A park source network load storage system multi-objective adaptive control method

Abstract

This invention discloses a multi-objective adaptive control method for a power generation, grid, load, and storage system in an industrial park, relating to the field of energy management technology. The method first constructs a generalized state space of physical equipment and virtualized production tasks, mapping the production tasks to a virtual energy storage model and constructing a congestion risk potential surface function. Then, it calculates and decomposes the total regret value of the strategy, dynamically updates the objective function weights based on external regrets, and corrects the virtual model parameters online based on internal regrets. Subsequently, it constructs a collaborative optimization proposition using the updated weights and parameters, incorporating congestion risk into the optimization objective, and solves for the optimal control sequence. Finally, it decouples the control sequence into physical charging / discharging commands and production operation commands. This invention achieves compatibility between discrete production and continuous energy control through virtualization mapping and utilizes a two-layer regret mechanism to achieve adaptive adjustment of strategies and parameters, improving the system's economy and robustness.

Description

Technical Field

[0001] This invention relates to the field of energy management technology, specifically to a multi-objective adaptive control method for a power generation, grid, load, and storage system in an industrial park. Background Technology

[0002] With the transformation of the energy structure and the advancement of the "dual carbon" goal, integrated source-grid-load-storage systems that integrate distributed renewable energy, electrochemical energy storage, industrial loads and main grid interaction have become the mainstream form of industrial park construction. The core of this type of system lies in coordinating the operation of each unit through an energy management system (EMS) to achieve multiple goals such as improving the renewable energy absorption rate, reducing energy costs and ensuring grid interaction friendliness.

[0003] In existing park energy management practices, control strategies often focus on the scheduling of the source and storage sides, usually treating industrial loads as uncontrollable rigid demands or flexible loads with only simple power reduction capabilities. However, industrial production is essentially composed of a series of discrete tasks with specific processes, output requirements, and delivery deadlines. Existing control methods often sever the coupling relationship between "energy flow" and "material flow," making it difficult to directly incorporate production tasks with discrete time-series characteristics (such as the number of processed parts or heating batches) into continuous power optimization models. This results in energy scheduling strategies failing to accurately respond to the dynamic demands of production plans, easily causing production delays or forcing high-power operation during peak electricity periods to meet deadlines, thus increasing operating costs.

[0004] Furthermore, traditional model predictive control (MPC) methods typically set production delivery dates as rigid end-point constraints. When faced with uncertainties such as fluctuations in renewable energy output or equipment failures, this hard constraint approach can easily lead to infeasibility during the rolling solution process, or cause severe power command fluctuations to meet constraints, impacting production equipment and the power grid. Most existing collaborative control systems rely on offline calibrated system parameters (such as equipment energy efficiency coefficients and battery capacity) and fixed optimization target weights. However, during long-term operation, production equipment may experience energy efficiency drift due to dust accumulation and wear, and changes in ambient temperature can affect the energy consumption model of thermal equipment. At the same time, adjustments to electricity pricing policies or operational strategies require the focus of optimization objectives to change dynamically. Static control strategies lacking online adaptive correction mechanisms can cause model predictions to gradually deviate from the actual physical state, resulting in significant deviations between the theoretically optimal solution and actual execution, making it difficult to maintain control accuracy and economy throughout the system's entire lifecycle. Summary of the Invention

[0005] This invention aims to solve the technical problems in the existing source-grid-load-storage coordinated control of industrial parks, such as the incompatibility between discrete time-series constraints of production load and continuous energy flow optimization, and the decrease in control accuracy caused by the drift of model parameters as equipment ages.

[0006] This invention provides a multi-objective adaptive control method for a power generation, grid, load, and storage system in an industrial park.

[0007] This method operates in an energy management and control system. It constructs a generalized state space encompassing physical devices and virtualized production tasks, and utilizes a two-layer regret feedback mechanism to achieve adaptive correction of the control strategy. The method specifically includes the following steps:

[0008] A power balance constraint model for the physical energy subsystem is constructed, and real-time operational status data is acquired. This real-time operational status data includes physical-side data covering renewable energy output, grid electricity prices, and physical energy storage status, as well as production-side data covering production task progress, equipment power characteristics, and task deadlines. The power balance constraint model establishes real-time active power balance conditions, limiting the algebraic sum of the predicted output of renewable energy generation units, the main grid interaction power value, and the output power value of electrochemical energy storage units to equal the sum of the non-adjustable base load power and the operating power of adjustable production equipment within the park. Specifically, the main grid interaction power value is defined as positive when purchasing electricity from the grid and negative when feeding electricity to the grid; the output power value of the electrochemical energy storage unit is defined as positive when discharging and negative when charging. Simultaneously, the evolution of the physical energy storage status follows a discrete-time integral law; the physical battery state of charge at the next moment is determined by subtracting the normalized battery output energy calculated using the comprehensive conversion efficiency from the current state.

[0009] A generalized state space is constructed using the real-time operating status data, and a virtual battery state equation is established to map production tasks to a virtual energy storage model. A congestion risk penalty function is also constructed. When establishing the virtual battery state equation, a virtual state of charge (VCC) is defined to represent the completion rate of the production task. Its update rule is as follows: the VCC at the next moment equals the VCC at the current moment plus the effective production progress increment calculated using virtual charging efficiency, minus the progress loss caused by virtual self-discharge. The progress loss caused by virtual self-discharge is calculated based on the equipment's comprehensive heat loss coefficient, ambient temperature, and internal equipment temperature, and is used to quantitatively characterize the energy dissipation of the production equipment in standby or interrupted states.

[0010] To address the delivery deadline constraints of production tasks, this invention handles them by constructing a congestion risk penalty function. Specifically, it calculates the difference between the production task's deadline and the current time as the remaining production window, and calculates the minimum theoretical time required to complete the remaining task. The risk state is determined based on the ratio of the remaining production window to the minimum theoretical time, and a potential energy function is constructed that dynamically changes with time and production progress. When production delays cause the ratio to approach a critical threshold, the function's value increases non-linearly, thus forming a numerical gradient in the optimization solution space that promotes increased production progress.

[0011] The total policy regret value is calculated based on the virtual battery state equation, and regret decomposition is performed. This step, based on a sliding window mechanism, calculates the difference between the cumulative loss generated by the actual policy execution within a set time window and the cumulative loss generated by the theoretically optimal policy, thus obtaining the total policy regret value. The total policy regret value is then decomposed into two independent components: first, the gradient vector of the loss function with respect to the weight vector, serving as the external regret value characterizing the multi-objective weight allocation bias; second, the deviation between the measured value of the virtual state of charge and the predicted value obtained from the model prediction at the previous time step, serving as the internal regret value characterizing the model parameter mismatch.

[0012] The objective function weight vector is updated based on the external regret value, and the virtual capacity parameters of the virtual energy storage model are corrected based on the internal regret value. During the weight update process, an online gradient ascent algorithm is used to calculate the intermediate weight vector in the unconstrained state based on the external regret value and a learning rate that decays over time. The intermediate weight vector is then mapped to a feasible region satisfying non-negativity and normalization constraints by solving a projection optimization problem. During parameter correction, the internal regret value is used as the observation residual to construct a parameter sensitivity matrix containing the partial derivatives of the observed variables with respect to the unit product energy consumption coefficient and virtual conversion efficiency. The adaptive correction gain is calculated using the error covariance matrix. The unit product energy consumption coefficient and virtual conversion efficiency are then posteriorly corrected based on the internal regret value, and the corrected unit product energy consumption coefficient is used to update the virtual rated capacity in the virtual capacity parameters to compensate for model biases caused by equipment aging or environmental changes.

[0013] A collaborative optimization problem is constructed using the updated objective function weight vector and the corrected virtual capacity parameters. This step involves constructing a linearized cumulative objective function, accumulating the weighted operating economic cost, grid interaction fluctuation penalty term, and congestion risk potential surface function value in the prediction time domain, and adding a smoothing penalty term to limit the adjustment range of production power. Simultaneously, constraints are set for the collaborative optimization problem, including mutual exclusion constraints for charging and discharging of physical energy storage units, state-of-charge capacity constraints for physical energy storage units, and virtual terminal state constraints requiring the virtual state of charge to reach a preset threshold at the task deadline. By solving this optimization problem, the optimal control sequence that satisfies the preset future time steps is obtained.

[0014] The decoupling and issuance of control commands are performed. The physical charging power, physical discharging power, and production line operating power at the current moment are extracted from the optimal control sequence obtained from the solution. The difference between the physical charging power and physical discharging power is calculated and sent as a net power command to the energy storage converter. Simultaneously, using the corrected unit product energy consumption coefficient at the current moment, the production line operating power is divided by this coefficient, mapped to the production line conveyor speed setpoint and other production equipment control commands. After performing dead-zone correction and acceleration limiting on these commands, they are written into the production equipment controller.

[0015] This invention provides a multi-objective adaptive control method for a power generation, grid, load, and storage system in an industrial park. It offers the following advantages:

[0016] 1. This invention innovatively introduces the unit product energy consumption coefficient and virtual conversion efficiency by establishing a virtual battery mapping mechanism. It quantifies and equates industrial production tasks with discrete time-series characteristics into virtual energy storage models with charge-discharge characteristics, capacity, and self-discharge rate, and incorporates them into a unified generalized state space. This isomorphic modeling enables the controller to simultaneously solve the charging and discharging strategies of the physical battery and the operating cycle of the production line within a linear mathematical framework. This solves the problem of optimization difficulty caused by rigid production constraints in traditional methods and realizes efficient coordination of source, grid, load, and storage.

[0017] 2. This invention constructs a spatiotemporally coupled congestion risk potential energy surface function that dynamically changes with time and production progress. This transforms the hard constraint of the production task's deadline into a nonlinear gradient penalty term in the objective function. In the early stages of production, the system prioritizes economic efficiency. When the schedule lags behind and approaches the deadline, the potential energy surface automatically generates a strong gradient signal, forcing the optimizer to allocate more power to the production equipment. This mechanism not only reduces the situation where the optimizer is unable to find a solution due to constraint conflicts under extreme operating conditions, but also automatically balances the risk of delay and operating costs, improving the system's robustness in dealing with uncertainties.

[0018] 3. This invention introduces a two-layer online correction mechanism based on regret theory. On the one hand, it uses the external regret gradient to adjust the weight vector of the multi-objective function in real time to adapt to changes in electricity price fluctuations and load characteristics. On the other hand, it uses extended Kalman filtering to process the internal regret observation residuals, identify and correct the energy consumption coefficient and efficiency parameters of the virtual model online. This enables the control system to have the ability to "self-evolve" and automatically compensate for model mismatch caused by equipment aging, changes in ambient temperature, or differences in raw materials, ensuring optimal control accuracy and economic benefits throughout the entire life cycle. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention;

[0020] Figure 2 This is a flowchart of the virtual battery mapping mechanism of the present invention;

[0021] Figure 3 This is a flowchart of the loss function construction and regret value calculation decomposition of the present invention;

[0022] Figure 4 This is a flowchart illustrating the adaptive update of the external regret of the optimization target weights in this invention.

[0023] Figure 5 This is a flowchart of the internal regret drift correction for virtual model parameters in this invention;

[0024] Figure 6 A flowchart illustrating the construction of the rolling collaborative optimization proposition of this invention;

[0025] Figure 7 This is a flowchart of the quadratic programming solution and instruction decoupling execution of the present invention. Detailed Implementation

[0026] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] See attached document Figure 1 This invention provides a multi-objective adaptive control method for a power generation, grid, load, and storage system in an industrial park. This method operates within an energy management and control system and includes the following steps:

[0028] S1, obtain the real-time operating status data of the park's source-grid-load-storage system at the current moment. The real-time operating status data includes physical side data and production side data. The physical side data includes renewable energy output, grid electricity price and physical energy storage status, while the production side data includes production task progress, equipment power characteristics and task deadline.

[0029] S2, based on real-time operating status data, constructs a generalized state space that includes physical equipment and virtualized production tasks. By establishing a virtual battery state equation, production tasks with discrete time-series characteristics are mapped to a continuous virtual energy storage model. A spatiotemporal coupling blocking risk potential energy surface function is constructed to characterize the urgency and interruption risk of production tasks.

[0030] S3, calculate the total regret value between the actual executed strategy and the theoretical optimal strategy at the previous control time, and decompose the total regret value into external regret value, which represents the deviation of multi-objective weight allocation, and internal regret value, which represents the deviation of virtualization model parameter mismatch.

[0031] S4, perform two-layer adaptive correction. On the one hand, based on the external regret value, use the online gradient descent algorithm to update the objective function weight vector for the next time step. On the other hand, based on the internal regret value, perform reverse drift correction on the virtual capacity parameters of the virtual energy storage model.

[0032] S5, based on the updated objective function weight vector and the corrected virtual capacity parameters, constructs a collaborative optimization problem based on model predictive control, incorporates the blocking risk potential surface function into the optimization objective, and solves the optimal control sequence for the future preset time step under the premise of satisfying the system power balance constraint and virtual production constraint.

[0033] S6, extract the first set of control vectors in the optimal control sequence, decouple them into physical device control commands sent to the battery management system and production equipment control commands sent to the production equipment controller, and issue them for execution. Then the system enters the next sampling time.

[0034] To enable those skilled in the art to more clearly understand the specific implementation of the present invention in terms of virtualization modeling, regret feedback mechanism and collaborative optimization solution, the core algorithm principles and technical details involved in the above steps S1 to S6 will be described in detail below.

[0035] In one embodiment of the present invention, the data definition and system modeling process of the generalized state space is the foundation for subsequent collaborative optimization. This process is carried out in a discrete-time frame, with the system's time step set to [value missing]. The construction of the generalized state space is not limited to traditional electrical parameters, but further encompasses non-electrical parameters related to the production process, thus incorporating the physical energy subsystem and the virtualized production subsystem into a unified mathematical description system. The specific data definition and modeling process can be described through the following sub-steps S101 to S103:

[0036] S101. Construct a power balance constraint model for the physical energy subsystem. At the point of common coupling (PCC) of the park's microgrid, real-time active power balance conditions must be met. The system collects real-time power data from renewable energy generation units, main grid interaction power, electrochemical energy storage unit charging and discharging power, and park load power, establishing the following power balance equation:

[0037] ;

[0038] In the formula, Let t be the predicted output value of renewable energy sources such as photovoltaic or wind power; Let t be the power exchange value between the park and the main power grid at time t. Here, a positive value indicates that the park purchases electricity from the grid, and a negative value indicates that the park feeds electricity to the grid. Let t be the output power of the electrochemical energy storage unit at time t, where discharge is defined as positive and charging as negative. This refers to the power of the rigid, non-adjustable foundation loads such as lighting and office equipment within the park. The operating power of adjustable production equipment to be included in the scope of collaborative control.

[0039] S102 defines the dynamic state evolution equation of the physical energy storage unit. The state of charge (SOC) of the electrochemical energy storage unit reflects the energy buffering capacity of the physical system, and its state change follows integral characteristics. Based on the current and voltage data fed back by the battery management system (BMS), the system estimates the current SOC using the ampere-hour integral method or the open-circuit voltage method, and describes its evolution according to the following discrete-time equation:

[0040] ;

[0041] In the formula, Let be the state of charge of the physical battery at time t, and its value range is... ; This refers to the rated capacity of the physical battery. The overall conversion efficiency function related to charge and discharge power. The output power of the battery. is the discrete time step. When When >0 (discharge), Take the reciprocal of the discharge efficiency; when When <0 (charging), The charging efficiency is taken into account. For the specific charging and discharging physical characteristics and safety threshold settings of the physical energy storage unit, those skilled in the art can refer to existing battery management specifications, which will not be elaborated upon here.

[0042] S103 defines the state parameters of production-side tasks. To achieve production load virtualization, production tasks with discrete time-series characteristics need to be quantitatively defined. The system reads the operating parameters of key production equipment via an industrial control bus (such as OPC UA or Modbus). For each batch of production tasks, a total task quantity parameter is defined. This parameter characterizes the physical workload (such as total heating heat, number of processed parts, or liquid delivery volume) required to complete this batch of tasks, and defines the current task completion progress parameters. , which represents the amount of work completed up to time t.

[0043] To establish the mapping relationship between physical workload and energy consumption, a unit product energy consumption coefficient is defined. This coefficient represents the electrical energy required for a unit of physical work in production. Simultaneously, to quantify the thermodynamic losses or process rollback risks caused by production interruptions, the system collects environmental state parameters related to the production process. (such as ambient temperature) and internal equipment status parameters (e.g., furnace temperature), define the rated power parameters of the production equipment. And the running status bit. In addition, define the absolute cutoff time for the current production batch. This parameter is used to calculate the urgency of production tasks.

[0044] Combining the physical and production data mentioned above, the system constructs a generalized state vector X(t). This vector not only contains... Traditional energy state variables, as well as production task progress, are also included. and remaining time for the task Production state variables. This generalized state space definition allows the controller to simultaneously observe the energy system's supply and demand status and the production system's operational progress within a unified mathematical framework, providing data support for subsequently mapping heterogeneous production equipment into homogeneous virtual energy storage units.

[0045] See attached document Figure 2 After establishing the data definition of the generalized state space, this embodiment of the invention transforms production tasks with discrete temporal characteristics into a continuous virtual energy storage model through a virtualization mapping mechanism. The core of this mechanism lies in establishing a mathematical isomorphism between the physical operation process and the energy storage process, enabling the production equipment to behave as a virtual battery with charging and discharging characteristics at the control level. The specific implementation process includes the following sub-steps S201 to S205:

[0046] S201, determine the rated capacity parameters of the virtual battery. The virtual capacity corresponds to the total energy requirement needed to complete a specific batch of production tasks. The system is based on the total task parameters collected in step S203. and unit product energy consumption coefficient Calculate the virtual rated capacity of the current task. The initial estimation formula for this parameter is:

[0047] ;

[0048] In the formula, the unit product energy consumption coefficient The initial value is obtained by analyzing the ratio of energy consumption data to output over historical operating cycles of the production equipment, or by using the nominal energy consumption value in the equipment's factory technical specifications. In subsequent control processes, this parameter will act as an adaptive parameter, dynamically corrected based on internal deficit values ​​during actual operation to reflect actual energy consumption drift caused by equipment aging or changes in raw materials. The virtual rated capacity for the current task. This refers to the total number of tasks.

[0049] S202, Quantifying virtual conversion efficiency parameters. The energy utilization efficiency of production equipment directly determines the speed at which electrical energy is converted into production progress. The system defines the virtual charging efficiency. This parameter characterizes the proportion of input electrical energy converted into effective process progress. For electrically heated loads, The thermal efficiency of the corresponding equipment; for motor-driven loads, This corresponds to the electromechanical conversion efficiency. This parameter is obtained by fitting data from the equipment's nameplate or historical operating data.

[0050] S203, construct a virtual self-discharge model based on thermodynamic principles. Unlike an ideal battery, energy dissipation occurs during the production process. For example, a high-temperature furnace loses heat through thermal conduction during shutdown, requiring additional energy for reheating upon resumption of production. The system utilizes the environmental state parameters collected in step S103. and equipment internal status parameters Constructing a virtual self-discharge rate The calculation logic:

[0051] ;

[0052] In the formula, The minimum power threshold required to maintain the operation of the equipment; The overall heat loss coefficient of the equipment is expressed in watts per degree Celsius (W / ℃). These are environmental state parameters. These are the internal status parameters of the equipment. This is the virtual rated capacity for the current task. (This coefficient...) The energy loss due to ambient temperature difference is calculated using the device's surface area and thermal conductivity, or experimentally determined by monitoring the temperature drop curve while the device is off. This model demonstrates that when the device is in standby or interrupted mode, the energy loss is equivalently mapped to the natural decay of the virtual battery's state of charge.

[0053] S204, Establish the virtual battery state evolution equation. Based on the above considerations of capacity, efficiency, and self-discharge characteristics, the system establishes a dynamic state-space equation describing the production progress. Define the virtual state of charge. The production task completion rate is normalized to [0,1][0,1]. The evolution of this state follows the following discrete-time dynamic equation:

[0054] ;

[0055] In the formula, This represents the virtual state of charge at the next moment; The operating power allocated to the production equipment at the current moment; To control the cycle, For virtual charging efficiency, The virtual rated capacity for the current task. The equation represents the virtual self-discharge rate. It equates the operating power of the production equipment to the charging power of a virtual battery, the increase in production progress to the improvement in state of charge, and the downtime loss to the decrease in power caused by self-discharge.

[0056] S205 sets the operational constraint boundaries for virtual energy storage. To ensure the continuity and safety of the production process, boundary conditions need to be set for the charging and discharging behavior of the virtual battery. Specific constraints include upper and lower power limits and state constraints:

[0057] ;

[0058] ;

[0059] In the formula, The minimum sustaining power of the production equipment or zero; The rated maximum power of the equipment. This represents the virtual state of charge at time t. Furthermore, to meet delivery deadlines, a terminal state constraint is set, i.e., at the deadline... At the corresponding moment, the virtual state of charge A preset threshold must be reached. Through the above mapping mechanism, heterogeneous production loads are transformed into virtual energy storage models with capacity, efficiency, self-discharge rate, and power boundaries, thus enabling them to directly participate in subsequent collaborative optimization calculations.

[0060] After completing the virtualization mapping of the production task, in order to reflect discrete delivery date constraints in the continuous energy optimization process, this embodiment of the invention constructs a spatiotemporally coupled congestion risk potential energy surface. This potential energy surface is essentially a nonlinear penalty function embedded in a multidimensional state space. Its function is to transform the risk of the production task not being completed on time into a numerical penalty term in the optimization objective, thereby mathematically forming a gradient field that drives the increase in production progress. This process is described through the following sub-steps S206 to S208:

[0061] S206, Construct the spatiotemporal coupling constraint boundary for the production tasks. The virtual energy storage system must simultaneously satisfy both state and time constraints. The system defines the deadlines for the production tasks. The difference between the current time t and the current time t is the remaining production window. At the same time, define the minimum theoretical time required to complete the remaining tasks. This value equals the remaining uncompleted virtual power. Divide by the maximum operating power of the equipment .

[0062] The system is based on ratios Determine the risk status. When When the value approaches 1, it indicates that the task has entered a critical state; when A value less than 1 indicates a production blockage. This physical boundary parameter is used to guide the subsequent calibration of the potential energy surface function shape parameters.

[0063] S207, Construct the potential energy surface function for the blocking risk. Based on the above physical boundary, establish the potential energy function of the system in the state space. This function explicitly includes time and state variables, forming a numerical surface that dynamically changes over time. Its specific analytical expression is as follows:

[0064] ;

[0065] In the formula, Let time t and state be The potential energy penalty value at that time is incorporated into the subsequent overall optimization objective function. The basic penalty weighting coefficient has physical dimensions consistent with the operating cost term in the objective function (such as monetary units). The numerical setting must meet the following requirements: Under non-emergency conditions, the potential energy value Much lower than the cost of purchasing electricity from the grid; under critical conditions, The order of magnitude should be greater than the product of the maximum electricity price and the maximum power to ensure that the optimization solver prioritizes production schedules. This is a potential energy form factor used to adjust the curvature of the function's growth. The deadline for the production task. This represents the current production task completion status (virtual state of charge). In this embodiment, The value of is related to the average duration of the production task, and is usually set such that when When the value drops to the preset safety threshold, the exponential term begins to increase. This is the numerical stability constant, which takes the value of a very small positive number (e.g., 10). -4 ), used to guarantee when Approaching zero and When the denominator is small, it is non-zero to prevent calculation overflow.

[0066] The mathematical characteristics of this function are as follows: In the early stages of production or when the progress is relatively fast, the denominator is large, and the function value remains in a stable range with a low amplitude. At this time, the controller is mainly guided by economic indicators for optimization. As time goes on, if the progress... If the denominator term fails to increase, its value decreases, and the function value enters the region of exponential nonlinear growth, forming a region of high numerical penalty.

[0067] S208, Calculate the potential-induced optimization gradient. During the rolling optimization process, the optimization algorithm calculates the partial derivatives of the objective function with respect to the control variables. The existence of the blocking risk potential surface causes the system to generate a gradient vector in the solution space pointing in the direction of increasing production progress. Specifically, the potential function with respect to virtual power... The partial derivatives are as follows:

[0068] ;

[0069] In the formula, For virtual conversion efficiency, For production schedule, It is the numerical stability constant. For discrete time steps, For virtual battery capacity, For the penalty function, As a potential energy form factor, The term within parentheses represents the final deadline at which the production task must be completed. Since the term is negative, it indicates that the partial derivative (i.e., the gradient direction) is negative. This means that during the optimization process aimed at minimizing the objective function, this gradient term will influence the control variables. Increase in a positive direction, thereby improving When the system is in the critical region, the absolute value of this gradient increases sharply, dominating the total cost of the optimization objective. This mathematically forces the solver to allocate maximum power to the production equipment to escape the high-penalty region. This mechanism transforms rigid delivery constraints into soft constraints in the objective function, avoiding the problem of traditional hard constraints causing the solver to have no solution under extreme conditions. It also achieves an automatic trade-off between production risks and operating costs.

[0070] See attached document Figure 3 After constructing the generalized state space and completing the virtualization mapping of the production task, this embodiment of the invention quantitatively evaluates the control effect of the system by constructing a multidimensional loss function and a regret value feedback mechanism. This evaluation result is used not only for the current optimization solution but also to drive subsequent two-layer adaptive correction. The specific construction and calculation process is described through the following sub-steps S301 to S304:

[0071] S301, Construct a multidimensional instantaneous loss vector. At each control time t, the system calculates an instantaneous loss vector that includes economic costs, grid interaction stability, and production congestion risk. :

[0072] ;

[0073] In the formula, The formula for the operating economic loss term is as follows: .in, The electricity purchase price at time t, This represents the electricity price (usually a positive value; a negative value from the min function indicates a negative cost, i.e., revenue). Let be the actual grid interaction power at time t. The power grid interaction fluctuation penalty term is given by the following formula: .in, This refers to the baseline interactive power curve generated during the daytime dispatch phase or issued by the superior power grid. This is the actual power exchanged with the power grid at this moment. The value of the blocking risk potential surface function constructed in step S207.

[0074] S302, Generate the weighted scalar total loss function. To transform the multi-objective optimization problem into a single-objective programming problem, the system utilizes a dynamic weight vector. The loss vector is linearly weighted and aggregated into a scalar loss. :

[0075] ;

[0076] In the formula, To cover operating economic loss items, For power grid interaction fluctuation penalty items, The value of the blocking risk potential energy surface function constructed in step S207. Let be the normalized weight vector at time t, satisfying The scalar loss function This serves as the direct optimization objective for subsequent model predictive control.

[0077] S303, Define and calculate the total regret value. The regret value is an indicator of the quality of a control strategy. This embodiment uses a sliding window mechanism to calculate the recent total regret value. :

[0078] ;

[0079] In the formula, ( This is the loss calculation operator, whose input is the control command vector. This operator first calculates the system state response (i.e., the response generated under the input control command) based on the system dynamic equations from steps S101 to S104. and Then, the corresponding scalar loss value is calculated according to step S302. The length of the time window for calculating regret; For a moment The actual set of control commands issued; This is the theoretically optimal set of control instructions. This is obtained by solving a post-hoc deterministic optimization problem, i.e., within a known time window. The control sequence is obtained by solving the problem with the objective of minimizing the cumulative scalar loss within the given load data, photovoltaic output data, and electricity price data.

[0080] S304, decompose external and internal regrets. To achieve targeted two-layer correction, the system generates two types of regret feedback signals. The first type is the external regret feedback signal, used to characterize weight allocation bias. The system calculates the loss function. Regarding the weight vector gradient vector The gradient vector is used as a quantized representation of external regret to indicate the direction of weight updates and reduce future cumulative loss. The second type is the internal regret feedback signal. This is used to characterize the execution deviation caused by parameter mismatch in the virtual model. Its calculation formula is:

[0081] ;

[0082] In the formula, This is the measured value of the virtual state of charge. This value is derived from the physical production progress data collected in step S103. Total number of tasks The ratio is determined, that is The controller at the previous moment based on the current virtual model parameters Predicted time The state value. For internal regret feedback value, when When the deviation signal is non-zero, it drives the adaptive correction of the model parameters.

[0083] See attached document Figure 4 This invention, through an adaptive mechanism in the first channel, dynamically adjusts the weight vector of the objective function using external regret feedback to address strategy deviations caused by electricity price fluctuations or load forecasting errors. This process is based on online convex optimization theory and specifically includes the following steps S401 to S404:

[0084] S401, Obtain the regret gradient feedback data at the current time step. The two-layer adaptive correction module receives the weight vector from the previous time step transmitted by the regret feedback analysis module. and instantaneous loss vector Because the objective function It's about the weight vector. A linear function whose gradient vector with respect to the weights. Numerically equal to the instantaneous loss vector This embodiment employs an online gradient ascent strategy to determine the update direction, that is, for components with large instantaneous loss values ​​(such as those caused by a sudden increase in electricity prices). (Surge), in the next moment, increase its corresponding weight coefficient, thereby increasing the proportion of this component in the overall objective function, and guiding the collaborative optimization solver to prioritize the optimization of this component in subsequent control cycles.

[0085] S402, calculate the pre-update weight vector under unconstrained conditions. The system applies the online gradient ascent algorithm to calculate the intermediate weight vector. This calculation process incorporates a learning rate sequence that decays over time to ensure the convergence of the algorithm:

[0086] ;

[0087] In the formula, This is the intermediate weight vector that has not yet been normalized; This is the weight vector that is currently in effect; It is an instantaneous loss vector that includes economic costs, volatility penalties, and congestion risks; Let be the learning rate step size at time t. The initial learning rate constant is set to match the normalized magnitude (between 0 and 1) of the weight vector. Set to a real number between 0.01 and 0.1. As the number of steps k increases, the learning rate increases accordingly. Attenuation is used to balance the system's response speed to recent changes with its long-term operational stability.

[0088] S403 performs the probabilistic simplex projection operation. Since optimizing the objective function requires the weight vector... The nonnegativity constraint must be satisfied. and normalization constraints The aforementioned intermediate weight vector It is usually not usable directly. The system solves the Euclidean projection optimization problem to... Mapping back to the probabilistic simplex feasible region. The specific solution process uses a sorting threshold algorithm, which includes the following sub-steps: First, the vector... The elements in the array are sorted in descending order of value to obtain a sorted vector. Secondly, find those that meet the conditions. Maximum index value ( ∈{1,2,3})

[0089] in This is the cutoff position for non-zero weights. Let p be the value of the element at position p. Let be the value of the element at position j.

[0090] Then, the Lagrange multiplier threshold is calculated using the following formula:

[0091] ;

[0092] For the first The element value at each position. To determine the optimal cutoff position, finally, according to the formula:

[0093] ;

[0094] Calculate the final updated weight vector, where The threshold value for the Lagrange multipliers. Let be the final weight value of the i-th component that satisfies the constraints. This projection operation guarantees that the updated weight vector is geometrically closest to the gradient update result while strictly satisfying the mathematical constraints.

[0095] S404 outputs and applies the objective function parameters for the next time step. The weight vector is then adjusted by projection. The weights are transmitted to the rolling collaborative optimization module as known constant parameters for constructing the optimization proposition at time t+1. This weight update completes a closed-loop feedback of external regret, enabling the system to automatically adjust its optimization preferences based on the recent loss proportions of each dimension, thereby achieving a dynamic balance across multiple objectives.

[0096] See attached document Figure 5 This embodiment of the invention utilizes a feedback mechanism in the second channel to perform online correction of the deviation between the virtual battery model parameters and the actual characteristics of the physical production equipment. This process is based on the extended Kalman filter (EKF) principle, and uses the internal regret value calculated in step S304... As an observation residual, the energy consumption coefficient per unit product and virtual conversion efficiency Joint estimation is performed. The specific implementation process is described in the following sub-steps S405 to S407:

[0097] S405, construct the parameter sensitivity Jacobian matrix. This is to address the scalar's internal regret value. Mapping to the parameter vector space to be corrected, the system calculates the partial derivative matrix of the observed variables (virtual state of charge) with respect to the model parameters. Define the parameter vector to be estimated. Based on the state evolution equation in step S204, combined with The coupling relationship, the observation Jacobian matrix at time t of system construction. :

[0098] ;

[0099] In the formula, It is a 1×2 dimensional Jacobian matrix; This represents the average operating power applied to the equipment during the previous control cycle. To control cycle duration; This represents the total output of the current batch of tasks. and This represents the estimated prior parameters at the current time. This matrix... The numerical characteristics reflect the information richness of the observed data: when the equipment is operating at power When the matrix elements approach zero, it indicates that no substantial energy conversion has occurred and the observation residual does not contain effective parameter drift information. The correction gain calculated subsequently will automatically decrease, thereby avoiding the introduction of computational noise in the standby state.

[0100] S406, Calculate the adaptive correction gain vector. The system uses the error covariance matrix to calculate the Kalman gain. This gain determines the weights for correcting model parameters using internal regret values:

[0101] ;

[0102] ;

[0103] In the formula, It is a 2×1 dimensional gain vector; It is a 2×2 dimensional error covariance matrix; It is the identity matrix. This is a 1×2 dimensional Jacobian matrix. The tuning principles for the key filter parameters are as follows: R is a scalar of the observation noise variance, its value being set based on the measurement accuracy variance of the piece-counting sensors and meters on the production line (usually 10). -3 Up to 10 -2 (magnitude). The process noise covariance matrix is ​​set as a diagonal matrix. .in and Characterizes the expected rate of drift of physical device parameters over time (typically 10). -6 Up to 10 -5 (the smallest value of the order of magnitude), the existence of this term guarantees the covariance matrix. It will not converge to zero over time, thus ensuring that the algorithm always maintains its ability to track sudden equipment failures or rapid aging. Initial covariance matrix. Set as a large-value diagonal matrix (e.g., 10) This allows for rapid convergence of parameters in the initial stage.

[0104] S407 performs parameter drift correction and boundary constraints. Based on the calculated gain vector and internal regret value, the system performs posterior corrections to the model parameters and generates parameter setpoints for the next control cycle. :

[0105] ;

[0106] In the formula, These are the estimated model parameters for the next control cycle. These are the estimated model parameters for the current control cycle. Here is the gain matrix. This is an internal regret value.

[0107] For the corrected parameters, the system performs a physical constraint check; if the updated virtual conversion efficiency... Exceeding the (0,1] range, or the energy consumption coefficient If the value exceeds the ±20% tolerance range specified on the equipment's nameplate, it will be forcibly truncated to the boundary value. Corrected parameters. The data is transmitted in real time to the virtual battery mapping module to update the virtual rated capacity for the next moment. And the state-space equation coefficients. Through this closed-loop mechanism, when the thermal efficiency of physical equipment decreases due to dust accumulation, the accumulation of internal regret values ​​will drive the estimated value. By gradually reducing the power consumption, the power consumption characteristics of the virtual model can be made closer to the actual physical characteristics, thereby eliminating the state estimation bias under open-loop control.

[0108] See attached document Figure 6 In this embodiment of the invention, at each control time t, based on the update parameters provided by the aforementioned two-layer adaptive mechanism, a finite-time-domain multi-objective collaborative optimization problem (Finite Horizon Optimal Control Problem) is constructed. This problem, by introducing auxiliary decision variables, transforms the nonlinear cost function into a standard quadratic programming model. The specific construction process is described in the following sub-steps S501 to S504:

[0109] S501, Initialize the prediction state and decision variable space. The system reads the current generalized state vector. To accurately describe the differences in battery charging and discharging characteristics and the bidirectional billing characteristics of the power grid, the system defines an extended sequence of decision variables. :

[0110] ;

[0111] In the formula, and They are time points The charging power and discharging power of the physical battery are both non-negative variables. This refers to the operating power of the production line. and They are time points The power purchased and sold by the power grid are both non-negative variables. In this step, the optimizer loads the weight vector determined in step S401. and the model parameters determined by step S407 and , which serves as the known coefficient for constructing the optimization proposition at this moment.

[0112] S502, Constructing a linearized cumulative objective function. To adapt to the quadratic programming solver, the system uses the aforementioned extended variables to transform the original piecewise linear economic cost function into a linear form, and combines this with weighting coefficients to construct the overall objective function. :

[0113] ;

[0114] In the formula, Characterizes the economic cost of operation; Characterizes the penalty for power grid interaction fluctuations; For about Blocking risk items; A smoothing penalty coefficient for adjusting production power. This objective function is achieved by introducing... and Eliminated the original The max / min operators in the calculation ensure that the problem satisfies the mathematical requirements for solving convex optimization. Simultaneously, the power balance constraint is rewritten as a linear equality constraint.

[0115] + ;

[0116] In the formula, To purchase electricity, To sell electricity, For basic electrical load, To control the production load, For energy storage charging power, For energy storage discharge power, This refers to the photovoltaic power generation capacity.

[0117] S503 constructs decoupled physical and virtual state evolution constraints. Based on the decomposed decision variables, the system constructs accurate discrete-time state evolution equations: For physical energy storage batteries, a charge-discharge separation energy integral model is adopted.

[0118] ;

[0119] In the formula, For charging efficiency, The discharge efficiency (usually η<1). Rated total energy capacity Discharge power at any given time , The state of charge of the battery at any given time. The battery's state of charge at any given time is accurately reflected by this formula, which accurately reflects the internal energy loss during discharge. For virtual production batteries, a parameter-adaptive progress model is used.

[0120] ;

[0121] In the formula, for Production completion rate at time +1 for , Let k be the power input into the production process. This represents the total amount of the entire production task. The above equation explicitly includes corrected parameters representing the theoretical energy required to produce one unit of product. and This ensures that the prediction model can keep up with the actual aging state of the physical equipment in real time.

[0122] S504, apply mutual exclusion constraints and termination constraints. To ensure the correctness of the physical meaning and the completion of the task, the following inequality constraints must be applied:

[0123] 1. Physical mutual exclusion and boundary constraints:

[0124] ;

[0125] ;

[0126] In the formula, Let K be the discharge power. The rated maximum power of the energy storage converter (PCS) The minimum permissible state of charge, This represents the upper limit of the maximum permissible state of charge. A binary variable representing the charge / discharge state, used to prevent the battery from charging and discharging simultaneously;

[0127] 2. Virtual terminal constraints:

[0128] ;

[0129] In the formula, To be based on the production batch cutoff time The minimum schedule requirement for the calculation, for The virtual state of charge at time N. If the predicted time domain N covers... Then forced If not covered, then This means that the production schedule must at least meet the average linear rate, thereby preventing the optimizer from delaying tasks in order to save money in the short term.

[0130] See attached document Figure 7 , Figure 7 This is a flowchart of quadratic programming solution and instruction decoupling execution according to an embodiment of the present invention. After constructing the above-mentioned collaborative optimization proposition, the processor executes a numerical solution and instruction conversion program to map the mathematical optimization results into physical control signals. The specific process is described through the following sub-steps S601 to S604:

[0131] S601, the optimization proposition is parameterized and solved. The processor calls the embedded quadratic programming (QP) solver (such as OSQP or an embedded interior-point library) to load the matrix parameters H, f, A, b defined in steps S402 and S404 into the solution workspace. Since the Hessian matrix H of the objective function is used with the introduction of a regularization term... The positive definiteness is guaranteed by the smoothing penalty coefficient for production power adjustment. Under the constraint of satisfying the KKT (Karush-Kuhn-Tucker) conditions, the solver iteratively calculates the optimal decision variable sequence in the entire time domain. .

[0132] S602, extract the first control action in the rolling time domain. Following the rolling optimization principle of Model Predictive Control (MPC), the system extracts the first control action from the sequence. Only the control component corresponding to the current time t is extracted:

[0133] ;

[0134] In the formula, The theoretically optimal charging power at time t is... Let t be the theoretically optimal discharge power. This is a power command for non-energy storage controllable units. This command set is locked as the sole execution target for this control cycle, while the remaining predicted values ​​in the sequence are only used to evaluate future state trajectories and do not directly participate in current device control.

[0135] S603 performs the decoupling mapping between physical and virtual commands. The controller converts the optimization results in the power dimension into control signals specific to each subsystem. For the physical energy storage system, the net power command is calculated according to the formula:

[0136] ;

[0137] In the formula, Let t be the net power that the energy storage system should actually output or absorb at time t. The theoretically optimal charging power at time t is... The theoretically optimal discharge power at time t is given by the instruction sent to the energy storage converter (PCS) via the industrial bus. The internal controller of the PCS controls the IGBT bridge arm operation according to the instruction symbol: a positive value performs AC / DC rectification charging, and a negative value performs DC / AC inverter discharging.

[0138] For virtual production systems, the controller uses the corrected energy consumption coefficient at the current moment. The power command is reverse-mapped into the production line operating rate command:

[0139] ;

[0140] In the formula, Speed ​​setting for the production line conveyor belt (unit: pieces / hour); To optimize the output power setting (unit: kW); Energy consumption coefficient per unit product (unit: kWh / unit); This is a time unit conversion factor (1 if the control cycle matches the energy consumption statistics cycle, and 3600 or other corresponding values ​​if a conversion between seconds and hours is involved). This step ensures that the energy consumed by the actual production cycle of the physical equipment is strictly matched with the charging power of the "virtual battery" planned by the optimizer.

[0141] S604, Instruction Dead Zone Correction and Safety Limiting. For potentially small non-zero values ​​(numerical noise) or values ​​below the device startup threshold that may exist in the mathematically optimized solution, the system performs physical feasibility correction before issuing the instruction: 1. If the calculated... The corresponding power is less than the minimum sustaining power of the production line. If so, the instruction is deemed invalid and forced to be reset. =0, to prevent the equipment from frequently starting and stopping or overheating in the inefficient zone; 2. Calculate the difference between the instruction and the value executed at the previous moment. If the acceleration exceeds the maximum limit of the mechanical transmission system, a smoothed instruction is output according to the maximum permissible slope. The corrected final instruction is written into the PLC register to drive the motor and complete the closed-loop control of this cycle.

Claims

1. A multi-objective adaptive control method for a power generation, grid, load, and storage system in an industrial park, characterized in that, Includes the following steps: S1. Construct a power balance constraint model for the physical energy subsystem and obtain real-time operating status data that includes physical side data and production side data; S2. Construct a generalized state space using the real-time operating status data, establish a virtual battery state equation to map the production task to a virtual energy storage model, and construct a blockage risk penalty function. S3. Calculate the total regret value of the strategy based on the virtual battery state equation, and decompose the total regret value of the strategy into an external regret value that represents the weight bias and an internal regret value that represents the parameter mismatch. S4. Update the objective function weight vector based on the external regret value, and correct the virtual capacity parameters of the virtual energy storage model based on the internal regret value; S5. Construct a collaborative optimization problem using the updated objective function weight vector and the virtual capacity parameter, and incorporate the blocking risk penalty function into the optimization objective to solve for the optimal control sequence; S6. Decouple the first set of control vectors in the optimal control sequence of the optimization objective into physical equipment control commands and production equipment control commands, and issue them.

2. The multi-objective adaptive control method for a source-grid-load-storage system in a park according to claim 1, characterized in that, The steps in step S1 for constructing the power balance constraint model of the physical energy subsystem include: A real-time active power balance condition is established, wherein the algebraic sum of the predicted output value of the renewable energy generation unit, the main grid interaction power value, and the output power value of the electrochemical energy storage unit is equal to the sum of the non-adjustable base load power and the operating power of the adjustable production equipment in the park. It is stipulated that the main grid interaction power value is positive when purchasing power from the grid and negative when feeding power to the grid, and the output power value of the electrochemical energy storage unit is positive when discharging and negative when charging.

3. The multi-objective adaptive control method for a source-grid-load-storage system in a park according to claim 1, characterized in that, The real-time operating status data in step S1 also includes the physical energy storage state. The evolution of the physical energy storage state follows the following rule: the physical battery state of charge at the next moment is equal to the physical battery state of charge at the current moment minus the normalized battery output energy. The battery output energy is determined by the battery output power, time step, and the comprehensive conversion efficiency related to the charge and discharge state.

4. The multi-objective adaptive control method for a source-grid-load-storage system in a park according to claim 1, characterized in that, The steps for establishing the virtual battery state equation in step S2 are as follows: The virtual state of charge is defined as the production task completion rate. The update rule of the virtual state of charge is that the virtual state of charge at the next moment is equal to the virtual state of charge at the current moment plus the effective production progress increment calculated by virtual charging efficiency, and minus the progress loss caused by virtual self-discharge. The progress loss caused by virtual self-discharge is determined based on the comprehensive heat loss coefficient of the equipment, the ambient temperature and the internal temperature of the equipment, and is used to characterize the energy dissipation of the equipment in standby or interrupted state.

5. The multi-objective adaptive control method for a source-grid-load-storage system in a park according to claim 1, characterized in that, The steps for constructing the blocking risk penalty function in step S2 include: The difference between the production task deadline and the current time is calculated as the remaining production window, and the minimum theoretical time required to complete the remaining task is calculated. The risk status is determined based on the ratio of the remaining production window to the minimum theoretical time. A function that dynamically changes with time and production progress is constructed as the blocking risk penalty function. When the production progress lags and the ratio of the remaining production window to the minimum theoretical time approaches the critical threshold, the function value of the blocking risk penalty function increases non-linearly, thereby forming a numerical gradient in the optimization solution space that promotes the increase of production progress.

6. A multi-objective adaptive control method for a source-grid-load-storage system in a park, as described in claim 4, is characterized in that... The step of calculating and decomposing the total regret value in step S3 includes: Based on the sliding window mechanism, the difference between the cumulative loss generated by the actual execution strategy within a set time window and the cumulative loss generated by the theoretically optimal strategy is calculated to obtain the total regret value of the calculated strategy. The gradient vector of the loss function with respect to the weight vector is calculated as the external regret value, which is used to indicate the direction of weight update. The deviation between the measured value of the virtual state of charge and the predicted value obtained based on the model prediction at the previous moment is calculated as the internal regret value, which is used to drive the model parameter correction.

7. A multi-objective adaptive control method for a source-grid-load-storage system in a park, as described in claim 5, is characterized in that... The process of updating the objective function weight vector in step S4 includes: Using the online gradient ascent algorithm, the intermediate weight vector in the unconstrained state is calculated based on the external regret value and the learning rate that decays over time. By solving the projection optimization problem, the intermediate weight vector is mapped to the feasible region that satisfies the non-negativity constraint and the normalization constraint, thus obtaining the updated objective function weight vector. The learning rate decays as the number of running steps increases.

8. The multi-objective adaptive control method for a source-grid-load-storage system in a park according to claim 1, characterized in that, The process of correcting the virtual capacity parameters in step S4 includes: The internal regret value is used as the observation residual to construct a parameter sensitivity matrix. The parameter sensitivity matrix contains the partial derivatives of the observed variables with respect to the unit product energy consumption coefficient and the virtual conversion efficiency. The adaptive correction gain is calculated using the error covariance matrix. The unit product energy consumption coefficient and the virtual conversion efficiency are then corrected a posteriori by combining the internal regret value. The virtual rated capacity in the virtual capacity parameter is updated using the corrected unit product energy consumption coefficient.

9. A multi-objective adaptive control method for a source-grid-load-storage system in a park, as described in claim 1, is characterized in that... The steps in step S5 for constructing the collaborative optimization proposition include: A linearized cumulative objective function is constructed, which accumulates the weighted operating economic cost, grid interaction fluctuation penalty term, and the congestion risk potential surface function value in the prediction time domain, and adds a smoothing penalty term to limit the adjustment range of production power. Constraints are set for the collaborative optimization proposition, including mutual exclusion constraints for charging and discharging of physical energy storage units, state of charge capacity constraints of physical energy storage units, and virtual terminal state constraints requiring the virtual state of charge to reach a preset threshold at the task deadline.

10. A multi-objective adaptive control method for a source-grid-load-storage system in a park, as described in claim 1, is characterized in that... Step S6 includes: Extract the physical charging power, physical discharging power, and production line operating power at the current moment from the optimal control sequence obtained from the optimization objective. Calculate the difference between the physical charging power and the physical discharging power as the net power command and send the net power command to the energy storage converter. Using the corrected unit product energy consumption coefficient at the current moment, divide the production line operating power by the unit product energy consumption coefficient to map it into the production equipment control command. The production equipment control command is the speed setpoint of the production line conveyor belt. After performing dead zone correction and acceleration limiting on the production equipment control command, it is written into the production equipment controller.

Images