Optical storage power station time-of-use electricity price linkage control method and system and storage medium

Abstract

The application discloses a time-of-use electricity price linkage control method and system for a light storage power station and a storage medium, and belongs to the technical field of light storage optimization scheduling. Time series data of electricity price, photovoltaic and load are collected, an electricity price inflection point is identified through difference operation, and a self-adaptive electricity price feature vector is constructed by quantifying fluctuation intensity; a multi-objective rolling optimization model is built by fusing battery nonlinear aging and demand electricity charges, and a storage power benchmark trajectory is solved; the deviation of actual and benchmark power is compared in real time, and when the threshold is exceeded, reinforcement learning is enabled to complete instruction correction, which is issued to the storage converter for smooth execution and local fault locking. The system is hierarchically provided with data acquisition, electricity price extraction, rolling optimization, online correction and execution feedback modules, and uses a working condition synchronous iteration aging model and a reinforcement learning network to solve the defects of traditional schemes, such as fixed peak-valley division, simplified loss model, lack of hierarchical correction and inability to adapt to aging, and is suitable for long-term economic stable operation of industrial and commercial light storage under the power spot market.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic-storage optimization scheduling technology, specifically to a method, system, and storage medium for time-of-use pricing linkage control of photovoltaic-storage power plants. Background Technology

[0002] With the comprehensive advancement of the electricity spot market, photovoltaic (PV) and energy storage (ESS) power plants are gradually shifting from the traditional fixed-time-of-use (OTU) arbitrage model to a refined, coordinated control system based on real-time clearing prices. PV output and user load inherently exhibit strong random fluctuations, coupled with the continuous dynamic changes in spot prices, making traditional fixed-time-of-use energy storage dispatch strategies ill-suited to scenarios with frequent price fluctuations. Commercial and industrial PV-ESS power plants simultaneously face multiple cost constraints, including electricity purchase and sale costs, monthly demand charges, and the aging of energy storage batteries. Relying solely on day-ahead static optimization cannot balance long-term operational benefits with the lifespan of energy storage equipment. There is an urgent need to construct a real-time closed-loop dispatch control scheme that can adapt to price changes and accommodate multiple cost constraints. Existing PV-ESS dispatch control technologies are mainly divided into two categories: model-based rolling optimization methods and model-free reinforcement learning methods. Both types of schemes have inherent limitations, making it difficult to simultaneously ensure planning stability and real-time disturbance adaptability.

[0003] Existing technologies include "Research on Economic Dispatch of Energy Storage for Peak Shaving and Frequency Regulation Based on Dynamic Peak-Valley Period Division," which focuses on dynamic peak-valley identification and achieves adaptive switching of energy storage operation scenarios based on load time-series characteristics. Meanwhile, Selim A et al. published "Adaptive BESS and Grid Setpoints Optimization: A Model-Free Framework for Efficient Battery Management under Dynamic Tariff Pricing" in IEEE Transactions on Sustainable Energy, employing reinforcement learning to achieve real-time energy storage control under dynamic electricity pricing scenarios. These Chinese and English publications respectively explore two approaches: fixed-rule time-series division and model-free intelligent control. The Chinese literature relies on load curves to divide peak-valley intervals but does not utilize real-time electricity market clearing prices for time-series characteristic analysis, lacking methods for quantifying price fluctuation intensity. The foreign literature relies entirely on reinforcement learning to directly output charging and discharging commands, lacking upper-level planning baseline trajectories, which is prone to strategy oscillations. Neither type of research constructed a standardized electricity price feature vector as an optimization input. Most of the energy storage losses adopted simplified linear models, and the nonlinear aging models were not established by coupling the effects of discharge depth, rate, and temperature. The demand electricity cost constraints of industrial and commercial scenarios were generally ignored. There was no hierarchical control architecture based on trajectory deviation threshold-triggered intelligent correction, nor was a closed-loop mechanism for synchronous iteration of the aging model and reinforcement learning network driven by operating data designed. It is difficult to meet the long-term economic operation needs of industrial and commercial photovoltaic and energy storage power stations in the spot market.

[0004] Chinese invention patent (CN120300930A) discloses an intelligent scheduling and control method for an integrated photovoltaic-storage charging and discharging power station. It constructs a three-level control architecture with multi-timescale prediction, optimization, and correction, combining model predictive control to solve the energy storage charging and discharging plan, and employing reinforcement learning to adjust rolling optimization parameters online. This patented solution generates an energy storage power scheduling benchmark based on rolling optimization, possessing multi-timescale collaborative control capabilities. However, the peak and valley periods still rely on manual pre-defined divisions, failing to adaptively update peak and valley boundaries based on daily real-time electricity price patterns, and lacking a mechanism for quantifying price fluctuation intensity. Secondly, reinforcement learning continuously participates in the optimization calculations throughout, resulting in high computational costs and potential issues such as power exceeding limits and violations of safety constraints regarding state of charge. Furthermore, the solution lacks a data-driven dual-model synchronous iteration mechanism; the battery aging model parameters remain fixed and cannot dynamically update with long-term energy storage operating conditions. Optimization accuracy continuously decays during long-term operation, making it unsuitable for long-term operating scenarios where electricity price patterns evolve continuously and energy storage aging characteristics change slowly. Summary of the Invention

[0005] Based on the aforementioned technical problems, this application discloses a time-of-use pricing linkage control method, system, and storage medium for photovoltaic-storage power plants; the method for time-of-use pricing linkage control for photovoltaic-storage power plants includes:

[0006] Obtain real-time electricity market clearing price series, photovoltaic output forecast data, and load demand forecast data for the nodes where the photovoltaic and energy storage power stations are located;

[0007] Based on the real-time electricity market clearing price sequence, by extracting preset price pattern features, the effective peak-valley time period boundary and price fluctuation intensity factor of the day are dynamically identified, and an adaptive time-of-use electricity price feature vector is generated.

[0008] A multi-objective rolling optimization function is constructed, which includes a nonlinear battery aging cost model and a demand electricity penalty term. The adaptive time-of-use electricity price feature vector, photovoltaic power output prediction data and load demand prediction data are used as inputs to solve the energy storage charging and discharging power baseline trajectory within a future preset time window.

[0009] During the real-time control phase, the actual operating status of the photovoltaic-storage power station and the latest market price signal are collected, and the deviation between the actual operating trajectory and the charging and discharging power reference trajectory is calculated. When the deviation exceeds a preset threshold, an online strategy correction based on reinforcement learning agent is triggered, and an adjusted real-time charging and discharging command is output according to the current battery state of charge and safety constraint boundaries.

[0010] The real-time charge and discharge commands are sent to the energy storage converter to perform power regulation, and the actual benefits and battery state changes during the execution process are recorded. The recorded data are used to synchronously update the parameters of the battery nonlinear aging cost model and the value network of the reinforcement learning agent.

[0011] Preferably, the step of dynamically identifying the effective peak-valley time-of-use boundary and price fluctuation intensity factor of the day through preset price pattern feature extraction, and generating an adaptive time-of-use electricity price feature vector, specifically includes:

[0012] The real-time electricity market clearing price sequence is subjected to first-order and second-order difference operations to extract the set of inflection points of the price curve;

[0013] Based on the set of inflection points, an adaptive sliding window clustering algorithm is used to divide the price range, and the ranges that are continuously higher or lower than the dynamic baseline are marked as candidate peak periods and candidate valley periods, respectively.

[0014] Calculate the price volatility intensity factor for each candidate peak period and candidate trough period. The formula is as follows: ,in, For the first Number of sampling points within each candidate time period This is a set of time indices for that time period. for The clearing price at any given moment. and These represent the mean and standard deviation of prices during that period, respectively. This serves as a reference value for the center time of typical historical peak and trough periods. The time decay constant;

[0015] filter Candidate time periods exceeding a preset intensity threshold are designated as valid peak and valley time periods, and the start and end boundaries, duration, and corresponding information of each valid peak and valley time period are recorded. The values ​​are concatenated to form the adaptive time-of-use electricity price feature vector.

[0016] Preferably, the construction of the multi-objective rolling optimization function, which includes a battery nonlinear aging cost model and a demand-based electricity cost penalty term, specifically includes:

[0017] Define the nonlinear aging cost of batteries for: in, Replacement cost per unit capacity battery For rated capacity, for The increment of the depth of discharge at any given time, This is a reference value for the standard depth of discharge cycles. This represents the absolute value of the charging and discharging power. Rated power, These are the aging fitting coefficients. The temperature acceleration factor is given by the formula: , For battery temperature, For optimal operating temperature, Temperature sensitivity coefficient;

[0018] Define demand-based electricity pricing penalties for: in, This is the demand-based electricity price coefficient. For load power, Contribute to photovoltaic power and These are charging power and discharging power, respectively. This serves as the baseline for the declared demand.

[0019] To minimize the total cost over the rolling time domain, a multi-objective rolling optimization function is constructed. : in, To optimize the time domain length for scrolling, for Net cost of electricity purchase and sale at any given time These are the corresponding weighting coefficients.

[0020] Preferably, when solving the reference trajectory of energy storage charging and discharging power within a future preset time window, the following constraints are also introduced:

[0021] Continuity constraint of charged state: ;

[0022] Energy storage power upper and lower limit constraints: ;

[0023] State of charge safety boundary constraints: ;

[0024] in, and These refer to charging efficiency and discharging efficiency, respectively. For time step, and This is a binary variable representing the charging / discharging state. For the maximum allowable power of energy storage, and These are the upper and lower limits for the state of charge, respectively.

[0025] Preferably, the calculation of the deviation between the actual operating trajectory and the charging / discharging power reference trajectory specifically includes:

[0026] Collect the latest Actual charge and discharge power sequence at each time step With the corresponding reference power sequence Calculate the weighted Euclidean distance as the deviation. : in, The time decay weight is given by the formula: , As the attenuation factor, Corresponding to the latest time.

[0027] Preferably, in the online policy correction module based on reinforcement learning agents, the action space of the reinforcement learning agent is defined as the adjustment amount relative to the reference power. And satisfy ,in The limit for energy storage ramp rate;

[0028] The reward function of the reinforcement learning agent Defined as: ,in, for Real-time spot prices To control the time step, To adjust the state of charge of the battery, The state of charge corresponding to the reference trajectory. For state-of-charge tracking penalty coefficient, and These are the upper limit penalty coefficient and the upper limit penalty coefficient, respectively, for exceeding the limit under charged conditions.

[0029] Preferably, the step of synchronously updating the parameters of the battery nonlinear aging cost model and the value network of the reinforcement learning agent using the recorded data specifically includes:

[0030] Online identification of aging cost models using recursive least squares method with aging-sensitized forgetting factor. Parameters, the aging-sensitized forgetting factor The formula is dynamically adjusted based on the cumulative equivalent cycle count of the battery: in, The lower limit of the basal forgetting factor For the deadline The cumulative equivalent number of complete loops at any given moment. The number of characteristic loops is used for calibration;

[0031] Constructing a reinforcement learning value network loss function for the confidence of a coupled aging model The formula is: in, For soft target values, the formula is: , As a discount factor, Here is the entropy regularization coefficient. For policy networks, For the target value network, This represents the aging gradient regularization coefficient. For the aging model parameter vector, Let be the vector of partial derivatives of aging cost with respect to the parameters.

[0032] A time-of-use pricing linkage control system for photovoltaic and energy storage power plants, applicable to any of the time-of-use pricing linkage control methods for photovoltaic and energy storage power plants, includes a data acquisition module, a price feature extraction module, a rolling optimization module, an online correction module, and an online correction module, specifically:

[0033] The data acquisition module obtains the real-time electricity market clearing price sequence, photovoltaic output forecast data, load demand forecast data, and the actual operating status of the photovoltaic-storage power station at the node where the photovoltaic-storage power station is located.

[0034] The electricity price feature extraction module, based on the real-time electricity market clearing price sequence, dynamically identifies the effective peak-valley time-of-use boundary and price fluctuation intensity factor of the day through a preset price pattern feature extraction algorithm, and generates an adaptive time-of-use electricity price feature vector.

[0035] The rolling optimization module constructs a multi-objective rolling optimization function that includes a battery nonlinear aging cost model and a demand electricity penalty term. It takes the adaptive time-of-use electricity price feature vector, photovoltaic output prediction data and load demand prediction data as inputs to solve the energy storage charging and discharging power baseline trajectory within a future preset time window.

[0036] The online correction module is used to calculate the deviation between the actual running trajectory and the charging and discharging power reference trajectory, and when the deviation exceeds a preset threshold, it triggers an online strategy correction based on a reinforcement learning agent, and outputs an adjusted real-time charging and discharging command according to the current battery state of charge and safety constraint boundary.

[0037] The execution feedback module sends the real-time charging and discharging command to the energy storage converter to perform power regulation, and records the actual benefits and battery state change data during the execution process, and synchronously updates the parameters of the battery nonlinear aging cost model and the value network of the reinforcement learning agent.

[0038] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the time-of-use pricing linkage control methods for photovoltaic and energy storage power stations.

[0039] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any one of the time-of-use pricing linkage control methods for photovoltaic and energy storage power stations.

[0040] Compared with the prior art, the technical solution of this application has the following technical effects:

[0041] This invention analyzes real-time electricity price sequences based on price pattern feature extraction algorithms, which can automatically identify effective peak and valley time boundaries and quantify price fluctuation characteristics, forming a standardized adaptive electricity price feature vector input optimization model; the control process can dynamically adjust the interval division rules according to the daily electricity price time sequence changes, get rid of the adaptation limitations brought by the fixed time period division mode, and improve the adaptability of the dispatch strategy to changes in the price pattern of the electricity spot market.

[0042] This invention introduces a multi-dimensional cost model within the rolling optimization framework, and integrates nonlinear battery aging costs and demand electricity costs into a unified solution. Through complete energy storage operation boundary constraints, it ensures that power and state of charge are maintained within the allowable operating range, and outputs a stable and continuous energy storage charging and discharging power benchmark trajectory, providing a stable and reliable basic scheduling reference command for subsequent real-time regulation.

[0043] This invention calculates the power trajectory deviation in real time based on time-weighted Euclidean distance, and initiates the reinforcement learning online correction process only when the deviation exceeds the threshold. During normal operation, scheduling calculations are completed by rolling optimization, and intelligent strategies are activated for adjustment under disturbance conditions to achieve a hierarchical control operation mode, reduce the computing power consumption caused by continuous intelligent optimization, and ensure the stable operation of the control process.

[0044] This invention establishes a dual-model synchronous update mechanism based on actual operating data collected on-site, continuously correcting battery aging model parameters and reinforcement learning value network; the control system can continuously perform self-iterative optimization following the energy storage cycle aging state and external operating conditions, maintaining model accuracy during long-term operation, and constructing a closed-loop photovoltaic-storage linkage control system with adaptive evolution capabilities.

[0045] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0046] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0048] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0049] Figure 1 This is a flowchart illustrating a time-of-use pricing linkage control method for photovoltaic and energy storage power plants.

[0050] Figure 2 A multi-objective rolling optimization architecture diagram for time-of-use pricing linkage control methods for photovoltaic and energy storage power plants;

[0051] Figure 3 A flowchart for online strategy correction using reinforcement learning in a time-of-use pricing linkage control method for photovoltaic and energy storage power plants;

[0052] Figure 4 shows the simulation curves of energy storage charging and discharging power versus time-of-use electricity price under various scheduling models on a typical weekday.

[0053] Figure 5 A schematic diagram comparing the time-series evolution of the State of Charge (SOC) of energy storage during a continuous simulation process throughout the day using multiple comparative models;

[0054] Figure 6 This is a schematic diagram of the power tracking error probability density distribution of the complete control scheme and the ablation control group under the random disturbance of source load.

[0055] Figure 7 This is a schematic diagram of the target command, smoothed reference power, and actual output power tracking waveforms under the discrete PI closed-loop control architecture of an energy storage converter.

[0056] Figure 8 This is a schematic diagram of a time-of-use pricing linkage control system for photovoltaic and energy storage power plants.

[0057] Figure 9 This is a schematic diagram of the electronic equipment for time-of-use pricing linkage control of photovoltaic and energy storage power stations in this embodiment. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0059] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0060] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0061] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0062] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0063] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0064] Example 1

[0065] This embodiment mainly describes a time-of-use pricing linkage control method for photovoltaic and energy storage power stations, such as... Figure 1 As shown, it includes:

[0066] Obtain real-time electricity market clearing price series, photovoltaic output forecast data, and load demand forecast data for the nodes where the photovoltaic and energy storage power stations are located;

[0067] Based on the real-time electricity market clearing price sequence, by extracting preset price pattern features, the effective peak-valley time period boundary and price fluctuation intensity factor of the day are dynamically identified, and an adaptive time-of-use electricity price feature vector is generated.

[0068] A multi-objective rolling optimization function is constructed, which includes a nonlinear battery aging cost model and a demand electricity penalty term. The adaptive time-of-use electricity price feature vector, photovoltaic power output prediction data and load demand prediction data are used as inputs to solve the energy storage charging and discharging power baseline trajectory within a future preset time window.

[0069] During the real-time control phase, the actual operating status of the photovoltaic-storage power station and the latest market price signal are collected, and the deviation between the actual operating trajectory and the charging and discharging power reference trajectory is calculated. When the deviation exceeds a preset threshold, an online strategy correction based on reinforcement learning agent is triggered, and an adjusted real-time charging and discharging command is output according to the current battery state of charge and safety constraint boundaries.

[0070] The real-time charge and discharge commands are sent to the energy storage converter to perform power regulation, and the actual benefits and battery state changes during the execution process are recorded. The recorded data are used to synchronously update the parameters of the battery nonlinear aging cost model and the value network of the reinforcement learning agent.

[0071] At the data preprocessing level, multiple standardized data interaction channels are established. The first channel connects to the regional electricity spot market trading system interface. The communication link adopts a periodic polling mechanism to continuously collect the real-time electricity market clearing price of the node corresponding to the grid connection point of the photovoltaic and energy storage power station according to the system control step size, forming a continuous time-series price sequence. , It represents the time series set consisting of all sampling moments on that day.

[0072] The second channel connects to the photovoltaic power prediction system to obtain the time-by-time photovoltaic power output prediction sequence in the future rolling time domain. The third channel connects to the user load forecasting unit, outputting a load demand forecast sequence within the same time domain. The fourth channel connects to the energy storage battery management system, continuously reading static and dynamic boundary parameters such as individual battery cell voltage, individual cell temperature, total state of charge, and maximum allowable charge / discharge power. All raw data undergoes outlier cleaning, eliminating abrupt changes caused by communication failures or acquisition disturbances using a threshold method based on the difference between adjacent time-series sampling points. Missing sampling points are filled in using linear interpolation between preceding and following time series. After cleaning, all time-series data is uniformly mapped to the same time scale for time alignment. The aligned dataset is then fully transmitted to the electricity price feature analysis level for computation. The system synchronously adds time stamps and data validity flags to all input data; invalid data is directly masked and does not participate in any subsequent optimization calculations. The expression for linear interpolation to fill in missing data is defined as:

[0073]

[0074] in Represents any type of time-series power or price variable. , The effective sampling times are those on the left and right sides of the missing point. This indicates a moment when data is missing.

[0075] The electricity price feature analysis layer performs price pattern feature extraction operations, and calculates the first-order difference operator and the second-order difference operator for each time point of the preprocessed real-time electricity market clearing price sequence;

[0076] Define the first-order difference:

[0077]

[0078] Define second-order difference:

[0079]

[0080] Traverse all time-series sampling points and locate those that meet the conditions. All time points are summarized to form a set of price curve inflection points. The inflection point set divides the entire day's price time series into several continuous independent price intervals. Using each independent interval as the basic unit, an adaptive sliding window clustering algorithm is employed to calculate the price mean within each interval. The interval mean is continuously compared with the dynamic price baseline for the entire day. Intervals above the dynamic baseline are marked as candidate peak periods, and intervals below the dynamic baseline are marked as candidate trough periods. The dynamic price baseline is updated in real time based on the average of all valid price sampling points for the day. The dynamic baseline calculation formula is as follows: In the formula This represents the total number of valid price sampling points for the day.

[0081] Simultaneously remove short-term intervals with a duration less than the minimum effective duration threshold, eliminate invalid candidate periods caused by instantaneous price disturbances, and complete the marking process for each completed period. For each candidate time period, extract all sampling points within that time period to construct a subset of samples and calculate the price volatility intensity factor. , , ;in Representing the Total number of time-series sampling points within each candidate time period This candidate time period contains a set of time indices corresponding to the sampling points. for Real-time clearing prices in the electricity market; The average price over the candidate period. The standard deviation of prices for the candidate period; These are reference values ​​for the center times of typical peak and valley periods stored in the historical database. A preset time decay constant is used to control the decay rate of historical time series reference values. This is applied to each candidate time period. The values ​​are compared with a preset intensity threshold to obtain all valid peak and valley time periods. The start time boundary, end time boundary, duration, and corresponding information of each valid peak and valley time period are extracted. Numerical values ​​are used to sequentially concatenate the above-mentioned multiple time-series features to construct a complete adaptive time-of-use electricity price feature vector. The form of the electricity price feature vector is as follows: In the formula , Representing the first The effective peak and valley time periods are defined by their start and end times. All elements within the vector undergo min-max normalization to eliminate the interference caused by differences in the dimensions of different physical quantities on the solution accuracy of the optimization model. The normalized eigenvector serves as the core input to the downstream multi-objective rolling optimization model.

[0082] After the electricity price feature vector is generated, the scheduling process enters the rolling optimization baseline trajectory solution stage, constructing a battery nonlinear aging cost model. The battery nonlinear aging cost comprehensively couples the discharge depth effect, charge / discharge rate effect, and battery operating temperature effect, as shown in the following formula:

[0083]

[0084]

[0085]

[0086] in The replacement cost per unit capacity of energy storage batteries, This refers to the rated capacity of the energy storage system. for Increment in battery discharge depth at any given moment; This is a reference value for the standard depth of discharge cycle; This represents the absolute value of the charging and discharging power of the energy storage at the current moment. Rated power of energy storage; These are the battery aging fitting coefficients; It is a temperature acceleration factor; The measured average temperature of a single battery cell; This is the optimal operating temperature for the battery; This is a temperature sensitivity coefficient, used to control the aging acceleration rate when the temperature deviates from the optimal value.

[0087] Continue to develop demand charge penalty items adapted for industrial and commercial photovoltaic-storage power station scenarios: , ;in This is the demand-based electricity price coefficient; Real-time net interactive power at the power station's grid connection point; Represents the real-time power of the user load; Represents real-time photovoltaic output; Represents the energy storage charging power; Represents the energy storage discharge power; Baseline demand values ​​reported to the power grid in advance by industrial and commercial users; The operator guarantees that penalty costs are incurred only when the net interaction power exceeds the reporting baseline.

[0088] like Figure 2 As shown, a multi-objective rolling optimization objective function is constructed with minimizing the total comprehensive operating cost within the rolling time domain as the optimization objective:

[0089]

[0090]

[0091]

[0092] in Represents the length of the time domain for scroll optimization; Representing the Net cost of electricity purchase and sale at the power station at any given time; These correspond to the weights of electricity purchase and sale costs, battery aging costs, and demand-based electricity costs, respectively. All weights are greater than zero, and the sum of the three is strictly equal to 1. for Electricity prices will be cleared out by the market at any time.

[0093] Throughout the entire process of solving the objective function, the complete operational constraints of the energy storage system are continuously applied.

[0094] Energy storage state of charge timing continuity constraints: ;

[0095] Upper and lower limits of energy storage charging and discharging power constraints:

[0096]

[0097]

[0098]

[0099] Energy storage state of charge safety boundary constraints:

[0100]

[0101] in Improve energy storage charging conversion efficiency; Energy storage discharge conversion efficiency; The time step represents the unified control time step of the system. and These are binary 0-1 state variables, used to mark the energy storage charging and discharging operation states, respectively; constraints Ensure that energy storage cannot perform charging and discharging operations simultaneously at the same time; This represents the maximum permissible unidirectional charge and discharge power of the energy storage system; This represents the lower limit threshold of the energy storage's state of charge; This represents the upper threshold of the energy storage state of charge. The objective function is numerically solved under all constraints, and the solution outputs the baseline time-series trajectory vector of the energy storage charge and discharge power within a future complete time window. The baseline trajectory is synchronously stored in the global cache and continuously retrieved and used by the real-time monitoring module.

[0102] The scheduling process enters the real-time control and monitoring level. The system continuously collects the actual charging and discharging power uploaded by the energy storage converter, the real-time state of charge of the battery, and the latest real-time market price signals, and continuously reads the most recent continuous data. The actual charge / discharge power sequence corresponding to each control step size Simultaneously, the reference power sequence for the same time period is retrieved from the global cache. .

[0103] Calculation of the deviation between two power time-series trajectories based on the time-weighted Euclidean distance method: , ,in For time-series decay weights, the sum of all weights equals 1; This represents the weighted attenuation factor, which controls the proportion of the contribution of power deviation at historical moments to the overall deviation. For the latest sampling time, the weight value of the time series point that is further away from the current time is lower.

[0104] The results calculated in real time will be continuously processed. The system performs a comparison calculation with a preset deviation threshold. Only when the deviation value continuously exceeds the upper limit of the threshold and remains so for a specified period of time will the system activate the online policy correction module based on the reinforcement learning agent. Under normal operating conditions, the system only relies on the baseline trajectory output by the rolling optimization to complete the power regulation and does not activate the intelligent correction process to avoid unnecessary consumption of computing resources.

[0105] like Figure 3 As shown, the feasible region of actions is predefined within the reinforcement learning agent, and the agent's output variable is defined as the power adjustment amount relative to the reference power. The action space strictly satisfies the boundary constraints: ,in This is the maximum ramp rate limit for the power of the energy storage system. This constraint takes into account both the upper limit constraint of the energy storage power and the power change rate constraint, so as to avoid the single power adjustment range exceeding the physical operating limit of the equipment.

[0106] Reinforcement learning agents evaluate the quality of actions based on a pre-defined reward function, the formula of which is: ;in for Real-time spot electricity prices; The predicted state of charge of the battery after applying a power adjustment. The target charge state corresponding to the baseline scheduling trajectory; This is the penalty coefficient for state of charge tracking; This is the penalty coefficient for exceeding the limit under charged state; This is the penalty coefficient for exceeding the upper limit of the charged state.

[0107] The reinforcement learning agent, combining the current battery state of charge and the safety constraints of all energy storage devices, solves for the optimal power adjustment and superimposes it onto the baseline power to generate the corrected target real-time charge / discharge power command. The command calculation formula is as follows: ,right A boundary truncation check is performed again, using the amplitude limiting operator constraint command within the safe operating range of the equipment. The check formula is as follows: In the formula As a limiting operator, it ensures that the output command is strictly limited to the maximum charge and discharge power range of the energy storage. Within this range, the standardized power command, after verification, is transmitted down to the energy storage converter execution level to complete the pre-verification and output of the control command.

[0108] The power execution and control layer of the energy storage converter is responsible for the entire process of command parsing, operating mode identification, power smooth transition, closed-loop precise control, fault safety interlocking, and real-time data feedback. It is the core hardware control layer for implementing upper-level scheduling commands. The command operating status is parsed using a standardized mode identification formula, and the mode identification rules are as follows: ;

[0109] When the command is determined to be in charging mode, the energy storage converter absorbs surplus power from the photovoltaic side or off-peak power from the grid, and stores it in the battery module after AC-DC conversion. When the command is determined to be in discharging mode, the battery module releases DC power, which is converted into AC power by the converter and supplied to the user load or connected to the grid. When the command is determined to be in standby mode, the converter maintains a zero-power interaction state, only maintaining the basic operating power consumption of the equipment, and the energy storage system does not participate in the power regulation of the grid and load.

[0110] To prevent power surges from impacting the battery and grid, a power smoothing transition control logic is implemented. Based on the equipment ramp rate limit constraint and real-time power changes, the transition power iterative calculation formula is as follows: , ;in The actual output power of the energy storage in the previous control cycle. To achieve a smooth transition target power for the current cycle, This is a sign function used to determine the direction of power adjustment, ensuring that the rate of power change does not exceed the physical tolerance limit of the equipment throughout the process.

[0111] A high-precision discrete proportional-integral closed-loop control circuit is built inside the converter. Using the smooth transition power as the given value and the actual active power output on the AC side as the feedback value, power tracking deviation is eliminated in real time. The discrete PI control algorithm formula is as follows: , ,in To address real-time power tracking deviation, This represents the actual active power output from the AC side of the converter. This is the proportional control coefficient, used for rapid response to power deviation. This is the integral control coefficient, used to eliminate steady-state power error. This is the modulation signal for the power device output by the PI controller.

[0112] Based on the modulation signal, the AC / DC side energy balance regulation of the converter is achieved. The DC side power balance constraint formula is as follows: ,in This represents the real-time power on the DC side of the energy storage system. This refers to the DC bus voltage of the converter. Provide the real-time charging and discharging current for the battery module. The rated energy conversion efficiency of the converter is taken as the positive value under charging conditions and the reciprocal value under discharging conditions to adapt to the energy flow direction.

[0113] The system also features a real-time hardware fault monitoring and safety interlock mechanism. The fault flag discrimination formula is as follows:

[0114] in , These are the upper and lower safety thresholds for the DC bus voltage, respectively. This is the battery's maximum safe operating current. When the fault indicator... Immediately freeze all power adjustment commands, force the issuance of a zero-power standby command, lock the converter's operating state, prevent the escalation of overvoltage and overcurrent faults, and suspend upper-level optimization calculations until the fault flag is cleared and the equipment returns to normal operation. All time-by-time operating power, electrical parameters, and fault status data are timestamped and transmitted back to the global cache in real time, providing a complete operating condition dataset for subsequent model iterations and updates.

[0115] After completing full-time power regulation and data acquisition, the system initiates a dual-model synchronous iterative update process involving both the battery nonlinear aging model and the reinforcement learning network, achieving closed-loop self-evolutionary optimization of the control system. For the battery nonlinear aging cost model, a recursive least squares algorithm with an adaptive forgetting factor is used to dynamically identify core aging parameters online and construct an accumulation formula for the battery's equivalent cycle count. ,in This represents the current number of equivalent complete cycles accumulated by the battery. The equivalent number of cycles for the previous control cycle is accumulated in real time by incrementing the discharge depth at each moment, thus accurately quantifying the degree of battery aging.

[0116] The adaptive forgetting factor is updated based on the equivalent number of cycles, and the formula is as follows:

[0117]

[0118] in The lower limit of the basal forgetting factor The number of characteristic cycles for battery calibration is determined, and the forgetting factor adaptively increases as the battery ages, ensuring a reasonable weight distribution between samples from new and old operating conditions. Based on the updated forgetting factor, a recursive least squares matrix iterative operation is performed. The complete parameter identification iterative equation set is as follows: , , ,in This is the updated battery aging parameter vector at the current moment. For the working condition regression vector, This is a measured value of battery aging loss. Adjust the gain matrix for the parameters. As the covariance matrix, iterative calculations are used to achieve real-time and accurate updates of aging parameters, adapting to the long-term aging and degradation characteristics of batteries.

[0119] Simultaneously, gradient iterative updates are performed on the reinforcement learning value network. A customized loss function is constructed that couples the gradient information of the battery aging model, enabling deep collaborative iteration between the two models. The complete formula for the loss function is as follows: , ,in To strengthen the sample pool for replaying learning experiences, For soft objective value function, As a value discount factor, Here is the entropy regularization coefficient. For the probability distribution of actions in the policy network, The current value network output value, Output value for the target value network. Here is the aging gradient regularization coefficient. Construct the partial derivative matrix of the aging cost function with respect to the aging parameter vector to achieve model gradient coupling, as shown in the formula: ;

[0120] To ensure the timeliness of experience samples, the sample pool adopts a capacity limit constraint and a first-in-first-out update mechanism. The sample update determination formula is as follows: ;in To maximize the capacity of the experience replay sample pool These are the earliest historical operating condition samples stored in the sample pool. Once the sample pool is full, old and invalid samples are automatically removed, and high-quality real-time operating condition samples are continuously updated.

[0121] This embodiment automatically identifies peak-valley boundaries and constructs an electricity price feature vector based on real-time market clearing price sequences. It integrates battery nonlinear aging costs and demand electricity costs to construct a multi-objective rolling optimization model to generate a benchmark trajectory for energy storage power. A trajectory deviation threshold is set as a trigger condition for the reinforcement learning module to achieve on-demand online correction. The aging model parameters and reinforcement learning value network are updated synchronously using field operation data to form a complete self-evolving closed-loop control link, solving the technical problems of insufficient electricity price adaptability, poor constraint security, and inability of the model to continuously adapt and evolve in existing technologies.

[0122] Example 2: This example uses a Python simulation platform to build a digital simulation environment for the joint scheduling of industrial and commercial grid-connected photovoltaic and energy storage systems. The hardware parameters of the simulation system remain constant throughout the process. The photovoltaic installed capacity is 1.2MW, the lithium-ion energy storage system has a rated capacity of 600kW·h and a rated charge / discharge power of 300kW, and the energy storage charging efficiency is [not specified]. Discharge efficiency Constraints on the operating range of energy storage in the state of charge The system uniformly controls the step size. The simulation duration is 24 hours per day. The industrial and commercial load time-series data used in the simulation is the publicly available academic dataset EWELD, published in *Scientific Data*. This dataset includes 15-minute resolution electricity consumption time-series data from 386 industrial and commercial users in southern my country, covering load samples under various seasonal and extreme weather conditions. The simulated electricity price time-series data uses the ERCOT (Electric Power Reliability Council of Texas) publicly available spot market day-ahead node marginal price dataset, selecting the 15-minute resolution clearing price for the ERCOT real-time market load partition. The simulation experiment uses 90 consecutive days of valid time-series samples, randomly partitioned into a 72-day training sample set and an 18-day test sample set.

[0123] To verify the comprehensive advantages of the proposed time-of-use pricing-linked closed-loop control method compared to existing mainstream scheduling algorithms in terms of economy, control accuracy, energy storage lifetime loss, and real-time computing performance, this experiment selected four mature and widely used existing technical solutions as control models: traditional Model Predictive Control (MPC), Deep Deterministic Policy Gradient (DDPG), Soft Actor Commentator (SAC), and Rule-Based Fixed-Time Scheduling (RBS). The traditional MPC scheme uses rolling time-domain linear programming to solve the day-ahead scheduling plan, generating a baseline power curve only based on predicted information, and lacks a real-time feedback correction mechanism. DDPG and SAC are classic model-free deep reinforcement learning scheduling algorithms, using minimizing system operating costs as the reward objective to complete policy training, without introducing a price feature analysis module or power closed-loop smoothing constraints for energy storage converters. RBS is a commonly used engineering benchmark scheme for commercial and industrial photovoltaic-storage projects, relying on fixed peak and valley periods to achieve simple charge-discharge arbitrage of energy storage, and lacks the ability to dynamically adapt to price fluctuations. During the simulation, all comparative models were subjected to identical energy storage power boundary constraints, SOC safety constraints, photovoltaic output time series, load time series, and time-of-use pricing time series. The model hyperparameters were configured according to the standards of the original publicly available literature of various algorithms to ensure the fairness of the comparative experiments. To intuitively quantify the steady-state performance of different control strategies, the steady-state performance summary data of multiple models was obtained based on the simulation results of all simulations over an 18-day test set, and the relevant index values ​​are summarized in Table 1.

[0124] Table 1 Summary of Steady-State Simulation Performance Indicators for Different Scheduling Models

[0125] Scheduling Model TOC (RMB / day) RES (%) RMSE (kW) BDR (%) CTC (ms) RBS 1726.43 62.74 31.72 12.83 0.74 MPC 1483.71 71.39 22.47 9.16 46.38 DDPG 1391.26 75.62 17.35 7.42 16.51 SAC 1347.54 77.18 14.69 6.75 18.26 This application 1251.68 82.46 8.34 4.21 22.73

[0126] TOC is the average daily total operating cost, RES is the photovoltaic self-consumption rate, RMSE is the root mean square error of energy storage power tracking, BDR is the proportion of ineffective cycle loss of energy storage, and CTC is the average solution time for a single control cycle.

[0127] The quantitative indicators in Table 1 provide a clear visual observation of the performance differences among various schemes. RBS relies solely on fixed time sequences for energy storage scheduling, failing to dynamically capture intraday price fluctuations. It exhibits the lowest photovoltaic absorption rate, frequent ineffective charge-discharge cycles in energy storage, and the highest power tracking deviation. MPC obtains a better day-ahead plan through rolling optimization, but lacks a real-time trajectory deviation correction link. When photovoltaic and load prediction deviations occur, the power tracking error increases significantly. Furthermore, online rolling solutions consume substantial computing power, with single-cycle solutions taking significantly longer than reinforcement learning schemes. DDPG and SAC achieve real-time adaptive optimization through reinforcement learning, but both benchmark reinforcement learning models lack a price feature vector extraction link, making it impossible to quantitatively model price inflection points and fluctuation intensity. Additionally, they lack energy storage converter level ramp-up smoothing constraints, easily leading to frequent power oscillations. The energy storage ineffective cycle loss index still has significant room for optimization.

[0128] This application's solution comprehensively outperforms all comparison models in all core indicators. The average daily total operating cost is reduced by 7.09% compared to the optimal baseline model SAC, the photovoltaic self-consumption rate is increased by 5.28 percentage points, the root mean square error of energy storage power tracking is reduced by 43.23%, the proportion of ineffective cycle losses in energy storage decreases by 37.63%, and the solution time remains within the engineering-usable range, balancing optimization effects with real-time control requirements. After completing the quantitative comparison of indicators, based on typical weekday samples from the test set, 24-hour time-series simulation results were extracted, and the real-time charging and discharging power time-series data of energy storage and the time-of-use electricity price sequence for the entire day were compiled, as follows: Figure 4 As shown, the curves are plotted sequentially for RBS, MPC, SAC, and the energy storage power curves corresponding to the method of this application. The time-of-use electricity price curve of the day is superimposed on the figure as a reference benchmark. By comparing the trends of multiple time-of-use curves in the figure, it can be clearly observed that the traditional control scheme has a large number of non-optimal energy storage charging and discharging actions. In some periods, there is reverse scheduling behavior of energy storage discharging during low electricity price periods and energy storage charging during high electricity price periods. The method of this application can accurately follow the time-of-use characteristics of electricity price, only perform charging in the low electricity price period and orderly discharging in the high electricity price period. The power curve changes smoothly and there are no frequent large-amplitude power oscillations, which fully reflects the scheduling strategy optimization effect brought by the electricity price characteristic analysis module.

[0129] At the level of energy storage state of charge evolution, simulation time-series data for the same typical workday are extracted, and all comparative models' daily SOC change time-series data are summarized to plot a daily SOC change comparison curve, such as... Figure 5As shown, the curves simultaneously display the SOC evolution trajectories of four benchmark models and the method of this application. Continuous observation of the time-series trajectory reveals that traditional schemes repeatedly approach the upper and lower safety boundaries of SOC, and in some periods, short-term boundaries are repeatedly touched, which accelerates the aging rate of the battery. The method of this application relies on the nonlinear battery aging cost term embedded in the multi-objective optimization to achieve a smooth change in SOC, reserve sufficient safety margin throughout the process, and avoid the energy storage from operating in the extreme charge range for a long time, thus suppressing battery degradation from the perspective of operation strategy.

[0130] Based on the simulation environment, system hardware parameters, EWELD industrial and commercial load dataset, and ERCOT electricity price time series dataset, more in-depth and refined verification experiments were conducted. These experiments compared the steady-state indices of different algorithms, focusing on ablation experiments of the core innovative modules within this application, robustness testing under dynamic disturbance conditions, and performance verification of the power closed-loop tracking of the energy storage converter. This further demonstrated the technical roles of the electricity price feature analysis unit, the reinforcement learning real-time correction link, and the PCS-level power smoothing control architecture. Four ablation control groups were set up in the ablation experiments: Ablation scheme 1 removed the electricity price inflection point identification and fluctuation intensity factor calculation modules, directly using the original electricity price time series as the optimization input; Ablation scheme 2 retained the electricity price feature analysis module, deleted the real-time trajectory deviation monitoring and reinforcement learning correction branches, and relied solely on rolling optimization to generate a fixed baseline power trajectory; Ablation scheme 3 retained all upper-layer optimization links, removed the energy storage converter power ramping smoothing constraint and discrete PI closed-loop control module, and directly issued and executed the upper-layer optimized power command; Ablation scheme 4 represented the complete closed-loop control architecture of this application, with all functional modules operating normally. The ablation experiments uniformly followed the 90-day dataset partitioning rule and were conducted under completely identical input time sequences. The simulation indicators of all samples in the test set were summarized to obtain the performance comparison results of the ablation experiments, which are presented in Table 2.

[0131] Table 2 Summary of Simulation Performance Indicators for Multiple Ablation Experiments

[0132] Experimental group TOC (RMB / day) RES (%) RMSE (kW) BDR (%) Instruction trace latency (s) Ablation scheme 1 1374.29 76.21 15.37 6.94 2.71 Ablation scheme 2 1326.75 78.53 13.12 5.87 2.68 Ablation scheme 3 1284.37 80.15 11.46 5.13 5.94 Complete protocol (ablation 4) 1251.68 82.46 8.34 4.21 2.76

[0133] Combining ablation experiments with quantitative data allows for a step-by-step clarification of the performance gains brought by each innovative module. Ablation scheme 1 lacks structured electricity price feature vector construction logic, the optimization model cannot identify electricity price inflection points and fluctuation strength, the scheduling strategy does not fully utilize electricity price time-series information, and the most significant increase in operating costs is observed. Ablation scheme 2 relies solely on rolling optimization of the baseline trajectory, lacking a real-time deviation correction mechanism based on time-series weighted Euclidean distance. When there are predicted disturbances in photovoltaics and loads, the upper-level optimization plan cannot dynamically adjust to the on-site operating conditions, resulting in a significant increase in power tracking error. Ablation scheme 3 has a complete upper-level optimization link, but the lower-level execution stage lacks power smoothing transition logic. Power commands are directly issued in a step, requiring the energy storage converter to frequently and significantly adjust its output power, which not only expands the power tracking error but also generates longer command dynamic adjustment delays, while increasing the cycle losses caused by frequent power fluctuations in the battery. The complete scheme has all modules operating collaboratively. The upper layer generates a forward-looking baseline power trajectory based on electricity price feature analysis, the real-time monitoring module continuously compares the baseline trajectory with the actual operating status, and when the deviation threshold is exceeded, a reinforcement learning agent is activated to complete strategy fine-tuning. The lower-level PCS Closed-loop control enables smooth power tracking, and the upper-level optimization decision-making and the lower-level power execution form a complete closed loop, so that all indicators reach the optimal level.

[0134] To verify the robustness of the proposed method under random source-load disturbance scenarios, random Gaussian disturbances with amplitudes of 0 to ±15% were superimposed on the original time-series data for photovoltaic power output and industrial and commercial loads. This simulated the real-world engineering scenario of photovoltaic irradiance fluctuations and random changes in enterprise electricity load. A 30-day simulation of the disturbance conditions was continuously conducted, and time-series samples of the energy storage power tracking deviation between the complete scheme and ablation scheme 2 under the disturbance conditions were extracted. Based on the statistical probability density distribution characteristics of the massive error samples, such as... Figure 6 The power tracking error distribution comparison chart shown is for the disturbance operating conditions, including Figure 6 (a), (b) and (c), the two probability density curves in the figure intuitively show the difference in error distribution characteristics between the two schemes. The ablation scheme 2, which relies solely on rolling optimization, has a wider error distribution range and a significantly higher proportion of large deviation samples. The complete scheme of this application uses the real-time correction branch to continuously correct the scheduling strategy. Under most operating conditions, the power tracking error is concentrated near zero error. It has a stronger adaptive capability in the face of random source and load disturbances, and the control stability is effectively guaranteed.

[0135] A specific simulation was conducted on the power closed-loop execution performance of the energy storage converter. Given a continuously varying target power command sequence, all time-series data of the power commands issued from the upper layer, the converter's smooth transition reference power, and the actual active power output on the AC side were completely recorded. Based on the three sets of time-series waveforms, the PCS power closed-loop tracking simulation waveform diagram was plotted, as shown below. Figure 7 As shown, Figure 7 (a) is the timing sequence of the target power command issued. Figure 7 (b) is the smooth transition reference power after adding the ramp constraint. Figure 7 (c) To compare the smoothed reference power with the actual output active power curve of the AC side of the converter, the discrete PI closed-loop controller can be directly observed to continuously eliminate steady-state tracking deviation based on the timing comparison of the three sets of waveforms. The power change strictly follows the ramp rate limit throughout the process, and there is no power step change phenomenon. If the smoothing transition link is removed, the actual power frequently oscillates significantly, accompanied by obvious overshoot. This fully verifies the supporting role of the power smoothing control logic of the energy storage converter level in reducing the dynamic loss of energy storage and improving the command tracking accuracy.

[0136] Based on the comprehensive simulation results, horizontal multi-model comparisons demonstrate that the closed-loop control method of this application has more outstanding economic efficiency, photovoltaic absorption capacity, and energy storage-friendly operation characteristics compared with existing mainstream scheduling algorithms. Ablation experiments and disturbance condition simulations verify the technical contributions of each core innovative module layer by layer. At the same time, the effectiveness of the underlying execution level mathematical constraints and closed-loop control architecture is verified through energy storage converter-specific tracking simulations. This fully demonstrates that the control method disclosed in this application has engineering feasibility and can be adapted to the actual operation scenarios of industrial and commercial grid-connected photovoltaic and energy storage power stations participating in the electricity spot time-of-use pricing market for optimized scheduling.

[0137] Example 3 describes in detail a time-of-use pricing linkage control system for photovoltaic-storage power stations, such as... Figure 8 As shown, it includes a data acquisition module, an electricity price feature extraction module, a rolling optimization module, an online correction module, and an execution feedback module, specifically:

[0138] The data acquisition module obtains the real-time electricity market clearing price sequence, photovoltaic output forecast data, load demand forecast data, and actual operating status of the photovoltaic and energy storage power station at the node where the power station is located.

[0139] Based on the real-time electricity market clearing price sequence, the electricity price feature extraction module dynamically identifies the effective peak-valley time boundary and price fluctuation intensity factor of the day through a preset price pattern feature extraction algorithm, and generates an adaptive time-of-use electricity price feature vector.

[0140] The rolling optimization module constructs a multi-objective rolling optimization function that includes a nonlinear battery aging cost model and a demand electricity penalty term. It takes the adaptive time-of-use electricity price feature vector, photovoltaic output prediction data and load demand prediction data as inputs to solve the energy storage charging and discharging power baseline trajectory within a future preset time window.

[0141] The online correction module is used to calculate the deviation between the actual running trajectory and the charging and discharging power reference trajectory, and when the deviation exceeds a preset threshold, it triggers an online strategy correction based on a reinforcement learning agent, and outputs an adjusted real-time charging and discharging command according to the current battery state of charge and safety constraint boundary.

[0142] The execution feedback module sends the real-time charging and discharging command to the energy storage converter to perform power regulation, and records the actual benefits and battery state change data during the execution process, and synchronously updates the parameters of the battery nonlinear aging cost model and the value network of the reinforcement learning agent.

[0143] The data acquisition module serves as the system's overall perception entry point, establishing four independent standardized communication channels to connect to the electricity spot trading platform, photovoltaic forecasting terminals, load acquisition terminals, and energy storage BMS and PCS devices, respectively. It continuously collects four types of basic time-series datasets: real-time electricity price time-series datasets at nodes. Photovoltaic prediction series Load forecast sequence Energy storage field operation status parameters, including real-time energy storage charging and discharging power. Battery state of charge Battery temperature DC bus voltage Key identifiers, etc. The module internally incorporates timing alignment and anomaly repair logic, employing methods to address missing timing points. Linear interpolation logic fills in all missing samples, while setting an interpolation threshold to filter out abrupt and abnormal data. All cleaned time series data are uniformly stamped with a synchronization timestamp. After removing invalid labeled samples, the data is completely sent to the downstream electricity price feature extraction module, providing an unbiased and standardized raw data source for feature calculations and avoiding time series misalignment interference caused by asynchronous acquisition from multiple devices from the source.

[0144] The electricity price feature extraction module has completed the cleaning process. The time series is used as the sole input carrier, based on the first-order difference. Second-order difference Complete the set of electricity price inflection points The positioning and splitting process divides the entire day's electricity price curve into independent time-series intervals, and then combines this with dynamic average prices. Candidate peak and trough time periods are divided, and invalid intervals with insufficient duration are eliminated through threshold screening. The fluctuation intensity coefficient is then calculated for each of the remaining valid candidate time periods. The module will display the start and end times of each time period. , Corresponding Ordered concatenation generates a standardized electricity price feature vector. The process involves performing minimum-maximum normalization to eliminate differences in the dimensions of various features. The normalized vector no longer relies on manually preset peak and valley periods and can automatically update the interval division rules according to daily spot electricity price fluctuations. The generated vector... Synchronous binding , The two sets of predicted time series are packaged and transmitted to the rolling optimization module as the core input feature set of the multi-objective function.

[0145] Scroll optimization module receives After determining the photovoltaic and load forecasting time series, a time-domain optimization architecture is built with the goal of minimizing overall operating costs, internally encapsulating... Battery non-linear aging cost, The two core cost calculation logics of demand-based electricity charges and penalties are combined and included in the cost of purchasing and selling electricity. Constituting the overall goal The operation time domain is fixed as Each control step size. The module continuously binds a complete set of energy storage safety constraints throughout the optimization process, including... upper and lower boundaries , Upper limit of charging and discharging power Charge and discharge mutual exclusion binary constraints The recursive relationship of the state of charge is used to complete numerical optimization under all constraint boundary conditions, and output the continuous time-domain energy storage reference power trajectory vector. , The entire data is stored in the global time-series cache, which serves as a baseline for trajectory comparison for the online correction module and as a basis for long-term scheduling planning of the underlying equipment. This ensures that the daily energy storage charging and discharging operations achieve the triple goals of electricity price arbitrage, peak electricity cost suppression, and battery aging suppression.

[0146] The online correction module continuously reads two sets of timing data from the cache in both directions; one set is the rolling optimization output baseline sequence. One set consists of data acquisition modules that transmit the actual operating sequence of the energy storage in real time. The trajectory deviation index is solved in real time using weighted Euclidean distance logic. The module has a built-in static threshold. Continuously perform numerical comparisons. When If the current source load and electricity price fluctuations are determined to be relatively small, the existing data can be directly used. As a control benchmark, no additional intelligent computation is initiated to save computing power; when And while maintaining the specified duration, the system activates the reinforcement learning agent to conduct online strategy fine-tuning, with the agent's action adjustment amount being... The range of motion is affected Climbing speed Dual boundary constraints, combined with real-time State substitution into built-in reward function Complete motion optimization and output the corrected original instructions. Then through The limiting operator receives a safety instruction. The data is then fully distributed to the execution feedback module to carry out underlying power control.

[0147] The execution feedback module undertakes five core functions: instruction parsing, PCS underlying power closed-loop control, fault interlocking, operating condition data feedback, and dual-model synchronous update. It first receives... Complete the identification of three operating modes: charging / discharging, and standby, based on ramp constraints. Generate smooth transition power To avoid power surges impacting the battery; based on discrete PI control logic, Given, AC measured power To provide real-time feedback and eliminate tracking deviations, while continuously monitoring , Generate fault identifier When a fault is confirmed, the module directly switches to zero-power standby protection mode. The module continuously collects time-series samples throughout the entire control process, including time-by-time power... Changes, periodic purchase and sale revenue, battery temperature The timestamp is sent back to the global cache, which is then used to supply two sets of model iteration logic: one of which uses an adaptive forgetting factor. Recursive least squares algorithm for updating aging parameter vector Relying on the cumulative equivalent number of iterations Dynamically adjust sample weights; its two-way reinforcement learning sample pool Store the new operating condition tuple, and apply the loss function based on the coupled aging gradient. The parameters of the policy network and value network are updated synchronously, and all updated model parameters are synchronously fed back to the rolling optimization module and the online correction module to participate in the next round of cycle calculation, forming a complete self-learning closed-loop link.

[0148] This application further provides a computer device, such as... Figure 9 As shown, the system includes a processor and a memory connected in communication with the processor. The memory stores computer programs that can run on the processor. When the processor executes the program, it can implement the aforementioned time-of-use pricing linkage control method for photovoltaic-storage power stations. The processor can complete the entire process of calculation, including electricity price feature extraction, rolling optimization solution, power deviation monitoring, and energy storage converter instruction calculation. It receives various time-series data and outputs control instructions. Based on hardware resources, it implements multi-level scheduling logic operation, providing hardware support for real-time optimization and control of photovoltaic-storage power stations.

[0149] This application further provides a computer-readable storage medium storing a computer program. When executed by a processor, the program performs the time-of-use pricing linkage control method for photovoltaic-storage power stations disclosed in this application. This storage medium can permanently store the control algorithm program code and can be mounted on hardware platforms such as industrial control computers and edge controllers. After the program runs, it can achieve multi-objective optimization, power smoothing calculation, fault diagnosis, and data interaction functions, facilitating program portability, deployment, and reuse. It eliminates the limitations of dedicated hardware and broadens the application scenarios for the engineering implementation of this control scheme.

[0150] This embodiment completes the electricity price analysis, optimization planning, real-time correction and underlying power regulation in a layered and collaborative manner. Based on the bidirectional data closed loop, the aging model and reinforcement learning network are updated synchronously, which can adapt to the fluctuation of spot electricity prices, effectively reduce energy storage cycle losses and industrial and commercial demand costs, and the power regulation is smooth and shock-free. The overall scheduling economy, operation stability and equipment adaptability are significantly improved.

[0151] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A time-of-use pricing linkage control method for photovoltaic-storage power stations, characterized in that, Includes the following steps: Obtain real-time electricity market clearing price series, photovoltaic output forecast data, and load demand forecast data for the nodes where the photovoltaic and energy storage power stations are located; Based on the real-time electricity market clearing price sequence, by extracting preset price pattern features, the effective peak-valley time period boundary and price fluctuation intensity factor of the day are dynamically identified, and an adaptive time-of-use electricity price feature vector is generated. A multi-objective rolling optimization function is constructed, which includes a nonlinear battery aging cost model and a demand electricity penalty term. The adaptive time-of-use electricity price feature vector, photovoltaic power output prediction data and load demand prediction data are used as inputs to solve the energy storage charging and discharging power baseline trajectory within a future preset time window. During the real-time control phase, the actual operating status of the photovoltaic-storage power station and the latest market price signal are collected, and the deviation between the actual operating trajectory and the charging and discharging power reference trajectory is calculated. When the deviation exceeds a preset threshold, an online policy correction based on a reinforcement learning agent is triggered, and an adjusted real-time charge and discharge command is output according to the current battery state of charge and safety constraint boundaries. The real-time charge and discharge commands are sent to the energy storage converter to perform power regulation, and the actual benefits and battery state changes during the execution process are recorded. The recorded data are used to synchronously update the parameters of the battery nonlinear aging cost model and the value network of the reinforcement learning agent.

2. The time-of-use pricing linkage control method for photovoltaic-storage power stations according to claim 1, characterized in that, The process involves extracting preset price pattern features, dynamically identifying the boundaries of effective peak and valley periods and price fluctuation intensity factors for the day, and generating an adaptive time-of-use electricity price feature vector. Specifically, this includes: The real-time electricity market clearing price sequence is subjected to first-order and second-order difference operations to extract the set of inflection points of the price curve; Based on the set of inflection points, an adaptive sliding window clustering algorithm is used to divide the price range, and the ranges that are continuously higher or lower than the dynamic baseline are marked as candidate peak periods and candidate valley periods, respectively. Calculate the price volatility intensity factor for each candidate peak period and candidate trough period. The formula is as follows: ,in, For the first Number of sampling points within each candidate time period This is a set of time indices for that time period. for The clearing price at any given moment. and These represent the mean and standard deviation of prices during that period, respectively. This serves as a reference value for the center time of typical historical peak and trough periods. The time decay constant; filter Candidate time periods exceeding a preset intensity threshold are designated as valid peak and valley time periods. The start and end boundaries, duration, and corresponding information for each valid peak and valley time period are then recorded. The values ​​are concatenated to form the adaptive time-of-use electricity price feature vector.

3. The time-of-use pricing linkage control method for photovoltaic-storage power stations according to claim 2, characterized in that, The construction of the multi-objective rolling optimization function, which includes a battery nonlinear aging cost model and a demand-based electricity cost penalty term, specifically includes: Define the nonlinear aging cost of batteries for: in, Replacement cost per unit capacity battery For rated capacity, for The increment of the depth of discharge at any given time, This is a reference value for the standard depth of discharge cycles. This represents the absolute value of the charging and discharging power. Rated power, These are the aging fitting coefficients. The temperature acceleration factor is given by the formula: , For battery temperature, For optimal operating temperature, Temperature sensitivity coefficient; Define demand-based electricity pricing penalties for: in, This is the demand-based electricity price coefficient. For load power, Contribute to photovoltaic power and These are charging power and discharging power, respectively. This serves as the baseline for the declared demand. To minimize the total cost over the rolling time domain, a multi-objective rolling optimization function is constructed. : in, To optimize the time domain length for scrolling, for Net cost of electricity purchase and sale at any given time These are the corresponding weighting coefficients.

4. The time-of-use pricing linkage control method for photovoltaic-storage power stations according to claim 3, characterized in that, When solving for the reference trajectory of energy storage charging and discharging power within a future preset time window, the following constraints are also introduced: Continuity constraint of charged state: ; Energy storage power upper and lower limit constraints: ; State of charge safety boundary constraints: ; in, and These are charging efficiency and discharging efficiency, respectively. For time step, and This is a binary variable representing the charging / discharging state. For the maximum allowable power of energy storage, and These are the upper and lower limits for the state of charge, respectively.

5. The time-of-use pricing linkage control method for photovoltaic-storage power stations according to claim 1, characterized in that, The calculation of the deviation between the actual operating trajectory and the charging / discharging power reference trajectory specifically includes: Collect the latest Actual charge and discharge power sequence at each time step With the corresponding reference power sequence Calculate the weighted Euclidean distance as the deviation. : in, The time decay weight is given by the formula: , As the attenuation factor, Corresponding to the latest time.

6. The time-of-use pricing linkage control method for photovoltaic-storage power stations according to claim 5, characterized in that, In the online policy correction module based on reinforcement learning agents, the action space of the reinforcement learning agent is defined as the adjustment amount relative to the reference power. And satisfy ,in The limit for energy storage ramp rate; The reward function of the reinforcement learning agent Defined as: ,in, for Real-time spot prices To control the time step, To adjust the state of charge of the battery, The state of charge corresponding to the reference trajectory. For state-of-charge tracking penalty coefficient, and These are the upper limit penalty coefficient and the upper limit penalty coefficient, respectively, for exceeding the limit under charged conditions.

7. The time-of-use pricing linkage control method for photovoltaic-storage power stations according to claim 6, characterized in that, The method of synchronously updating the parameters of the battery nonlinear aging cost model and the value network of the reinforcement learning agent using recorded data specifically includes: Online identification of aging cost models using recursive least squares method with aging-sensitized forgetting factor. Parameters, the aging-sensitized forgetting factor The formula is dynamically adjusted based on the cumulative equivalent cycle count of the battery: in, The lower limit of the basal forgetting factor For the deadline The cumulative equivalent number of complete loops at any given moment. The number of characteristic loops is used for calibration; Constructing a reinforcement learning value network loss function for the confidence of a coupled aging model The formula is: in, For soft target values, the formula is: , As a discount factor, Here is the entropy regularization coefficient. For policy networks, For the target value network, This represents the aging gradient regularization coefficient. For the aging model parameter vector, Let be the vector of partial derivatives of aging cost with respect to the parameters.

8. A time-of-use pricing linkage control system for photovoltaic-storage power plants, applicable to the time-of-use pricing linkage control method for photovoltaic-storage power plants as described in any one of claims 1-7, characterized in that, It includes a data acquisition module, an electricity price feature extraction module, a rolling optimization module, an online correction module, and an online adjustment module, specifically: The data acquisition module obtains the real-time electricity market clearing price sequence, photovoltaic output forecast data, load demand forecast data, and the actual operating status of the photovoltaic-storage power station at the node where the photovoltaic-storage power station is located. The electricity price feature extraction module, based on the real-time electricity market clearing price sequence, dynamically identifies the effective peak-valley time-of-use boundary and price fluctuation intensity factor of the day through a preset price pattern feature extraction algorithm, and generates an adaptive time-of-use electricity price feature vector. The rolling optimization module constructs a multi-objective rolling optimization function that includes a battery nonlinear aging cost model and a demand electricity penalty term. It takes the adaptive time-of-use electricity price feature vector, photovoltaic output prediction data and load demand prediction data as inputs to solve the energy storage charging and discharging power baseline trajectory within a future preset time window. The online correction module is used to calculate the deviation between the actual running trajectory and the charging and discharging power reference trajectory, and when the deviation exceeds a preset threshold, it triggers an online strategy correction based on a reinforcement learning agent, and outputs an adjusted real-time charging and discharging command according to the current battery state of charge and safety constraint boundary. The execution feedback module sends the real-time charging and discharging command to the energy storage converter to perform power regulation, and records the actual benefits and battery state change data during the execution process, and synchronously updates the parameters of the battery nonlinear aging cost model and the value network of the reinforcement learning agent.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a time-of-use pricing linkage control method for photovoltaic and energy storage power stations as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a time-of-use pricing linkage control method for photovoltaic and energy storage power stations as described in any one of claims 1 to 7.

Images