the adaptive function is as follows:

[0012] In the above formula, The radius of the farthest forecast period corresponding to the forecast lead time. The radius of the nearest forecast period corresponds to the forecast lead time, where k is the decay constant and the parameter is... The statistical variance of historical prediction error data under different prediction lead times is fitted and calibrated.

[0013] Furthermore, when solving the day-ahead adaptive sub-Blule bar optimization model, the specific steps are as follows: First, the day-ahead adaptive sub-Blule bar optimization model is reconstructed. Then, the solver of the decomposition algorithm is used to solve the reconstructed model to obtain the day-ahead energy storage power scheduling plan and the state of charge reference trajectory. The reconstruction of the day-ahead adaptive sub-Blule bar optimization model includes the following steps: The maximization problem contained in the aforementioned adaptive sub-Brussels bar optimization model is transformed into its dual form, as shown in the following mathematical expression:

[0014] in, It is a non-negative dual variable. It is an adaptive function. For the day-ahead decision variables in the first stage, For an uncertain random variable in the next 24 to 48 hours, This is a function for the total daily operating cost. For an uncertain set of random variables, For nodes i Uncertain variables, To support the set; Substituting the dual form into the objective function of the day-ahead adaptive sub-Brussels bar optimization model, and introducing auxiliary variables for linearization, we obtain the reconstructed objective function, the mathematical expression of which is as follows:

[0015] In the above formula, As an auxiliary variable, And simultaneously satisfy the following constraints: .

[0016] Furthermore, the constraints of the hierarchical programming model include: The dynamic equilibrium constraint of the charged state is mathematically expressed as follows:

[0017] In the above formula, , For different states of charge at different times, For charging efficiency, For discharge efficiency, For charging power, For discharge power, For equipment capacity, t For time, For a period of time; The mathematical expressions for charging / discharging power and state constraints are as follows:

[0018]

[0019]

[0020] In the above formula, This is the maximum charging power. This is the maximum discharge power. For whether to charge, use a 0-1 variable. The variable is 0-1 indicating whether or not to discharge. The SOC safety range and ramp rate constraints are expressed mathematically as follows:

[0021]

[0022] In the above formula, In a charged state, , For the minimum and maximum states of charge, , For charging or discharging power at different times, This represents the upper limit of the power phase difference; The system power balance constraint is mathematically expressed as follows:

[0023] In the above formula, This is the predicted generator output value. For the uncertainty error in generator output prediction, For discharge power, For charging power, This represents the load forecast.

[0024] Furthermore, when the day-ahead energy storage power scheduling plan and the state of charge reference trajectory are used as the central reference and terminal constraint of the model predictive controller, respectively, the day-ahead energy storage power scheduling plan is used as the central reference power trajectory of the model predictive controller, and the state of charge reference trajectory is used as the terminal constraint of the model predictive controller. The mathematical expression is as follows:

[0025]

[0026] in, It is the center reference power trajectory. and These are the day-ahead energy storage power dispatch plans. The discharge power and charging power in the middle, It is based on the current power plan The corresponding charged state trajectory was derived. The value of H in the future prediction time domain for the control period k. It is the terminal charge state constraint of the model predictive controller.

[0027] Furthermore, the energy storage power station data includes ultra-short-term forecast information. When the model predictive controller performs rolling optimization in each control cycle, it specifically uses the ultra-short-term forecast information to solve the optimization problem of the objective function of the model predictive controller under the physical constraints and terminal constraints of the energy storage power station, obtains the optimal power sequence, and selects the first optimal power as the power setpoint corresponding to the current control cycle. The mathematical expression of the objective function of the model predictive controller is as follows:

[0028] Where k represents the k-th control cycle, and H represents the prediction time domain. The energy storage power sequence optimized by the model predictive controller in this time domain; It is the center reference power trajectory. It is based on the latest ultra-short-term forecast calculations of real-time energy imbalance costs. It's the cost of battery degradation. These are the weighting coefficients for each item.

[0029] Furthermore, the mathematical expression for the fine-tuning power is as follows:

[0030] in, It is the rated frequency of the power grid. It is the real-time measured power grid frequency. It is the droop coefficient.

[0031] This invention also proposes a multi-timescale optimal scheduling system for energy storage power plants that integrates adaptive distributed bar and model predictive control, including a processor and a computer-readable storage medium. The computer-readable storage medium stores a computer program, which is executed by the processor to implement the steps of the multi-timescale optimal scheduling method for energy storage power plants that integrates adaptive distributed bar and model predictive control.

[0032] Compared with the prior art, the advantages of the present invention are as follows: This invention captures long-term uncertainties through week-ahead sub-Blule bar optimization, constructs an initial robust boundary, and lays the foundation for global optimality in terms of economy. It tightens or relaxes the robustness level in real time through day-ahead adaptive sub-Blule bar optimization to avoid over-conservatism. The two-layer programming structure of week-ahead sub-Blule bar plus day-ahead adaptive sub-Blule bar makes robustness change from static preset to dynamic adaptive.

[0033] This invention uses the day-ahead scheduling plan as the central benchmark to guide short-term decisions to approach economic optimality, uses the state of charge reference trajectory as the terminal constraint to ensure that long-term energy storage capacity does not exceed the limit, embeds the global optimality under a large time scale into the real-time solution under a small time scale, and the model predictive controller re-solves based on actual measurement feedback in each control cycle, using online optimization to compensate for model errors and disturbances in offline planning, and eliminates scheduling gaps in multi-scale switching.

[0034] This invention superimposes a fine-tuning power amount onto the power setpoint output by the model predictive controller, thereby responding to ultra-short-term random fluctuations. Without disrupting the current economic framework, it transforms robustness into executable fine-grained actions, balancing safety constraints and dynamic response speed. Attached Figure Description

[0035] Figure 1 This is a schematic diagram illustrating the technical framework of the method according to an embodiment of the present invention.

[0036] Figure 2 This is a simplified flowchart of the method according to an embodiment of the present invention.

[0037] Figure 3 This is a detailed flowchart of the method according to an embodiment of the present invention. Detailed Implementation

[0038] The present invention will be further described below with reference to the accompanying drawings and specific preferred embodiments, but this does not limit the scope of protection of the present invention.

[0039] To address the increasing uncertainty in power systems, academia and industry have developed various advanced optimization and scheduling methods. Among them, stochastic programming (SP) aims to optimize the expected operating cost across all scenarios by sampling the probability distributions of uncertain variables (such as wind and solar power output and electricity prices) across numerous scenarios. However, the performance and reliability of stochastic programming methods are highly dependent on an accurate understanding of the true probability distributions of uncertain variables. In practical applications, the prediction error distribution of renewable energy output often exhibits complex characteristics such as non-Gaussianness and leptokurtic peaks with heavy tails, making it difficult to accurately describe using standard analytical functions. Incorrect distribution assumptions can lead to significant deviations between scheduling results and actual operating conditions, thereby reducing the reliability of the solution. Furthermore, the large number of scenarios generated to fully characterize uncertainty can trigger the "curse of dimensionality," causing the computational complexity of model solving to increase exponentially, limiting its feasibility in applications requiring rapid decision-making.

[0040] Unlike stochastic programming, robust optimization (RO) provides a solution that does not rely on probability distribution information. It constructs a deterministic set of uncertainties (e.g., a set of boxes or polyhedra based on upper and lower bounds) and seeks the optimal solution within this set that still satisfies all operational constraints under all possible worst-case scenarios. A key advantage of this approach is its ability to provide strict feasibility guarantees for scheduling schemes, effectively "immune" to worst-case scenarios. However, a major drawback of robust optimization is that its results are often overly conservative. Because the model aims to withstand extremely low-probability, low-probability, extreme uncertainties, it can lead to excessively high operating costs for scheduling schemes, resulting in underutilization of energy storage assets and significant economic losses under most normal operating conditions. While two-stage or multi-stage adaptive robust optimization alleviates this conservatism to some extent, its characterization of the uncertain set still fails to effectively utilize the valuable distribution information contained in historical data.

[0041] To achieve a better balance between the economy of stochastic programming and the reliability of robust optimization, Distributed Robust Optimization (DRO) has emerged as a cutting-edge theoretical framework that bridges the gap between the two. DRO does not require a perfectly accurate probability distribution; instead, it constructs a "fuzzy set" containing all possible true distributions and then seeks the optimal solution under the worst-case probability distribution within this fuzzy set. This approach avoids the strong dependence of stochastic programming on precise distributions and, by utilizing partial distribution information in the data (such as mean, variance, or historical samples) to limit the size of the fuzzy set, it is more economical than traditional robust optimization. Currently, constructing a fuzzy set based on Wasserstein distance (also known as "bulldozer distance") is a proven and highly effective method. It constructs a Wasserstein sphere as the fuzzy set, centered on an empirical distribution composed of historical data. Decision-makers can intuitively and effectively weigh the robustness and economic cost of scheduling schemes by flexibly adjusting the radius of this sphere.

[0042] Meanwhile, the actual operation of energy storage power stations exhibits significant multi-timescale characteristics. To achieve global optimum, the decision-making chain needs to span multiple levels, including long-term planning and real-time control. Model Predictive Control (MPC) is an online optimization technique widely used in real-time control. Its core idea is to utilize the latest prediction information to perform rolling optimization of the control problem within a finite future time (i.e., the prediction time domain) in each control cycle, and execute only the first control command of the optimized sequence. This closed-loop control mechanism based on feedback correction enables it to effectively cope with real-time prediction errors. However, using MPC alone also has significant limitations. Its limited prediction time domain may lead to "short-sighted" decision-making, failing to guarantee global optimum over the entire scheduling cycle (e.g., 24 hours), and its performance is highly dependent on the accuracy of short-term predictions and the quality of the tracked reference trajectory.

[0043] Although the above methods have achieved certain application results in the field of energy storage scheduling, existing technologies still have the following key technical gaps and shortcomings in building an optimization framework that can truly adapt to the dynamic characteristics of the real world and achieve seamless multi-scale coordination: First, existing DRO (Distributed Renewable Energy Output) applications suffer from a mismatch between static uncertainty characterization and dynamically changing prediction accuracy. Current DRO methods based on Wasserstein distance typically apply a fixed, uniform fuzzy set radius to all uncertainties across the entire 24-hour scheduling time domain when performing day-ahead scheduling. However, a fundamental and crucial physical reality is that prediction uncertainty dynamically changes over time—the prediction error for renewable energy output 24 hours later is statistically far greater than the prediction error 1 hour later. Using a fixed radius means the model employs the exact same metric for uncertainty across all timeframes, which is logically flawed. This static model inevitably faces a dilemma: either a large radius is set to cope with high long-term uncertainty, leading to overly conservative and costly near-term decisions; or a small radius is set to balance near-term economics, sacrificing the robustness required for long-term decisions and failing to achieve a globally optimal risk-cost balance.

[0044] Secondly, the application of Dependent Robust Optimization (as a long-term planning tool) and Model Predictive Control (as a real-time control tool) is fragmented, failing to form an effective synergistic mechanism. Existing research typically uses DRO for long-term planning and scheduling such as day-ahead, while using MPC for short-term rolling control such as intraday or real-time. The two are often treated as independent tools for solving problems at different time scales, lacking organic integration. MPC's performance is limited by the quality of its reference trajectory, which is usually generated by simple deterministic predictions, making it vulnerable to significant, unforeseen long-term deviations. On the other hand, while DRO can generate a robust long-term plan that considers various uncertainties, it is itself open-loop, lacking a mechanism for feedback correction based on real-time information. Organically combining the two—using the robust day-ahead plan generated by DRO as the central reference trajectory and boundary constraints for MPC—can significantly improve MPC's performance, ensuring it remains anchored within a long-term robust "safe corridor" during real-time corrections. This hierarchical synergistic control concept has not been fully elucidated or effectively implemented in existing technologies.

[0045] In summary, existing technologies suffer from limitations due to overly strong model assumptions (such as stochastic programming), overly conservative results (such as robust optimization), or a failure to dynamically adjust their robustness levels to adapt to the time-varying characteristics of prediction accuracy (such as static DRO). Furthermore, DRO and MPC, two powerful optimization tools, are often used in isolation, failing to leverage their inherent complementary advantages.

[0046] To address the conservative decision-making issues caused by static uncertainty models in existing technologies, as well as the technical shortcomings of long-term planning and real-time control being disconnected and lacking effective coordination, this embodiment proposes a multi-timescale optimization scheduling method for energy storage power plants that integrates adaptive distributed robust optimization (WDRO) and model predictive control. It innovatively improves and systematically applies the Wasserstein WDRO theory, deeply integrating it with model predictive control technology to construct a hierarchical collaborative control architecture. Its aim is to achieve a dynamic balance between the robustness and economy of energy storage system operation in complex and uncertain power system environments, thereby maximizing the economic value of energy storage assets while ensuring the safe and reliable operation of the power grid.

[0047] First, combine Figure 1 The following three logically progressive and complementary core parts will be used to systematically explain the concept: The first aspect will introduce in detail a multi-timescale scheduling architecture based on Vastel-Brussels bar optimization. Its purpose is to build a complete technical framework from pre-week macro planning to intraday real-time control, and to establish Vastel-Brussels bar optimization as its core theoretical foundation. The second aspect will elaborate on a day-ahead adaptive sub-Bruker optimization model, which aims to make key innovations to the standard Vastel sub-Bruker optimization theory by introducing a dynamic adaptive radius mechanism to solve the core technical pain point that static models cannot match the accuracy of time-varying predictions. The third aspect will present a complete closed-loop collaborative mechanism and solution process for day-ahead planning and real-time control. Its purpose is to clarify the collaborative interfaces and data flows between scheduling levels in detail, and to provide a complete algorithmic solution for the entire set of complex optimization problems in an engineering-oriented and efficient manner.

[0048] 1. A multi-time-scale scheduling architecture based on Vassile bar optimization.

[0049] Firstly, this embodiment provides a multi-timescale scheduling architecture that organically integrates multiple decision-making stages of an energy storage power station, from pre-weekly macro-planning to real-time precise execution. The establishment of this architecture aims to ensure the continuity and consistency of decisions at each scale through a hierarchical collaborative mechanism. Its core theoretical basis is Wasserstein Distributionally Robust Optimization (WDRO). The construction of this architecture can be achieved through the following steps: Step S1.1: Construct the pre-week split-bar optimization layer; Step S1.2: Establish the current-day refined planning layer; Step S1.3: Establish an intraday rolling correction layer; Step S1.4: Deploy the real-time control execution layer and build the complete architecture.

[0050] Step S1.1 serves as the top-level macro-planning for the entire scheduling architecture. Its core task is to generate a highly robust energy storage plan and weekly state of charge (SOC) trajectory for the next 168 hours, based on long-term market and renewable energy forecasts and applying a standard distributed cyclic optimization (DRO) model. The objective function of this week-long DRO model aims to minimize the expected total operating cost under the worst-case probability distribution. Its mathematical expression is as follows: (1) in, These represent the first-stage decision variables within the previous week's scheduling cycle (such as the approximate daily charging and discharging schedule). Its feasible domain. For uncertain random variables in the coming week (such as weekly average wind and solar power output error). This is the function for the total weekly operating cost. Let be the expectation under the true probability distribution P. It is a Wasserstein fuzzy set used to characterize long-term uncertainty, based on empirical distribution. Centered on, with radius The probability distribution sphere. This set is an empirical distribution consisting of N historical long-term prediction error samples. A Wasserstein sphere is constructed with the following definition: (2) in, Represents all definitions in the support set The set of probability distributions on the x-axis. P represents a probability distribution.

[0051] At the planning level a week ago, due to the long prediction time span and high degree of uncertainty, the radius of the fuzzy set... It is set to a large, fixed constant. This ensures that the generated weekly plan can withstand significant long-term forecast biases and has sufficient conservatism and robustness. After solving this level, its key outputs, such as the initial SOC target for the first day (i.e., T-24 hours), will serve as boundary conditions and inputs for the next level of day-ahead finer planning.

[0052] Step S1.2 is the core link between long-term planning and short-term execution. Its task is to develop a more refined and accurate day-ahead power scheduling plan for the next 24 to 48 hours, based on the macro-level guidance provided by the pre-week layer.

[0053] In this architecture, the day-ahead layer also employs a distributed bar optimization model. However, directly applying a fixed-radius DRO model from the week-ahead layer to day-ahead scheduling has fundamental technical drawbacks. This is because the prediction accuracy is time-varying within the day-ahead scheduling cycle: the uncertainty in predicting output after 24 hours is much greater than the uncertainty in predicting output after 1 hour. Using a uniform fixed radius either leads to overly conservative short-term decisions to cope with high long-term uncertainty, or sacrifices the robustness of long-term decisions to pursue short-term economic efficiency. Therefore, this architecture introduces an adaptive concept in the day-ahead layer. The conceptual structure of its model is as follows: (3) The key difference from step S1.1 is that the radius of the fuzzy set is no longer a fixed constant. Instead, it became a predictor of the early stage. Related adaptive functions This design, which aims to dynamically adjust the robustness level of the model, is one of the core innovations of this invention. The design of this adaptive radius function and the complete construction of the model will be described in detail in the second aspect of this invention.

[0054] Step S1.3 aims to address ultra-short-term forecast updates and system disturbances that occur intraday and were not anticipated in the previous day's plan. This architecture employs Model Predictive Control (MPC) technology at this level for rolling online optimization and correction.

[0055] MPC operates within a short, forward-scrolling forecast time domain (e.g., the current time period plus 2-4 hours). At the beginning of each control cycle (e.g., 15 minutes), it solves an optimization problem once using the latest system state and ultra-short-term forecast information, and executes only the first control command of that optimization sequence. In each control cycle k, MPC solves an optimization problem within a finite time domain H to minimize a weighted sum of multiple objectives. Its typical objective function can be expressed as: (4) In the above formula This is the energy storage power sequence optimized by MPC in this time domain; The power reference trajectory obtained from the day-ahead planning layer (step S1.2) is a key link to ensure consistency between long-term and short-term goals. It is a real-time energy imbalance cost calculated based on the latest ultra-short-term forecasts. It's the cost of battery degradation. where represents the weighting coefficients for each term. This optimization problem is solved under the physical constraints of the energy storage power station and includes a critical terminal SOC constraint derived from the day-ahead plan: .

[0056] Step S1.4 is the final execution stage of the dispatching instructions. This step is responsible for accurately tracking the upper-level instructions and responding to the real-time needs of the power grid. Its control logic can be mathematically represented by a superposition principle: (5) in It is the actual power that the energy storage system ultimately outputs to the power grid. The minute-level power setting value issued by the MPC layer (step S1.3) within the day. This is a second-level or sub-second-level power adjustment generated by the real-time controller. This adjustment is a function of the real-time state of the power grid; for example, when providing primary frequency regulation services, it can be expressed as: (6) in, It is the rated frequency of the power grid. It is the real-time measured power grid frequency. It is the droop coefficient.

[0057] In summary, this invention constructs a complete, top-down hierarchical collaborative scheduling architecture through the integration of the above four levels and the clear definition of the mathematical model. In this architecture, the pre-week DRO layer (step S1.1) provides the macroscopic boundary for the day-ahead planning layer (step S1.2); the robust plan generated by the day-ahead planning layer provides the core reference benchmark for the intraday MPC layer (step S1.3); and the intraday MPC layer provides the precise power setpoint for the real-time control layer (step S1.4), ultimately forming a highly collaborative optimization control system with a closed loop of decision flow and data flow.

[0058] 2. A day-to-day adaptive split-bar optimization model.

[0059] Secondly, this embodiment provides an innovatively designed day-ahead adaptive dichotomous bar optimization (DRO) model as the core of a multi-timescale scheduling architecture. This model aims to make key improvements to the standard Vassite dichotomous bar optimization theory by introducing a dynamically adaptive fuzzy set radius mechanism to accurately address the core technical challenge of static uncertainty models failing to match the accuracy of time-varying predictions. The complete construction and solveable transformation of this model can be achieved through the following steps: Step S2.1: Define the adaptive Wasserstein fuzzy set; Step S2.2: Construct a two-stage adaptive DRO objective function and constraints; Step S2.3: Perform dual reconstruction of the model; Step S2.4: Form a solvable mixed integer linear programming (MILP) paradigm.

[0060] Step S2.1 aims to construct an uncertainty set that can dynamically reflect the evolution of prediction accuracy over time. First, this model is based on sample data of N historical prediction errors. Construct an empirical probability distribution This distribution assigns an equal probability of 1 / N to each historical sample point.

[0061] Secondly, regarding this distribution of experience A spherical region, or Wasserstein fuzzy set, is constructed using the Wasserstein distance as a metric. However, unlike the standard DRO model which uses a fixed radius, the adaptive Wasserstein fuzzy set proposed in this invention has a radius... It is no longer a fixed constant, but is defined as the prediction lead time. An adaptive function for (i.e., the time difference between the current scheduling time and the future execution time). Its mathematical definition is as follows: (7) A preferred implementation that accurately captures the characteristics of prediction uncertainty decay is to use the adaptive radius function. Defined as an exponentially decaying function: (8) In the above formula Corresponding to the farthest prediction period (e.g.) The radius (hours) represents the highest level of uncertainty. Corresponding to the farthest prediction period (e.g.) The radius (in hours) represents the highest level of uncertainty. k is a positive decay constant used to control the rate at which the radius shrinks from its maximum to its minimum.

[0062] These parameters The model can be fitted and calibrated based on the statistical variance of historical prediction error data under different lead times, thus providing a solid data-driven foundation. Through this adaptive radius function, the model can construct a larger fuzzy set for long-term decisions to ensure robustness, while constructing a smaller fuzzy set for short-term decisions to pursue economy, thereby achieving refined and dynamic management of risk and cost.

[0063] Step S2.2 constructs a two-stage partial Bruker optimization problem. Its objective is to minimize the expected total running cost under the worst-case probability distribution. The complete mathematical objective function is: (9) in, For the first phase of day-ahead decision variables (such as charge / discharge power plans) wait), Let be the total cost function. This optimization problem must satisfy the following physical operational constraints of the energy storage power station and system: State of Charge (SOC) dynamic equilibrium constraints: (10) In the above formula, in the above formula, , For different states of charge at different times, For charging efficiency, For discharge efficiency, For charging power, For discharge power, Let t be the equipment capacity and t be the time. For a time period.

[0064] Charge / discharge power and state constraints: (11) (12) (13) In the above formula, in the above formula, For charging power, For discharge power, This is the maximum charging power. This is the maximum discharge power. For whether to charge, use a 0-1 variable. The variable is 0-1 indicating whether or not discharge occurs, and t is time.

[0065] SOC safety range and ramp rate constraints: (14) (15) In the above formula, In a charged state, , For the minimum and maximum states of charge, , For charging or discharging power at different times, t represents the upper limit of the power difference, and t represents time.

[0066] System power balance constraints (this is the second-stage constraint, which needs to be applied to all...) (All of these conditions must be met) (16) In the above formula, This is the predicted generator output value. For the uncertainty error in generator output prediction, For discharge power, For charging power, t represents the load forecast value, and t represents time.

[0067] Step S2.3 performs dual reconstruction of the model. The objective function defined in step S2.2 involves an infinite-dimensional maximization problem with respect to the probability distribution P, which cannot be solved directly. This step utilizes the strong duality of optimal transport theory to reconstruct the internal maximization problem. This is precisely transformed into its dual problem, thus converting infinite-dimensional optimization into finite-dimensional optimization. Its dual form is: (17) In the above formula, It is a non-negative dual variable. It is an adaptive function. For the day-ahead decision variables in the first stage, For an uncertain random variable in the next 24 to 48 hours, This is a function for the total daily operating cost. For an uncertain set of random variables, Let i be an uncertain variable. To support the set. This is an introduced nonnegative dual variable, which in economics can be interpreted as the marginal price of uncertainty. This dual form transforms the original problem into a problem about... The minimization problem has a clearer structure, but it still contains an internal sup term that requires further processing. Through the above steps, this invention constructs a mathematically rigorous and physically complete multi-timescale decomposition bar optimization scheduling model. This model can not only handle multi-source, coupled uncertainties, but also lays the foundation for the subsequent design of efficient decomposition algorithms through its hierarchical structure.

[0068] Step S2.4 Substitutes the dual form from step S2.3 into the original objective function and performs linearization to obtain a final normal form that can be computed by the standard solver.

[0069] First, substituting the dual problem, the entire two-stage DRO problem is transformed into a single-layer optimization problem that only involves the minimum: (18) To handle the internal sup term, we introduce a set of auxiliary variables. And add the following constraints: (19) The above constraints are semi-infinite (requiring consideration of all...). (This holds true). For a linear cost function... and polyhedral support set This constraint can be equivalently replaced by only applying it to other constraints. A finite set of linear constraints holds at the vertices. Meanwhile, L1 in the expression... - Normative terms It can also be precisely linearized by introducing additional auxiliary variables and constraints.

[0070] Through the aforementioned series of mathematical transformations, including dual reconstruction, introduction of auxiliary variables, and linearization, this invention ultimately reconstructs the original, complex, nonlinear infinite-dimensional DRO problem with adaptive radius into a large-scale but structurally standard mixed integer linear programming (MILP) problem. This problem can be efficiently solved using decomposition algorithms such as column sum and constraint generation (C&CG) combined with mature commercial optimization solvers (such as Gurobi and CPLEX), thereby completely solving the computational feasibility problem of the model and laying a solid mathematical foundation for the engineering application of the technical solution of this invention.

[0071] 3. A closed-loop collaborative mechanism and solution process for day-ahead planning and real-time control.

[0072] Thirdly, this embodiment provides a closed-loop collaborative mechanism that tightly couples day-ahead robust planning with intraday and real-time control, and elaborates on the integrated algorithm flow and system implementation scheme for efficiently and automatically solving the entire set of multi-timescale, multi-model optimization problems. Its purpose is to clarify in detail the collaborative interfaces and data flows between various scheduling levels, and to transform the technical solution of this invention into a complete, engineering-deployable solution. The specific implementation steps are as follows: Step S3.1: Define the collaborative interface between day-ahead blue bar optimization and intraday model predictive control; Step S3.2: Describe the rolling optimization process of intraday model predictive control; Step S3.3: Plan the integrated multi-timescale optimization solution algorithm flow; Step S3.4: Construct a system implementation plan.

[0073] Step S3.1 is key to realizing the multi-timescale collaboration of the present invention and solving the problem of the separation between long-term and short-term decisions in the prior art. The present invention establishes a clear and strongly coupled collaboration interface to directly transform the output of the day-ahead adaptive DRO layer (described in the second aspect) into the core input and boundary of the intraday MPC layer (described in the first aspect).

[0074] First, a reference trajectory is defined. After solving the adaptive DRO model, the optimal first-stage decision variables obtained include the energy storage power scheduling plan for the next 24-48 hours. This plan was directly used as the central reference power trajectory for the intraday MPC model. : (20) in, and These are the day-ahead energy storage power dispatch plans. The discharge power and charging power in the process.

[0075] When performing rolling optimization, MPC's objective function will include minimizing the impact on the reference trajectory. The tracking deviations ensure that intraday corrections always prioritize achieving a long-term robust plan.

[0076] Then define the terminal state constraints based on the day-ahead power plan. The derived trajectory of the corresponding state of charge (SOC) This is used as the terminal SOC constraint for the MPC model. During any control period k within a day, when the MPC optimizes its future prediction time domain H, the following terminal constraint must be satisfied: (twenty one) in, It is based on the current power plan The corresponding charged state trajectory was derived. The value of H in the future prediction time domain for the control period k. It is the terminal state of charge constraint of the model predictive controller. This strong constraint is a key mechanism to prevent the MPC from making "greedy" decisions (such as over-discharging) for short-term gains (such as dealing with temporary electricity price spikes). It forces the short-term behavior of the MPC to serve the strategic goal of long-term SOC state management, effectively avoiding the problem of "shortsightedness".

[0077] Step S3.2, based on the cooperative interface defined in step S3.1, involves the intraday MPC layer executing cyclically over a forward-rolling time domain. In each control cycle k (e.g., T-1 hour, T-0.75 hour...), the MPC performs the following optimizations: (twenty two) At the same time, it is necessary to meet the dynamic constraints of the energy storage power station and the terminal SOC constraints.

[0078] MPC uses the latest ultra-short-term forecast information (e.g., higher-precision wind and solar forecasts for the next 1-4 hours) to solve the above optimization problem and obtain the optimal power sequence. However, according to the basic principles of MPC, only the first control instruction in the sequence is valid. The command is then sent to the real-time control layer for execution. Subsequently, the system state is updated, and the entire prediction time domain is rolled forward by one step, entering the optimization solution for the next control cycle.

[0079] Step S3.3 outlines a complete, end-to-end integrated solution algorithm flow, including the following steps: 1) Data preparation and input: The algorithm begins by importing the basic data required by the system, including the technical parameters of the energy storage power station, historical renewable energy output and load prediction and measured data, and the latest long-term and ultra-short-term forecast information.

[0080] 2) Pre-week macro planning (T-7 days to T-24 hours): Run the pre-week DRO model (as described in the first aspect), use long-term forecast information to solve for the macro energy plan for the coming week, and determine the initial SOC target for the next day (T-24 hours). .

[0081] 3) Detailed planning before the deadline (T-24 hours): a. Adaptive radius calculation: based on the predicted lead time (From 1 to 24 hours), using the adaptive function in step S2.1 The fuzzy set radius is calculated for each hour in the scheduling time domain.

[0082] b. Model Solving: Since the current adaptive DRO model described in the second aspect is a large-scale mixed integer linear programming (MILP) problem, direct solution may be inefficient. Therefore, this embodiment adopts an iterative algorithm based on column and constraint generation (C&CG) for efficient solution. The algorithm operates through an iterative framework of "main problem-subproblem": First, initialize the iteration parameters, setting the upper bound of cost UB to infinity and the lower bound LB to negative infinity; then, iteratively execute the following steps until convergence: (i) Solve a main problem (MP) containing only partially known worst-case scenarios to obtain a preliminary scheduling plan x* and an updated lower bound of cost LB; (ii) Based on the plan x*, solve a subproblem (SP) to find a new worst-case u* that maximizes the cost, and update the upper bound of cost UB accordingly; (iii) Determine whether the upper and lower bounds have converged. If convergence is achieved, the iteration ends; otherwise, the new worst-case u* is added back to the main problem as a constraint, and the next iteration begins. Finally, after convergence, a robust day-ahead power reference plan is output. and SOC reference trajectory .

[0083] 4) Intraday rolling adjustment (T-24 hours to T-0 hours): a. Start the MPC controller and process the data generated recently. and SOC reference trajectory As the core input.

[0084] b. Cyclic Execution: At a fixed frequency (e.g., every 15 minutes), the MPC controller receives the latest system status and ultra-short-term forecast information, performs the rolling optimization described in step S3.2, and outputs a power setpoint for the next 15 minutes. And based on the real-time state of the power grid (such as frequency deviation), a fine-tuning power amount is added. Ultimately, this forms and executes the actual output power of the energy storage system.

[0085] Step S3.4 Constructing the System Implementation Scheme. The entire method of this invention can be embedded in a system for multi-timescale optimized scheduling of energy storage power stations. In terms of hardware, this system includes at least one or more processors, and a memory for storing program instructions and data. In terms of software, the system can be constructed as a series of modular programs working in concert, integrated into the energy management system (EMS) of the energy storage power station, specifically including: 1) Data processing and prediction module: responsible for collecting and cleaning data and generating the required prediction information.

[0086] 2) Optimization Model Building Module: Responsible for automatically building mathematical programming models for pre-week DRO, pre-day adaptive DRO, and intraday MPC based on input data.

[0087] 3) Solver Interface Module: Responsible for calling commercial optimization solvers (such as Gurobi, CPLEX) to efficiently solve the constructed MILP and QP problems.

[0088] 4) Scheduling instruction execution module: responsible for parsing the optimization results and converting them into power instructions that can be sent to the real-time controller.

[0089] Based on the above, such as Figure 2 As shown, the method in this embodiment includes the following steps: S1) Construct a hierarchical sub-Bluer bar optimization planning model: Construct a hierarchical planning model that includes pre-week sub-Bluer bar optimization and day-ahead adaptive sub-Bluer bar optimization, obtain energy storage power station data and solve the hierarchical planning model that includes pre-week sub-Bluer bar optimization and day-ahead adaptive sub-Bluer bar optimization to obtain a robust day-ahead energy storage power scheduling plan and state of charge reference trajectory. S2) Apply Model Predictive Control (MPC) to perform rolling online corrections within the day: The daytime energy storage power scheduling plan and the state of charge reference trajectory are used as the central reference and terminal constraint of the model predictive controller, respectively. The model predictive controller is used to perform rolling optimization solutions in each control cycle to obtain the corresponding power setpoint. S3) Superimpose real-time adjustment and execute optimal power command: Superimpose the power setpoint with fine-tuning power to generate the final energy storage system output power command and execute the energy storage system output power command to complete the scheduling closed loop.

[0090] In step S1, the energy storage power station data includes the technical parameters of the energy storage power station, historical renewable energy output and load forecasts and measured data, and the latest long-term and ultra-short-term forecast information. When solving the hierarchical planning model that includes pre-weekly sub-Blule bar optimization and day-ahead adaptive sub-Blule bar optimization, specifically, the pre-weekly sub-Blule bar optimization model of step S1.1 is first solved using long-term forecast information to obtain the energy storage plan and state of charge trajectory for the coming week, and the target result (i.e., the initial SOC target for the next day) is selected from it. The boundary conditions and input data of the day-ahead adaptive sub-Blu-rod optimization model are used as the boundary conditions. Then, according to the formula (8) above, the adaptive function of the day-ahead adaptive sub-Blu-rod optimization model is calculated based on the predicted lead time. Finally, the day-ahead adaptive sub-Blu-rod optimization model in step S1.2 is solved. When solving the day-ahead adaptive sub-Blu-rod optimization model, the day-ahead adaptive sub-Blu-rod optimization model is reconstructed according to step 2.3 above. Then, according to step S3.3 above, the solver of the decomposition algorithm is used to solve the reconstructed model to obtain the day-ahead energy storage power scheduling plan and the state of charge reference trajectory, such as Figure 3 As shown.

[0091] In step S2, when the day-ahead energy storage power scheduling plan and the state of charge reference trajectory are used as the central reference and terminal constraint of the model predictive controller, respectively, the day-ahead energy storage power scheduling plan is used as the central reference power trajectory of the model predictive controller according to the formula (20) above, and the state of charge reference trajectory is used as the terminal constraint of the model predictive controller according to the formula (21) above. In step S2, when the model predictive controller performs rolling optimization in each control cycle, it specifically follows step S3.3 above to use ultra-short-term prediction information to solve the optimization problem of the objective function of the model predictive controller under the physical constraints and terminal constraints of the energy storage power station, obtains the optimal power sequence, and selects the first optimal power as the power setpoint corresponding to the current control cycle.

[0092] This embodiment also proposes a multi-timescale optimal scheduling system for energy storage power stations that integrates adaptive split-bar and model predictive control. The system includes a processor and a computer-readable storage medium. The computer-readable storage medium stores a computer program, which is executed by the processor to implement the steps of the multi-timescale optimal scheduling method for energy storage power stations that integrates adaptive split-bar and model predictive control as described in this embodiment. In addition to the functional module fingerprints described in step S3.4 above, the system may also include the following functional modules: A hierarchical planning module is configured to execute step S1 to construct and solve a hierarchical planning model that includes adaptive sub-Browser optimization for the week-ahead and day-ahead periods, thereby generating a robust day-ahead reference trajectory; a rolling correction module is configured to execute step S2 to use the day-ahead reference trajectory as the core input and apply model predictive control (MPC) for intraday rolling online correction to generate a power setpoint; and an instruction execution module is configured to execute step S3 to superimpose the power setpoint with the real-time fine-tuned power amount to form and execute the final energy storage power instruction, thereby completing the entire scheduling closed loop.

[0093] To further verify the effectiveness and superiority of the "multi-timescale optimization scheduling method and system for energy storage power stations" described in this invention through specific numerical values, this embodiment constructs a simplified 4-hour scheduling scenario for case study and quantitatively compares the performance of the method of this invention with three benchmark methods.

[0094] The example in this embodiment is based on a 4-hour dispatch scenario that includes changes in electricity prices and wind power forecasting errors. Its core parameters and data settings are shown in the table below.

[0095] Table 1 Parameter Table of Energy Storage Power Station

[0096] The data for a 4-hour dispatch scenario is shown in the table below: Table 2 4-hour dispatch scenario data table

[0097] Subsequently, a comparison was made between the day-ahead planned power and the actual operating power under different strategies. The table below details the day-ahead planned power and the actual operating power after considering the prediction error and MPC correction for the energy storage power station under four different strategies.

[0098] Table 3. Comparison of planned and actual energy storage capacity under different strategies (MW)

[0099] Note: The planned power of baseline 3 is the reference trajectory value in parentheses.

[0100] Based on the simulation results above, the core performance indicators of each strategy were calculated and summarized in the table below.

[0101] Table 4. Quantitative Comparison of Final Performance of Four Scheduling Strategies

[0102] Although the independent MPC strategy achieves the highest apparent profit in this specific example scenario due to its opportunistic nature, the superiority of the proposed "Adaptive DRO+MPC" lies not in the highest profit in a single scenario, but in its comprehensive performance as a professional "risk manager": First, based on the bibliometric optimization theory, this invention provides performance guarantees under the "worst-case" scenario, possessing the risk resistance and robustness lacking in independent MPC; second, and most importantly, the day-ahead average dispatch deviation generated by this invention is the lowest among all methods. In a real electricity market environment, lower dispatch deviation means lower penalty costs and higher reliability; therefore, the -104,000 yuan profit is a more realistic and reliable "net profit." Quantitative analysis through this specific embodiment shows that the comprehensive benefits of the proposed adaptive DRO+MPC method are significantly better than existing technologies, clearly demonstrating the significant and beneficial effects achieved by this invention in terms of robustness and economy in dynamic balance operation.

[0103] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0104] The above description is merely a preferred embodiment of this practice. The scope of protection of this practice is not limited to the above embodiments. All technical solutions falling within the scope of this practice are protected. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of this practice should also be considered within the scope of protection of this practice.

Claims

1. A method for multi-time scale optimal scheduling of energy storage power station fusing adaptive distribution robustness and model predictive control, characterized in that, The method comprises the following steps: obtaining energy storage power station data and solving a hierarchical planning model comprising week-ahead distribution robust optimization and day-ahead adaptive distribution robust optimization to obtain a robust day-ahead energy storage power scheduling plan and a state of charge reference trajectory; using the day-ahead energy storage power scheduling plan and the state of charge reference trajectory as a center reference and a terminal constraint of a model predictive controller respectively, and using the model predictive controller to solve a rolling optimization in each control period to obtain a corresponding power set value; superimposing a fine power amount on the power set value to generate a final energy storage system output power instruction and executing the energy storage system output power instruction.

2. The method of claim 1, wherein the method is characterized by, The energy storage power station data comprises long-term prediction information, and when solving the hierarchical planning model comprising week-ahead distribution robust optimization and day-ahead adaptive distribution robust optimization, the long-term prediction information is used to solve a week-ahead distribution robust optimization model to obtain an energy storage energy plan and a state of charge trajectory for a future week, and a target result is selected from the energy storage energy plan and the state of charge trajectory as a boundary condition and input data of a day-ahead adaptive distribution robust optimization model, an adaptive function of the day-ahead adaptive distribution robust optimization model is calculated according to a prediction lead time, the prediction lead time is a time difference between a current scheduling time and a future execution time, and finally the day-ahead adaptive distribution robust optimization model is solved.

3. The method of claim 2, wherein the method is characterized by, A mathematical expression of an objective function of the week-ahead distribution robust optimization model is as follows: in, This represents the first-stage decision variables within the pre-week scheduling cycle. Its feasible region; For an uncertain random variable in the coming week, This is a function for the total weekly operating cost; It is a Wasserstein fuzzy set used to characterize long-term uncertainty, based on empirical distribution. Centered on, with radius The probability distribution sphere; Let P be the expected value under the true probability distribution P. A mathematical expression of an objective function of the day-ahead adaptive distribution robust optimization model is as follows: where, is the day-ahead decision variable for the first stage, is its feasible region; is the uncertainty random variable for the next 24 to 48 hours, is the total operation cost function for the day-ahead; is the adaptive Wasserstein fuzzy set, mathematically expressed as follows: wherein, represents a set of all probability distributions defined on the support set represents a set of all probability distributions defined on the support set represents that P is a probability distribution, represents an adaptive function that predicts the lead time .

4. The method of claim 3, wherein, A mathematical expression of the adaptive function is as follows: In the above formula, a radius corresponding to the farthest prediction period in the prediction lead time, a radius corresponding to the nearest prediction period in the prediction lead time, k is an attenuation constant, and parameters based on fitting and calibrating the statistical variance of historical prediction error data at different prediction lead times.

5. The method of claim 4, wherein, When solving the day-ahead adaptive distribution robust optimization model, the day-ahead adaptive distribution robust optimization model is first reconstructed, and then a solver of a decomposition algorithm is used to solve the reconstructed model to obtain the day-ahead energy storage power scheduling plan and the state of charge reference trajectory, and when reconstructing the day-ahead adaptive distribution robust optimization model, the following steps are included: A maximization problem contained in the day-ahead adaptive distribution robust optimization model is converted into a dual form, and a mathematical expression is as follows: wherein, is a non-negative dual variable, is an adaptive function, is a day-ahead decision variable of the first stage, is an uncertainty random variable in the next 24 to 48 hours, is a total hourly operation cost function, is a set of uncertain random variables, is an uncertain variable of node i, is a support set; The dual form is substituted into an objective function of the day-ahead adaptive distribution robust optimization model, and an auxiliary variable is introduced for linearization processing to obtain a reconstructed objective function, and a mathematical expression is as follows: In the above formula, are auxiliary variables, while satisfying the constraints: .

6. The method of claim 1, wherein, Constraint conditions of the hierarchical planning model include: A state of charge dynamic balance constraint, a mathematical expression is as follows: In the above formula, , is the state of charge at different times, is the charging efficiency, is the discharging efficiency, is the charging power, is the discharging power, is the device capacity, t is time, is the time period; A charging and discharging power and state constraint, a mathematical expression is as follows: In the above formula, Pmax is the maximum charging power, Pmax is the maximum discharging power, C is a 0-1 variable for whether to charge, D is a 0-1 variable for whether to discharge, and t is time. An SOC safety range and climbing rate constraint, a mathematical expression is as follows: in the above formulae, is the state of charge, , is the minimum, maximum value of the state of charge, , is the charging or discharging power at different times, is the upper limit value of the power difference, t is time; A system power balance constraint, a mathematical expression is as follows: In the above formula, is a generator output prediction value, is a generator output prediction uncertainty error, is a discharge power, is a charge power, is a load prediction value, and t is time.

7. The method of claim 1, wherein, When the day-ahead energy storage power scheduling plan and the state of charge reference trajectory are used as a center reference and a terminal constraint of a model predictive controller respectively, the day-ahead energy storage power scheduling plan is used as a center reference power trajectory of the model predictive controller, and the state of charge reference trajectory is used as a terminal constraint of the model predictive controller, and a mathematical expression is as follows: in, It is the center reference power trajectory. and These are the day-ahead energy storage power dispatch plans. The discharge power and charging power in the middle, It is based on the current power plan The corresponding charged state trajectory was derived. The value of H in the future prediction time domain for the control period k. It is the terminal charge state constraint of the model predictive controller.

8. The method of claim 1, wherein, The energy storage power station data includes ultra-short-term prediction information, and when a model predictive controller is used to solve a rolling optimization in each control period, specifically, the ultra-short-term prediction information is used to solve an optimization problem of an objective function of the model predictive controller under physical constraints and terminal constraints of the energy storage power station, to obtain an optimal power sequence and select a first optimal power in the optimal power sequence as a power set value corresponding to a current control period, and a mathematical expression of the objective function of the model predictive controller is as follows: where k denotes the kth control period, H denotes the prediction horizon, the sequence of energy storage powers optimized by the model predictive controller in this time horizon; is the central reference power trajectory, is the real-time energy imbalance cost calculated based on the latest ultra-short-term prediction, is the degradation cost of the battery, are the weight coefficients of each term.

9. The method of claim 1, wherein, A mathematical expression of the fine-tuning power amount is as follows: wherein, is the nominal frequency of the power grid, is the real-time measured frequency of the power grid, is the droop coefficient.

10. A multi-time scale optimal scheduling system for energy storage power station fusing adaptive distribution robustness and model predictive control, characterized in that, The method comprises a processor and a computer readable storage medium, the computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the steps of the energy storage power station multi-time scale optimization scheduling method based on fusion of adaptive distribution robustness and model predictive control according to any one of claims 1-9.

Images