Photovoltaic output smooth regulation and control method, system, equipment and medium

Abstract

The invention belongs to the technical field of photovoltaic power generation, and discloses a photovoltaic output smooth regulation and control method, system, equipment and medium, and the method comprises the steps: firstly carrying out the low-pass filtering of an original photovoltaic output signal, generating a reference motion, inputting a state vector to a pre-trained PPO reinforcement learning model, outputting a residual motion, carrying out the superposition, and forming a synthetic motion instruction, and then correcting by a power limit and SOC boundary constraint mechanism, and outputting by a teacher model. And based on the distillation loss function, training a lightweight student model by taking teacher output as a soft label for real-time regulation and control. According to the method, low-pass filtering provides a stable reference, learning residual error compensation high-frequency fluctuation and non-stationary characteristics are reinforced, and a constraint mechanism protects the service life of the battery; complex model knowledge is migrated to a lightweight model in the distillation process, so that the calculation burden is reduced. By adopting the method, the smoothness performance is remarkably improved, the control failure is avoided, the battery health is effectively protected, the efficient edge deployment is realized, and the control performance, the service life management and the engineering feasibility can be considered.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of photovoltaic power generation, and particularly relates to a photovoltaic output smoothing control method, system, device and medium. BACKGROUND

[0002] With the global energy structure transforming to clean and low-carbon, photovoltaic power generation, as an important part of renewable energy, has achieved rapid expansion in installed capacity. However, photovoltaic power generation output is significantly affected by natural factors such as light intensity, cloud shadow, and weather changes, and has prominent time-varying, volatility, and uncertainty. This characteristic has become a major challenge to the safe and stable operation of the power grid. Short-term and large fluctuations in photovoltaic output not only cause power quality problems such as power grid voltage fluctuation, frequency deviation, and harmonics, but also significantly increase the operating cost of power grid frequency regulation, standby, and peak shaving. In extreme cases, it may even lead to the phenomenon of light and electricity being abandoned, causing serious waste of valuable energy resources. In order to effectively smooth the fluctuations in photovoltaic output, battery energy storage systems (BESS) have become a key technical means recognized by the industry due to their flexible charge and discharge adjustment capabilities. By buffering the instantaneous output changes of photovoltaic power, they can significantly improve the grid stability of photovoltaic power generation and reduce the impact on transformer, transmission line, and other downstream power grid facilities.

[0003] Current technical solutions for photovoltaic output smoothing control have many shortcomings and cannot meet the comprehensive needs of engineering practice. Traditional filtering and rule / strategy-based energy storage control methods, although they generate a smoothed reference by moving average, low-pass filtering, and other methods, and execute control in combination with rules such as state of charge (SOC) upper and lower limits, have the advantages of easy implementation and strong interpretability, but are limited by constraints such as battery power upper limit and energy capacity. Therefore, the filtered reference power cannot always be executed by the battery, resulting in insufficient smoothing effect or control action failure. Moreover, these methods do not fully consider factors such as battery cycle depth and charge / discharge frequency that affect the life, which can accelerate battery aging. Control methods for energy storage systems introduce SOC and charge / discharge depth constraints to alleviate battery wear, but they are limited in control performance in extreme weather or high-frequency fluctuation scenarios due to the lack of effective characterization of the non-stationary characteristics of photovoltaic output. In recent years, intelligent control methods have emerged, but model predictive control relies on photovoltaic output and weather prediction results, which requires high prediction accuracy and model fitting. It has problems such as large computational overhead and difficulty in tuning in strong nonlinear and complex constraint scenarios. Reinforcement learning-based methods generally face the difficulties of slow training convergence, poor strategy interpretability, and high computational demand, making it difficult to directly deploy and land on edge devices in photovoltaic power stations.

[0004] Therefore, the existing photovoltaic output smoothing and energy storage control technology cannot effectively balance the control performance, battery life and edge deployment feasibility in engineering practice. SUMMARY

[0005] The application provides a photovoltaic output smoothing regulation method, system, device and medium, which can effectively solve the problem that the existing photovoltaic output smoothing and energy storage control technology cannot effectively balance the control performance, battery life and edge deployment feasibility in engineering practice.

[0006] In order to achieve the above purpose, the application adopts the following technical solutions: A photovoltaic output smoothing regulation method, comprising: Performing low-pass filtering processing on the original photovoltaic output signal to generate a reference action; Inputting the original state space vector into a reinforcement learning agent of a pre-constructed teacher model to output a residual action; wherein the reinforcement learning agent adopts an algorithm based on PPO; Superimposing the reference action and the residual action to generate a synthesized action instruction; Correcting the synthesized action instruction by using a constraint mechanism of power limit and SOC boundary to generate a final action instruction; the final action instruction is output by the teacher model; Training a student model based on a distillation loss function, taking the final action instruction output by the teacher model as a soft label, to obtain a lightweight regulation model; Regulating the real-time photovoltaic output signal by using the lightweight regulation model to output a real-time regulation instruction.

[0007] Further, the low-pass filtering processing on the original photovoltaic output signal to generate a reference action comprises: Obtaining the original photovoltaic output signal; Filtering the original photovoltaic output signal by using a low-pass filter to generate a reference action, and a specific first-order discrete filtering formula is as follows:

[0008] In the formula, Refers to the reference action; Refers to the filtering coefficient; Refers to the original photovoltaic output signal; Refers to the current time; Refers to the last time.

[0009] Further, the inputting the original state space vector into a reinforcement learning agent of a pre-constructed teacher model to output a residual action comprises: obtaining an original state space vector; wherein the original state space vector comprises an environment state and a battery state of charge; inputting the original state space vector into a reinforcement learning agent of a pre-constructed teacher model, and outputting a residual action, wherein: A specific formula of the optimization objective function of the reinforcement learning agent is as follows:

[0010] In the formula, represents the optimization objective function; represents a ratio of new and old policies; represents an advantage function estimate; represents a clipping range; represents an expectation; represents a numerical clipping function; represents a minimum function; represents a policy parameter; represents a current time; The state space vector A specific expression of the state space vector is as follows:

[0011] In the formula, represents a change amount of the original photovoltaic output signal; represents a battery state of charge; represents a reference action; represents a final action instruction; represents a historical state amount of photovoltaic power; represents a historical state amount of a photovoltaic power change amount; represents a current time; represents a previous time.

[0012] Further, a specific calculation formula of the step of superimposing the reference action and the residual action to generate a synthesized action instruction is as follows:

[0013] In the formula, represents a synthesized action instruction; represents a reference action; represents a residual action; represents a current time.

[0014] Further, in the step of correcting the synthesized action instruction by using a power limit and an SOC boundary constraint mechanism to generate a final action instruction, a specific correction formula is as follows:

[0015]

[0016]

[0017] In the formula, represents the final action instruction; represents the synthesized action instruction; represents the maximum allowed discharge power; represents the maximum allowed charging power; represents the rated limit of battery power; represents the minimum function; represents the maximum function; represents the battery state of charge; and respectively represent the minimum and maximum state of charge allowed by the battery; represents the rated capacity of the battery; represents the control cycle time step, represents the current time.

[0018] Further, in the step of correcting the synthesized action instruction using the power limit and SOC boundary constraint mechanism to generate the final action instruction, an SOC control factor is also introduced in the reward function to correct the residual action; wherein the cumulative loss of SOC is calculated by the rainflow counting method, and the cumulative loss The specific calculation formula is as follows:

[0019] In the formula, represents the total number of cycles identified by the rainflow counting; represents the depth of the th charge and discharge cycle; represents the number of occurrences of the th charge and discharge cycle; represents the number of rated life cycles of the battery under the cycle depth .

[0020] Further, the light-weight regulation and control model is obtained by training the student model based on the distillation loss function, taking the final action instruction output by the teacher model as a soft label, comprising: inputting the original photovoltaic output signal into the pre-constructed student model to obtain the output result of the student model; taking the final action instruction output by the teacher model as a soft label for training the student model; wherein the final action instruction is the output result of the teacher model; training the student model by using the distillation loss function in combination with the output result of the teacher model and the output result of the student model to obtain the light-weight regulation and control model, wherein: The distillation loss function The specific expression is as follows:

[0021] In the formula, is the number of samples, and the value is , is the total number of training samples; represents the first output result of the student model; represents the first output result of the teacher model; represents the first output result of the student model; represents the first output result of the teacher model; represents the student model parameter.

[0022] A photovoltaic output smoothing regulation system, comprising: A filtering module for performing low-pass filtering processing on an original photovoltaic output signal to generate a reference action; A reinforcement learning module for inputting an original state space vector into a reinforcement learning agent of a pre-constructed teacher model to output a residual action; wherein the reinforcement learning agent adopts an algorithm based on PPO; An action superposition module for superimposing the reference action and the residual action to generate a synthesized action instruction; A correction module for correcting the synthesized action instruction by using a constraint mechanism of power limit and SOC boundary to generate a final action instruction; the final action instruction is output by the teacher model; A model training module for training a student model based on a distillation loss function, taking the corrected final action instruction output by the teacher model as a soft label, to obtain a lightweight regulation model; A regulation module for regulating a real-time photovoltaic output signal by using the lightweight regulation model to output a real-time regulation instruction.

[0023] A photovoltaic output smoothing regulation device, comprising: A memory for storing a computer program; A processor for executing the computer program to realize the steps of the above-mentioned photovoltaic output smoothing regulation method.

[0024] A computer readable storage medium, the computer readable storage medium stores a computer program, the computer program is executed by the processor to realize the steps of the above-mentioned photovoltaic output smoothing regulation method.

[0025] Compared with the prior art, the present application has the following beneficial effects: This invention provides a method for smoothing photovoltaic (PV) power output control. First, the original PV power output signal is low-pass filtered to generate a baseline action. This baseline action is then input into a pre-trained PPO reinforcement learning model, where it is fed into the output residual action. These residual actions are then superimposed to form a synthetic action command, which is further corrected by power limit and SOC boundary constraints. The final action command is then output by the teacher model. Subsequently, a lightweight student model is trained using the teacher output as a soft label based on a distillation loss function for real-time control. In this method, low-pass filtering provides a stable baseline, reinforcement learning residuals compensate for high-frequency fluctuations and non-stationary characteristics, and the constraint mechanism protects battery life. The distillation process transfers knowledge from complex models to a lightweight model, reducing computational burden. This method significantly improves smoothing performance, avoids control failure, effectively protects battery health, and achieves efficient edge deployment, balancing control performance, lifespan management, and engineering feasibility. Attached Figure Description

[0026] Figure 1 A schematic diagram illustrating the principle of a photovoltaic power output smoothing control method provided in an embodiment of the present invention; Figure 2 A flowchart illustrating the implementation of a photovoltaic power output smoothing control method provided in an embodiment of the present invention; Figure 3 A flowchart of a photovoltaic power output smoothing control method provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a photovoltaic power output smoothing control system provided in an embodiment of the present invention. Detailed Implementation

[0027] To further understand the content of this invention, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention.

[0028] The technical terms involved in this invention are explained below: SOC: State of Charge; SOC is a core parameter for measuring the remaining usable capacity of a battery, representing the percentage of the battery's total capacity when fully charged.

[0029] PPO stands for Proximal Policy Optimization, a reinforcement learning algorithm that is an improved version of the policy gradient method.

[0030] LPF: Short for Low-Pass Filter, it is a filter that allows low-frequency signals to pass through while suppressing high-frequency signals.

[0031] BESS stands for Battery Energy Storage System.

[0032] As mentioned in the background section, existing photovoltaic (PV) power plants typically rely on battery energy storage systems for output smoothing due to the intermittent and fluctuating nature of solar energy, thus maintaining grid operational safety. However, traditional methods have limitations in both efficiency and control effectiveness. While filtering methods, such as low-pass filtering, are simple to implement and computationally inexpensive, their performance is severely limited when considering multiple constraints such as battery power, energy capacity, and remaining charge thresholds, often failing to guarantee a consistently effective smoothing effect. Furthermore, these methods do not adequately consider the depth and frequency of battery charging and discharging, easily leading to frequent start-ups and shutdowns, thereby shortening battery life. To overcome the shortcomings of traditional filtering methods, reinforcement learning has been gradually introduced into PV output smoothing control. Its advantages lie in its adaptive and dynamic optimization capabilities, enabling it to learn optimal strategies from historical experience and environmental feedback without relying on explicit mathematical models, thus theoretically better addressing the random fluctuations in PV power and various complex nonlinear constraints. However, directly employing end-to-end reinforcement learning methods still has several limitations: first, the training convergence speed is slow, making it difficult to quickly obtain usable policies; second, the lack of interpretability makes policy adjustment and engineering verification difficult; and third, the high model complexity and computational requirements make it difficult to deploy directly on power plant edge devices and respond quickly to control needs. Therefore, the current challenge is to design a photovoltaic power output smoothing control method that combines the stability of filtering methods with the adaptability of reinforcement learning, while considering both battery capacity and lifespan, and possesses both interpretability and convergence efficiency, allowing for efficient deployment in practical engineering.

[0033] To achieve the aforementioned objectives, this embodiment provides a photovoltaic power output smoothing control method. This method offers a photovoltaic power output smoothing control framework combining LPF prior and reinforcement learning residual correction. A filter provides a stable and reliable benchmark action, and the action output by the reinforcement learning agent is used as a residual term to achieve fine-grained action correction. This ensures both the interpretability and convergence efficiency of the control process, while also guaranteeing physical feasibility at the action level through power limits and SOC constraints. Simultaneously, the reward function design incorporates power ramp-up constraints and SOC health factors (control factors) as dual considerations to avoid frequent deep charge-discharge cycles, thereby extending battery life. Furthermore, this invention utilizes a "teacher-student" policy distillation mechanism to compress complex residual reinforcement learning strategies into lightweight end-to-end student models, enabling independent deployment in real-world power plants with limited computing resources. This achieves rapid inference and control, balancing performance, reliability, and engineering feasibility.

[0034] like Figure 3As shown, this embodiment provides a method for smooth control of photovoltaic power output, including: The original photovoltaic output signal is low-pass filtered to generate a reference action; The original state space vector is input into the reinforcement learning agent of the pre-constructed teacher model, and the residual action is output; wherein, the reinforcement learning agent adopts the PPO-based algorithm; The baseline action and the residual action are superimposed to generate a synthetic action command; The synthesized motion commands are modified using a constraint mechanism based on power limits and SOC boundaries to generate the final motion commands; the final motion commands are output by the teacher model. Based on the distillation loss function, the corrected final action command output by the teacher model is used as a soft label to train the student model, so as to obtain a lightweight control model. A lightweight control model is used to regulate the real-time photovoltaic power output signal and output real-time control commands.

[0035] The control method provided in this embodiment will be further explained below with reference to the accompanying drawings: like Figure 1 As shown, Figure 1 The diagram below illustrates the principle of the photovoltaic output smoothing control method provided in this embodiment. As can be seen, the entire architecture consists of an LPF module, a residual reinforcement learning module, and a constraint mechanism module. The LPF module processes the original photovoltaic output signal (original photovoltaic power signal) to generate a baseline action, ensuring the overall output has basic stability. The residual reinforcement learning module outputs a correction based on the baseline action and explicitly incorporates a SOC health factor (SOC control factor) into the reward function to compensate for the shortcomings of the filtering method, making the control more refined, thereby delaying battery aging and extending its service life, and possessing stronger dynamic adaptive capabilities. The constraint mechanism module applies battery power limits and SOC boundary conditions to the synthesized action commands, ensuring that all output synthesized action commands are physically executable. Furthermore, this embodiment adopts a two-stage "teacher-student" mechanism in its training strategy. The teacher model is trained based on the Proximal Policy Optimization (PPO) reinforcement learning algorithm and is obtained by training with complete features (including filtering and residual information). The student model learns the final action of the teacher model through policy distillation with simplified input. Compared with the teacher model, the student model adopts a smaller structure, which greatly reduces the number of parameters and significantly improves the inference speed, making it more suitable for deployment on edge devices in photovoltaic power plants.

[0036] It is evident that the core principle of this method lies in the organic combination of the stability and interpretability of LPF, the adaptive capability of reinforcement learning, and the lightweight characteristics of policy distillation.

[0037] like Figure 2 As shown, in this photovoltaic power output smoothing control method, the overall battery control action is composed of the benchmark action generated by the LPF and the residual correction amount output by the reinforcement learning agent. The complex model is transformed into a simplified model through the policy distillation mechanism, which facilitates practical deployment. The specific implementation process is as follows: The original photovoltaic output signal is low-pass filtered to generate a baseline action; the LPF (Local Perforated Filter) is used to suppress high-frequency fluctuations in the photovoltaic output signal while preserving low-frequency trends. Its mathematical expression can be represented as a first-order discrete filtering formula:

[0038] In the formula, Indicates the reference action; Indicates the filter coefficients; This indicates the original photovoltaic power output signal; Indicates the current time; Indicates the previous moment.

[0039] smaller This results in a smoother output, but introduces greater latency; a larger... The response is faster, but the smoothing effect is weakened. This embodiment obtains a stable baseline trajectory through LPF, providing a priori reference for the reinforcement learning agent (reinforcement learning model) to carry out policy optimization.

[0040] The original state space vector is input into the reinforcement learning agent of the pre-constructed teacher model, and the residual action is output. The residual action is calculated using the PPO algorithm, which constrains the update magnitude of the old and new policies to avoid instability caused by excessive policy updates. The optimization objective function of the reinforcement learning agent is:

[0041] In the formula, This represents the objective function to be optimized. This indicates the ratio between the old and new strategies; This represents the estimation of the advantage function; Indicates the cutting range; Expressing expectations; Represents a numerical clipping function; Describes the minimum value function; Indicates the strategy parameters; Indicates the current moment; where, ; Represents the new strategy function; Represents the old strategy function; Indicates the action taken; Indicates the system status.

[0042] In this embodiment, the objective function can improve policy performance while limiting the differences between the old and new policies, avoiding excessive parameter updates, thereby ensuring the stability and convergence of the policy learning process. In this embodiment, the input to the PPO agent is a state space vector: Output residual action ; used to compensate for the deficiencies of LPF. In the formula, This represents the change in the original photovoltaic output signal; Indicates the battery's state of charge; Indicates the reference action; Indicates the final action instruction; This represents the historical state of photovoltaic power. Historical state quantities representing changes in photovoltaic power; Indicates the current time; Indicates the previous moment.

[0043] The synthesized action instructions are determined by the following relationship:

[0044] In the formula, Indicates a combined action instruction; Indicates the reference action; Indicates residual action; Indicates the current moment.

[0045]

[0046]

[0047]

[0048] In the formula, Indicates the final action instruction; Indicates a combined action instruction; Indicates the maximum allowable discharge power; Indicates the maximum allowed charging power; Indicates the rated limit of battery power; Describes the minimum value function; Represents the maximum value function; Indicates the battery's state of charge; and These represent the minimum and maximum allowable state of charge of the battery, respectively. Indicates the rated capacity of the battery; Indicates the control cycle time step; Indicates the current moment.

[0049] In this embodiment, the synthesized action command undergoes a power limit and SOC constraint mechanism before output, thereby ensuring physical feasibility: In this embodiment, to further consider battery health, in the process of synthesizing action command constraints, not only are power limits and SOC boundaries applied, but an SOC control factor is also introduced into the reward function, thereby avoiding frequent deep cycles and reducing battery loss. In this method, the assessment of battery life loss (cumulative loss) adopts the rainflow counting method, which can decompose irregular charge-discharge cycles into equivalent standard cycles, and then estimate the cumulative loss of the battery.

[0050] The specific principle of the rainflow counting method is as follows: the equivalent cycle depth formed during the charging and discharging process is identified by rainflow counting. Based on the empirical lifetime model, the equivalent number of cycles and the corresponding lifetime attrition rate are calculated; the formula for calculating the cumulative loss L is:

[0051] In the formula, This indicates the total number of cycles identified by the rainflow counter; Indicates the first The depth of each charge-discharge cycle; Indicates the first The number of times a charge-discharge cycle occurs; This indicates that in a loop with a depth of The number of battery cycles under the rated life.

[0052] After the above steps are performed, the final action instructions are output through the teacher model.

[0053] Based on the distillation loss function, the final action command output by the teacher model is used as a soft label to train the student model, resulting in a lightweight control model. In this embodiment, the baseline action is removed from the student model to reduce its input state space.

[0054] Policy distillation is used to transform complex teacher models into lightweight student models. The basic idea is to use the teacher model's output actions on the training set as soft labels, and to achieve knowledge transfer by minimizing the difference between the student model's and the teacher model's outputs.

[0055] Let the student model parameters be... The first student model The output results are The first teacher model The output results are Distillation loss function Defined as:

[0056] In the formula, The number of samples, taking values... , This represents the total number of training samples.

[0057] In this method, the student model learns the control policy of the teacher model by minimizing the distillation loss function described above. In this embodiment, through policy distillation, the student model can significantly reduce the number of parameters and computational overhead compared to the teacher model. The distilled student model maintains smooth performance close to that of the teacher model while significantly improving inference speed, making it more suitable for deployment in edge control devices of photovoltaic power plants.

[0058] Finally, the trained student model, namely the lightweight control model, is deployed in the edge control equipment of the photovoltaic power station. The lightweight control model is used to regulate the real-time photovoltaic output signal and output real-time control commands.

[0059] In summary, the overall principle of this embodiment can be summarized as "filtering first, residual correction second, and distillation simplification third." Specifically, firstly, the original photovoltaic output signal is filtered using an LPF to obtain a smoothed baseline action. This process effectively suppresses high-frequency fluctuations, providing a stable prior for subsequent control. Secondly, the PPO-based reinforcement learning agent outputs residual actions under the constraints of the observed state and battery SOC health. Synthetic battery charging and discharging commands (synthetic action commands) are formed through action decomposition relationships. Before execution, the synthesized action commands undergo constraints based on power limits and SOC boundaries to ensure compliance with physical feasibility and battery safety. During the training phase, the teacher model learns a policy using the PPO in an environment containing LPF and residual information. Then, through a policy distillation mechanism, the teacher model's output on the training set is used as a "soft label" to guide the student model to learn the same control policy under simplified input features. The distilled student model has a smaller structure and lower computational cost, yet maintains a smoothness similar to the teacher model. This process, which uses LPF to provide baseline stability, PPO residual compensation to achieve adaptive optimization, and distillation mechanism to achieve lightweight deployment, enables the control method to balance performance, interpretability, and engineering feasibility.

[0060] like Figure 4As shown, this embodiment also provides a photovoltaic power output smoothing control system, including: a filtering module for low-pass filtering the original photovoltaic power output signal to generate a reference action; a reinforcement learning module for inputting the original state space vector into a pre-constructed teacher model reinforcement learning agent and outputting residual actions; an action superposition module for superimposing the reference action and the residual action to generate a synthetic action command; a correction module for correcting the synthetic action command using a constraint mechanism of power limit and SOC boundary to generate a final action command; the final action command is output by the teacher model; a model training module for training a student model based on a distillation loss function, using the corrected final action command output by the teacher model as a soft label to obtain a lightweight control model; and a control module for controlling the real-time photovoltaic power output signal using the lightweight control model and outputting a real-time control command.

[0061] The present invention also provides a photovoltaic power output smoothing control device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the photovoltaic power output smoothing control method.

[0062] The present invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the photovoltaic power output smooth control method.

[0063] When the processor executes the computer program, it implements the above-mentioned steps for smooth control of photovoltaic output, for example: The original photovoltaic power output signal is low-pass filtered to generate a baseline action. The original state space vector is input into a reinforcement learning agent of a pre-constructed teacher model, which outputs a residual action. The reinforcement learning agent uses a PPO-based algorithm. The baseline action and the residual action are superimposed to generate a synthetic action command. The synthetic action command is modified using a constraint mechanism based on power limits and SOC boundaries to generate a final action command. The final action command is output by the teacher model. Based on the distillation loss function, the final action command output by the teacher model is used as a soft label to train the student model to obtain a lightweight control model. The lightweight control model is used to regulate the real-time photovoltaic power output signal and output a real-time control command.

[0064] For example, the computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing preset functions, wherein the instruction segments describe the execution process of the computer program in the photovoltaic power output smoothing and control device. For example, the computer program can be divided into: a filtering module for low-pass filtering the original photovoltaic power output signal to generate a baseline action; a reinforcement learning module for inputting the original state space vector into a pre-constructed reinforcement learning agent of a teacher model and outputting residual actions; wherein the reinforcement learning agent adopts a PPO-based algorithm; an action superposition module for superimposing the baseline action and the residual action to generate a synthetic action command; a correction module for correcting the synthetic action command using a constraint mechanism of power limit and SOC boundary to generate a final action command; the final action command is output by the teacher model; a model training module for training a student model based on a distillation loss function, using the corrected final action command output by the teacher model as a soft label to obtain a lightweight control model; and a control module for controlling the real-time photovoltaic power output signal using the lightweight control model and outputting a real-time control command.

[0065] The photovoltaic power output smoothing and control device can be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. The photovoltaic power output smoothing and control device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above are examples of photovoltaic power output smoothing and control devices and do not constitute a limitation on the photovoltaic power output smoothing and control device. It may include more components than described above, or combine certain components, or different components. For example, the photovoltaic power output smoothing and control device may also include input / output devices, network access devices, buses, etc.

[0066] The processor referred to can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or any conventional processor, etc. The processor is the control center of the photovoltaic power output smoothing regulation, connecting various parts of the entire photovoltaic power output smoothing regulation equipment through various interfaces and lines.

[0067] The memory can be used to store the computer program and / or modules. The processor realizes various functions of the photovoltaic output smoothing control device by running or executing the computer program and / or modules stored in the memory and calling the data stored in the memory.

[0068] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function (such as sound playback, image playback, etc.). The data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). Furthermore, the memory may include high-speed random access memory and non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital cards (SD cards), flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0069] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the photovoltaic power output smoothing control method described above.

[0070] If the modules / units integrated in the photovoltaic output smoothing control system are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0071] Based on this understanding, the present invention can implement all or part of the processes in the above-mentioned photovoltaic power output smoothing control method, or it can be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the above-mentioned photovoltaic power output smoothing control method. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or a preset intermediate form, etc.

[0072] The computer-readable storage medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0073] It should be noted that the content contained in the computer-readable storage medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable storage medium does not include electrical carrier signals and telecommunication signals.

[0074] In summary, this method has the following significant advantages compared to traditional control methods: First, enhanced performance of action residuals: Through the design of "low-pass filtering baseline action + reinforcement learning residual correction", this invention combines the stability and interpretability of filtering methods with the flexibility and adaptability of reinforcement learning, which is significantly better than the simple LPF method and end-to-end reinforcement learning method, and performs better in terms of power smoothing performance and dynamic response.

[0075] Secondly, this invention also considers battery health protection: While implementing smooth control, it introduces a State of Charge (SOC) control factor and uses rainflow counting to assess battery loss. This method effectively reduces the number of deep charge / discharge cycles and high-frequency switching, significantly reducing cumulative battery loss, slowing down battery aging, and extending the lifespan of the energy storage system.

[0076] Third, lightweight design facilitates edge deployment: This invention adopts a "teacher-student" strategy distillation mechanism to compress complex residual reinforcement learning strategies into lightweight student models, significantly reducing the number of parameters and computational overhead, enabling the control model to run independently on power plant edge devices, and possessing rapid inference and real-time response capabilities.

[0077] The above embodiments are merely one of the implementation methods for achieving the technical solution of the present invention. The scope of protection claimed by the present invention is not limited to this embodiment, but also includes any variations, substitutions and other implementation methods that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for smoothing the output of a photovoltaic plant, characterized in that, The method comprises the following steps: performing low-pass filtering on an original photovoltaic output signal to generate a benchmark action; inputting an original state space vector into a pre-constructed reinforcement learning agent of a teacher model to output a residual action; wherein the reinforcement learning agent adopts an algorithm based on PPO; superimposing the benchmark action and the residual action to generate a synthesized action instruction; correcting the synthesized action instruction by using a constraint mechanism of a power limit and an SOC boundary to generate a final action instruction; the final action instruction is output by the teacher model; training a student model based on a distillation loss function, taking the final action instruction output by the teacher model as a soft label, to obtain a lightweight regulation and control model; regulating and controlling a real-time photovoltaic output signal by using the lightweight regulation and control model to output a real-time regulation and control instruction.

2. The photovoltaic power output smoothing method according to claim 1, characterized in that, The step of performing low-pass filtering on the original photovoltaic output signal to generate the benchmark action comprises the following steps: obtaining the original photovoltaic output signal; filtering the original photovoltaic output signal by using a low-pass filter to generate the benchmark action; a specific first-order discrete filtering formula is as follows: wherein represents a reference action; represents a filter coefficient; represents an original photovoltaic power signal; represents a current time instant; represents a previous time instant.

3. The photovoltaic power output smoothing method according to claim 1, characterized in that, The step of inputting the original state space vector into the pre-constructed reinforcement learning agent of the teacher model to output the residual action comprises the following steps: obtaining the original state space vector; wherein the original state space vector comprises an environmental state and a battery state of charge; inputting the original state space vector into the pre-constructed reinforcement learning agent of the teacher model to output the residual action; wherein: a specific formula of an optimization objective function of the reinforcement learning agent is as follows: wherein represents an optimization objective function; represents a ratio of new and old policies; represents a merit function estimate; represents a clipping range; represents an expectation; represents a numerical clipping function; represents a minimum function; represents a policy parameter; represents a current time; The state space vector The specific expression is as follows: wherein represents a variation of the original photovoltaic power signal; represents a battery state of charge; represents a reference action; represents a final action instruction; represents a historical state quantity of the photovoltaic power; represents a historical state quantity of the photovoltaic power variation; represents a current time instant; represents a previous time instant.

4. The photovoltaic power output smoothing method according to claim 1, characterized in that, The step of superimposing the benchmark action and the residual action to generate the synthesized action instruction comprises the following specific calculation formula: In the formula, represents a synthetic motion instruction; represents a reference motion; represents a residual motion; represents a current time.

5. The photovoltaic power output smoothing method according to claim 1, characterized in that, In the step of correcting the synthesized action instruction by using the constraint mechanism of the power limit and the SOC boundary to generate the final action instruction, a specific correction formula is as follows: wherein represents a final action command; represents a synthesized action command; represents the maximum allowed discharging power; represents the maximum allowed charging power; represents the rated limit of the battery power; represents a minimum function; represents a maximum function; represents the battery state of charge; and respectively represent the minimum and maximum state of charge allowed by the battery; represents the battery rated capacity; represents the control period time step, represents the current time.

6. The photovoltaic power output smoothing method according to claim 1, characterized in that, In the step of correcting the synthesized motion instruction by using the constraint mechanism of the power limit and the SOC boundary to generate the final motion instruction, an SOC control factor is introduced into a reward function to correct the residual motion; wherein the accumulated loss of the SOC is calculated by using a rain flow counting method, and the accumulated loss of the SOC is used as a constraint condition for the final motion instruction The specific calculation formula is as follows: wherein represents the total number of cycles identified by the rainflow counting; represents the depth of the th charge and discharge cycle; represents the number of times the th charge and discharge cycle occurs; represents the number of battery life cycles at a cycle depth of .

7. The photovoltaic power output smoothing method according to claim 1, characterized in that, The step of training the student model based on the distillation loss function, taking the final action instruction output by the teacher model as the soft label, to obtain the lightweight regulation and control model comprises the following steps: inputting the original photovoltaic output signal into the pre-constructed student model to obtain an output result of the student model; taking the final action instruction output by the teacher model as a soft label for training of the student model; wherein the final action instruction is an output result of the teacher model; training the student model by using the distillation loss function, combining the output result of the teacher model and the output result of the student model, to obtain the lightweight regulation and control model; wherein: The distillation loss function A specific expression of the distillation loss function is as follows: In the formula, is the number of samples, and the value , is the total number of training samples; represents the first output result of the student model; represents the first output result of the teacher model; represents the student model parameters.

8. A photovoltaic power output smoothing system, characterized by, The method comprises the following steps: a filtering module, configured to perform low-pass filtering on an original photovoltaic output signal to generate a benchmark action; a reinforcement learning module, configured to input an original state space vector into a pre-constructed reinforcement learning agent of a teacher model to output a residual action; wherein the reinforcement learning agent adopts an algorithm based on PPO; an action superimposition module, configured to superimpose the benchmark action and the residual action to generate a synthesized action instruction; a correction module, configured to correct the synthesized action instruction by using a constraint mechanism of a power limit and an SOC boundary to generate a final action instruction; the final action instruction is output by the teacher model; The model training module is configured to train the student model based on a distillation loss function, taking the corrected final action instruction output by the teacher model as a soft label, to obtain a lightweight regulation and control model. The regulation and control module is configured to regulate and control the real-time photovoltaic output signal by using the lightweight regulation and control model, and output a real-time regulation and control instruction.

9. A photovoltaic power output smoothing device, characterized by, The computer program is configured to implement the steps of the photovoltaic output smoothing regulation and control method according to any one of claims 1-7 when executed by the processor. The computer program is configured to implement the steps of the photovoltaic output smoothing regulation and control method according to any one of claims 1-7 when executed by the processor. ​ 10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. ​

Images